Glacial isostatic adjustment in Greenland is the result of Holocene ice thickness changes in North America and Greenland and late Holocene deglaciation in Greenland. There is also evidence for substantial lateral variation in viscosity, such as below viscosity in East Greenland possibly due to the passing of the former Iceland hotspot. Here we investigate how viscosity changes influence the uplift rates in Greenland, and we focus on the contribution of North America and the effect of stress-dependent viscosity.

We derive viscosity maps from temperature estimates which are obtained in two ways. One is based on gravity, seismic and petrology data (WINTERC-G). The other is based on seismic velocity anomalies from a new seismic model for the north Atlantic (NAT2020) that are converted to temperature anomalies. Viscosity maps are dependent on parameters in olivine flow law. From the various 3D maps, the ones with best fit to GPS velocities in East Greenland are selected. All models lead to a lithosphere thicker than 100 km, with a viscosity below that < 1019 Pa s.

We show that the viscosity estimates change in time, with 1 order of magnitude lower viscosity between 12 and 8 kyear before present. Stress due to North American deglaciation has negligible contribution to these viscosity changes, so these are caused by ice melt in Greenland. A relative uniform forebulge effect across Greenland is found of 1-2 mm/year. Both lateral viscosity variation and stress-dependent viscosity have a significant effect on present-day uplift rates.

The Aerosol Index (AER_AI) as measured by the Tropospheric Monitoring Instrument (TROPOMI) on board the Sentinel-5 Precursor (S5P) platform maintains heritage with aerosol index data records from other satellite instruments including TOMS (EP and Nimbus-7), GOME(-2), OMI and OMPS. The operational TROPOMI AER_AI dataset is available from May 2018 through to the present. It is useful not only for tracking ultraviolet (UV) absorbing aerosol plumes of desert dust, volcanic ash, and smoke from biomass burning but also for monitoring the quality of the TROPOMI Level 1b (L1b) data since the AER_AI calculation is very sensitive to the absolute calibration of irradiance and radiance. This work will also include a description of the impacts on AER_AI due to the observed, wavelength-dependent degradation in the TROPOMI measured irradiance and radiance. Results of recent updates in the L1b data which incorporate corrections for these degradation-driven features will be discussed.

With a nadir pixel size of 3.5 x 5.5 km, TROPOMI’s high spatial resolution presents specific challenges as non-Lambertian cloud features, cloud shadows, and 3-D effects of clouds are now visible in the TROPOMI AER_AI data. An overview of plans to address these features in future AER_AI updates will be given including a description of planned implementation of a new AI datafield utilizing two rather than one LER surfaces-- a so-called bi-reflector approach.

Because TROPOMI AER_AI contributes to the continuation of aerosol index measurements from other missions, it is important to compare TROPOMI to other data records. OMPS and OMI data will be used in case studies to lend insights about known biases in the TROPOMI AI dataset. Comparisons will be conducted for several aerosol plume and scene types.

FP_ILM algorithm applied to UVN sensors for the retrieval of GE_LER climatologies for UV/VIS trace gases and clouds

Ana del Águila(1)*, Pascal Hedelt(1), Klaus-Peter Heue(1),(2), Ronny Lutz(1), Víctor Molina García(1), Fabian Romahn(1), Jian Xu(1),(3), Ka Lok Chan(1),(4), Diego Loyola(1)

(1) Remote Sensing Technology Institute, German Aerospace Center (DLR), 82234 Weßling, Germany
(2) Technical University of Munich (TUM), 80333 Munich, Germany
(3) National Space Science Center, Chinese Academy of Sciences, Beijing, China
(4) Rutherford Appleton Laboratory Space, Harwell Oxford, United Kingdom

The accurate retrieval of trace gases, aerosols and clouds from UVN sensors require precise information on the surface properties at global scale which are traditionally obtained from Lambertian Equivalent Reflectivity (LER) climatologies. New satellite missions usually use LER climatologies based on previous missions encountering a number of drawbacks, such as: (a) lower spatial resolution, (b) different satellite-viewing geometry dependencies or (c) strong differences from the actual surface conditions under snow/ice conditions.
The accurate knowledge of surface reflectance at global scale is of great importance since the neglection of surface anisotropic effects can lead to sources of error in the retrievals. The well-stablished full-physics inverse learning machine (FP_ILM) algorithm [1] allows to retrieve the geometry-dependent effective Lambertian equivalent reflectivity (GE_LER) from UVN sensors in a monthly basis. The FP_ILM algorithm consists of a training phase with four steps: (1) forward model, (2) smart sampling, (3) feature extraction and (4) machine learning; and a retrieval phase. The GE_LER retrieval is solved using neural networks trained with radiative transfer model simulations with inputs similar to the AMF calculations [2]: DOAS fitting polynomial coefficients, fitted trace gas slant column amounts and satellite viewing geometry.
The FP_ILM GE_LER algorithm has been applied to TROPOMI/S5P measurements and it is being used for operational retrieval of total ozone and cloud properties. The TROPOMI GE_LER climatology results for the fitting windows in the UV/VIS corresponding to SO2 (311-325 nm), O3 (325-335 nm), HCHO (328.5-346 nm), NO2 (405-465 nm), H2O (430-466.5 nm) and clouds (755-771 nm) are presented. In addition, the retrieved GE_LER climatologies with grid resolution 0.1ºx0.1º from this study are compared with climatological GOME-2 and OMI LER data with grid resolutions of 0.25ºx0.25º and 0.5ºx0.5º, respectively. The results of the comparison show effective high-resolution GE_LER climatologies for UV/VIS trace gases and the operational UVN cloud product.

[1] Loyola, D. G., Xu, J., Heue, K.-P., and Zimmer, W.: Applying FP_ILM to the retrieval of geometry-dependent effective Lambertian equivalent reflectivity (GE_LER) daily maps from UVN satellite measurements, Atmos. Meas. Tech., 13, 985–999, https://doi.org/10.5194/amt-13-985-2020, 2020.
[2] Chan, K. L., Valks, P., Slijkhuis, S., Köhler, C., and Loyola, D.: Total column water vapor retrieval for Global Ozone Monitoring Experience-2 (GOME-2) visible blue observations, Atmos. Meas. Tech., 13, 4169–4193, https://doi.org/10.5194/amt-13-4169-2020, 2020.

The validation of water vapour isotopologue satellite data products (such as SRON's scientific TROPOMI HDO product and the S5P+I H2O-ISO product) is currently hindered by known biases in ground-based Fourier transform infrared (FTIR) reference data. For example, the HDO product by the Total Carbon Column Observing Network (TCCON) is lacking a so-called in situ correction factor that corrects biases due to uncertainties in the spectroscopy which tend to be highly reproducible. In order to derive this calibration, in situ profile measurements concurrently to TCCON observations are necessary, however no such profile measurements are available to date. Conventional in situ techniques utilize aircraft or large balloon platforms that are costly and need significant effort and infrastructure. Instruments for aircraft have to go through an extensive permission procedure. Large balloons can only be launched at a few dedicated launch sites and cannot be steered, making them unsuitable for reference measurements at many places such as most ground-based remote sensing stations. Due to these limitations, profile measurements of water vapour isotopologues are sparse. In the project Water vapour Isotopologue Flask sampling for the Validation Of Satellite data (WIFVOS), a novel balloon-borne instrument for small meteorological balloons (< 20kg payload) capable of measuring profiles of water vapour isotopologues is developed. The new instrument is much more cost-effective and flexible than previous ones. It is based on a flask-sampling technique proven on drones that is adapted to lower pressures and water vapour mixing ratios up to the tropopause. Air is sampled in flasks at pre-programmed pressure levels during descent of the payload. Sampling on descent avoids potential contaminations by humidity brought up by the balloon. After landing and recovery, the samples are analysed with a cavity-ringdown spectrometer. This contribution presents the instrument design, the calibration at a laboratory in Bergen and first results from a field campaign with concurrent TCCON measurements at Sodankylä in spring 2022.

FTIR Team: Carlos Aquino Bauer 2, Thomas Blumenstock 3, Martine De Mazière 1, Michel Grutter 4, James Hannigan 5, Nicholas Jones 6, Rigel Kivi 7, Emmanuel Mahieu 8, Maria Makarova 9, Isamu Morino 10, Isao Murata 11, Tomoo Nagahama 12, Justus Notholt 13, Ivan Ortega 5, Mathias Palm 13, Markus Rettinger 14, Amelie Röhling 3, Dan Smale 15, Wolfgang Stremme 4, Kim Strong 16, Youwen Sun 17, Ralf Sussmann 14, Yao Té 18, Pucai Wang 19, Tyler Wizenberg 16

1 Royal Belgian Institute for Space Aeronomy (BIRA-IASB), Brussels, Belgium; 2 Instituto Federal de Educaçao, Ciência e Tecnologia de Rondônia (IFRO), Porto Velho, Brazil; 3 Karlsruhe Institute of Technology (KIT), IMK-ASF, Karlsruhe, Germany; 4 Universidad Nacional Autónoma de México (UNAM), 04510 Mexico City, México; 5 National Center for Atmospheric Research (NCAR), Boulder, CO, USA; 6 Centre for Atmospheric Chemistry, University of Wollongong, Wollongong, Australia; 7 Finnish Meteorological Institute (FMI), Sodankylä, Finland; 8 Institute of Astrophysics and Geophysics, Université De Liège, Liège, Belgium; 9 Saint Petersburg State University, Atmospheric Physics Department, St Petersburg, Russia; 10 National Institute for Environmental Studies (NIES), Tsukuba, Japan; 11 Graduate School of Environment Studies, Tohoku University, Japan; 12 Institute for Space-Earth Environmental Research (ISEE), Nagoya University, Nagoya, Japan; 13 Institute of Environmental Physics, University of Bremen, Bremen, Germany; 14 Karlsruhe Institute of Technology (KIT), IMK-IFU, Garmisch-Partenkirchen, Germany; 15 National Institute of Water and Atmospheric Research Ltd (NIWA), Lauder, New Zealand; 16 Department of Physics, University of Toronto, Toronto, Canada; 17 Hefei Institute of Physical Sciences, Chinese Academy of_Sciences (CAS), Hefei, China; 18 LERMA-IPSL, Sorbonne Université, CNRS, PSL Research University, 75005 Paris, France; 19 Institute of Atmospheric Physics, Chinese Academy of Sciences (CAS), Beijing, China

Within the NIDFORVal project (S5P NItrogen Dioxide and FORmaldehyde VALidation using NDACC and complementary FTIR and UV-Vis DOAS ground-based remote sensing data), we have successfully obtained a harmonized HCHO data set from the FTIR network (Vigouroux et al., 2018). This has led to a comprehensive validation of the S5P HCHO product (Vigouroux et al., 2021), showing the strength of using more than 20 sites covering clean and polluted conditions. Within this NIDFORVAL project, the total, tropospheric, and stratospheric NO2 S5P products have been validated, against Direct-Sun, MAX-DOAS (Multi-Axis Differential Optical Absorption Spectroscopy), and zenith-sky DOAS measurements, respectively (Verhoelst et al., 2021).
Fourier Transform Infrared (FTIR) instruments have the capability to measure NO2, with a sensitivity mainly located in the stratosphere (e.g., Hendrick et al., 2012, Bognar et al., 2019). However, only a few FTIR sites exploited this until now, using different retrieval settings. For the present work, we have optimized the NO2 retrieval settings and applied them consistently to the whole FTIR network (mostly from NDACC, Network for the Detection of Atmospheric Composition Change, but also including additional NDACC candidate sites and TCCON sites operated in NDACC mode). We have obtained a unique harmonized NO2 data set covering 25 FTIR sites, ensuring consistency of the results if used as reference data for validation. This stratospheric NO2 data set can complement the zenith-sky DOAS data. Indeed, the zenith-sky DOAS observations are made during sunset and sunrise which imposes the use of a photochemical box model to adjust the observations to the time of the TROPOMI overpasses, while the FTIR measurements are made during the whole day, allowing direct comparison between measurements that are collocated in time.

In this presentation, we will show the validation results of more than three years of S5P stratospheric NO2 data, allowing robust statistics on the comparisons, and the seasonal cycles to be compared at 25 FTIR sites globally distributed. Diurnal cycles comparisons can also be obtained at high latitude sites. Conclusions about the accuracy and the precision of the S5P stratospheric NO2 products will be drawn and compared to the ones obtained using Zenith-sky DOAS data (Verhoelst et al., 2021).


Bognar et al.: Updated validation of ACE and OSIRIS ozone and NO2 measurements in the Arctic using ground-based instruments at Eureka, Canada, JQSRT, 238, https://doi.org/10.1016/j.jqsrt.2019.07.014, 2019.
Hendrick et al.: Analysis of stratospheric NO2 trends above Jungfraujoch using ground-based UV-visible, FTIR, and satellite nadir observations, Atmos. Chem. Phys., 12, 8851–8864, https://doi.org/10.5194/acp-12-8851-2012, 2012.
Verhoelst et al.: Ground-based validation of the Copernicus Sentinel-5P TROPOMI NO2 measurements with the NDACC ZSL-DOAS, MAX-DOAS and Pandonia global networks, Atmos. Meas. Tech., 14, 481–510, https://doi.org/10.5194/amt-14-481-2021, 2021.
Vigouroux et al.: NDACC harmonized formaldehyde time series from 21 FTIR stations covering a wide range of column abundances, Atmos. Meas. Tech., 11, 5049–5073, https://doi.org/10.5194/amt-11-5049-2018, 2018.
Vigouroux et al.: TROPOMI–Sentinel-5 Precursor formaldehyde validation using an extensive network of ground-based Fourier-transform infrared stations, Atmos. Meas. Tech., 13, 3751–3767, https://doi.org/10.5194/amt-13-3751-2020, 2020.

In the framework of the BAQUNIN (Boundary-layer Air Quality-analysis Using Network of Instruments, https://www.baqunin.eu) ESA project, three lidar system characterised by different design and measurement modes are operated in an urban (Rome downtown) and a rural (Montelibretti) sites of the Tiber Valley. This experimental setup is particularly useful for the evaluation of the impact of different environmental conditions (e.g. surface reflectivity, pollution load) on atmospheric composition satellite products. In this contribution we will discuss in detail the methods applied and the preliminary results of the validation of TROPOMI Level-2 Cloud products (offline), with particular focus on the estimatedcloud boundaries (top/bottom).
Two Lidar systems located in the Atmospheric Physics Laboratory (APL) of the Physics Department of Sapienza University (Rome downtown). At APL, a powerful multi-wavelength and multi-polarisation Lidar (MWL-LIDAR) has been designed and assembled using both custom made and commercial devices, while the controlling software has been developed in house.The system includes a large power pulsed laser, emitting pulses at 355, 532,1064 nm wavelengths, four receivers and twelve acquisition channels. MWL-LIDAR can be operative in no rain day and, for what concerns the Raman channels, in night conditions. In addition, a Raymetrics Aerosol Profiler (RAP) system, emitting at 1064 nm, is installed on the roof of APL a few meters apart from the MWL-LIDAR. Differently from the latter, RAP operates continuously (24/7). For the APL systems, the reference cloud boundary data are obtained from high time/space resolution Lidar range-corrected signal, providing very accurate estimates of the position of the cloud edges. The third system is a Vaisala CT25K ceilometer, located in the “Liberti” experimental campus of CNR-IIA, about 25 km distance from APL. This ceilometer is also operated in continuous mode and can provide reliable cloud bottom altitudes and, in case of optically thin clouds, the cloud top altitudes.
The TROPOMI Level-2 offline products are gathered from the ONDA-DIAS system using a dedicated selection/extraction tool called TROPEX (TROPomiEXtraction) developed in the context of the BAQUNIN project.
The validation approach consists in selecting the TROPOMI products for which the satellite observing view around the selected sites is almost nadir (viewing zenith angle < 10 degrees) and a variable horizontal search radius. The inter-comparison results are finally analysed in terms of satellite viewing geometry and cloud fraction, of relative distance from the site, and, in the case of APL, of the aerosol load retrieved from AERONET and EUROSKYRAD collocated systems (also part of BAQUNIN).

Antarctic geothermal heat flow (GHF) has often been derived indirectly from geophysical data with assumptions about a simplified and undifferentiated lithosphere, which resulted in weakly constrained and inconsistent models. From other continents, we know that thermal parameters and GHF can exhibit large spatial variations depending on geology and tectonic history.
Combining gravity and magnetic data in a joint inversion approach yields information on the crustal structure of East Antarctica and possible geological features become more evident. Both datasets are combined through a coupling method which increases the mutual information to get similar and statistically compatible inversion results. Therefore, we minimize data misfit and model similarity under the coupling constraint. The results show matching features of high magnitude density and susceptibility anomalies. Prominent structures are visible in NW – SE direction along the edge of the Mawson craton and at the presumed Australo-Antarctic and Indo-Antarctic terrane boundary.
With the structural similarity and resulting parameter relationship between inverted susceptibility and density, we aim to define a spatially variable heat production map, which in turn would lead to improved GHF estimates. For this, we rely on existing petrophysical and geochemical databases to correlate and confine thermal parameters with our results.

The observed secular variation of the geomagnetic field is assumed to reflect the convective dynamics of the Earth’s fluid outer core. Time-dependent models of the field are used to invert for the fluid flow just beneath the core surface. However, it is unclear how well rapid core motions on interannual or shorter time scales can be resolved with current geomagnetic field models.
To address this question, we present a study in which we use a state-of-the-art numerical geodynamo simulation that exhibits such rapid dynamics. From the simulation data, we generate synthetic magnetic field observations with a temporal and spatial distribution equivalent to ground and satellite records. We then construct a magnetic field model by solving the inverse problem constrained by this synthetic data. The obtained model is in turn inverted for the core surface motions. Based on the comparison of the inverted flow with the reference flow provided by the geodynamo simulation, we discuss how accurately we can expect to recover the Earth's core surface motions when applying the same methods to geomagnetic records. In particular, we focus on whether we can detect hydromagnetic waves such as geostrophic torsional Alfvén waves (torsional oscillations) and quasi-geostrophic Alfvén waves, the latter of which have been associated with geomagnetic jerks.

Due to improvements in earth-system modelling as well as to the increasing accuracy of geodetic observation systems, there is a rising demand for improved deformation and load-correction models. Usually, load deformations are quantified applying load Love numbers or Green's functions which are based on one dimensional (1D) parameterisations of the earth's structure. Improvements due to local deviations in deformational response can be achieved by adjusting the 1D parameterisation to the region of interest (e.g., Dill et al., 2015). Drawback of this procedure is that a specific deformation model holds only for a specific region and, therefore, it cannot be applied consistently to large scale processes like ocean-tide loading or when deriving global quantities like the geocentre motion. In a recent publication (Huang et al., 2021), we discussed the impact of a laterally heterogeneous anelastic mantle structure on ocean-tide loading and its feedback on ocean-tide dynamics applying the Dynamic ocean-tide model TiME (Sulzbach et al., 2021). Therein, we limited the forcing setup to the M2 ocean tide.

In this study which will be presented also in Huang et al. (2021, in prep.), we extend the model set up. In addition to mantle anelasticity, we consider additional structural features inside the lithosphere like sediment layers, differences in crustal composition and variations especially between continental and oceanic lithosphere. We discuss the impact on loading at selected sites as well as the feedback to tidal dynamics, and extend the forcing to the fortnightly Mf tide. Furthermore, the comparison of two different lithosphere models, WINTERC-G and LITHO1.0, resulted already in differences of up to 1 mm concerning tidal deformations. This result together with the interplay between different structural features of lithosphere and mantle demands a careful choice of the appropriate lithosphere structure.

This study contributes to the national Earth system modelling initiative, ESM Innovation Pool and PalMod, and is considered in the GRACE-FO project within the ESA Third Party Mission Program.

Lit.:
Dill R, Klemann V, Martinec Z, Tesauro M, 2015. Applying local Green's functions to study the influence of the crustal structure on hydrological
loading displacements, J. Geodyn. 88:14-22. https://doi.org/10.1016/j.jog.2015.04.005
Huang, P., Sulzbach, R. L., Tanaka, Y., Klemann, V., Dobslaw, H., Martinec, Z., Thomas, M. (2021). Anelasticity and lateral heterogeneities in Earth's upper mantle: Impact on surface displacements, self-attraction and loading and ocean tide dynamics. J. Geophys. Res.: Solid Earth, 126:e2021JB022332. https://doi.org/10.1029/2021JB022332
Sulzbach, R., Dobslaw, H., Thomas, M. (2021). High-resolution numerical modeling of barotropic global ocean tides for satellite gravimetry. J. Geophys. Res.: Oceans, 126:2020JC017097. https://doi.org/10.1029/2020JC017097

Several alternative gravity forward modelling methodologies and associated numerical codes with their own advantages and limitations are available for the Solid Earth community. With the upcoming state-of-the-art lithosphere density models and accurate global gravity field data sets it is vital to understand the opportunities and limitations of the various approaches. In this paper, we discuss the four widely used techniques: spherical harmonics, tesseroid integration, triangle integration, and hexahedral integration approach. A constant density shell benchmark shows that all four codes can produce similar gravitational potential fields suitable for reproducing satellite-acquired gravity data. Two additional shell tests were conducted with more complicated geological structures: lateral varying density structures and a Moho density interface between crust and mantle. The differences were all below 1.5 percent of the modeled gravity signal, where the tesseroid and global spherical harmonics approach produced the most similar results (less than 0.3 percent).

To examine the usability of the forward modelling codes for realistic geological structures, we use the global lithosphere model WINTERC-G constrained, among other data, by satellite gravity field data computed using a GSH approach. Short-wavelength noise is present between the spectral and tesseroid forward modelling approaches likely related to the different way in which the spherical harmonic analysis of the varying boundaries of the mass layer is performed. The latter produces small differences especially at sharp interfaces, introducing mostly short-wavelength differences. Nevertheless, both approaches result in accurate solutions of the potential field with reasonable computational resources. Differences below 0.5 percent are obtained, resulting in residuals of 0.076 mGal standard deviation.

The other two spatial forward modelling schemes (triangle-based and hexagonal-based) have more difficulty in reproducing similar gravity field solutions. They need to go to unrealistic high resolutions, resulting in enormous computation efforts. The biggest issue for the triangle approach is the characteristic pattern in the residuals that is related to the grid layout. Increasing the resolution and filtering allows for the removal of most of this erroneous pattern, but at the expense of higher computational loads with respect to the other codes. The hexahedron-based code performs worst in the forward modelling of the gravity signature of a density contrast of a depth-varying boundary. Care must be taken with any forward modelling software as the approximation of the geometry of the WINTERC-G model may deteriorate the gravity field solution, if the limitations of this model are not taken into account.

As of today a series of low Earth orbiting satellites (with the finished missions Oersted, Champ, CryoSat-2 and currently Swarm) have been almost continuously monitoring the geomagnetic field for about 2 decades. Using these observations along with ground-based measurements, we infer a model of the time-varying magnetic field above the core surface, dynamically related to core surface motions, using the pygeodyn data assimilation algorithm [1,2]. Assimilation of magnetic data is performed using an ensemble Kalman Filter (EnKF) where the forward stochastic model is constructed upon spatio-temporal statistical information extracted from a dynamo simulation of the latest generation [3]. Our algorithm is able to collect data presenting varying temporal coverage at any point in space. Following an augmented state approach, our re-analysis gives access to the core flow together with errors of representativeness (primarily composed of subgrid induction). Both contribute significantly to the observed secular variation (rate of change of the field). On top of ground-based observatory records, observations consist of geomagnetic virtual observatory series that interpolate magnetic data in a set of fixed locations at satellite altitude in four-monthly bins [4, 5]. These data are corrected as much as possible from contributions of the lithospheric field and external sources. From these the secular variation is computed as annual differences of these 4-monthly means, without further smoothing.
Based on a finite ensemble size, EnKF estimates may suffer from biases, such as ensemble collapse, or noise introduced by spurious correlations within the ensemble state members. In order to avoid these problems, we modify the canonical EnKf algorithm [6]. Rather than using localization (awkward for an application in the spectral domain) or ad hoc covariance inflation, we clean spurious correlations induced by the finite sampling size using the graphical lasso algorithm (“G-LASSO”, [7]). The method requires no fine tuning, as the extra parameter introduced by G-LASSO is estimated by cross-validation. This novel implementation first reduces significantly spurious high frequency variations otherwise introduced in the field and flow changes, in link with gaps present in the datasets, either in time towards the end/start of satellite missions, or in space towards high latitudes at epochs where no night-time data are available. G-LASSO furthermore enables the use of information contained in cross-correlations between in particular the core surface flow and errors of representativeness. These latter, also present in geodynamo simulation, are crucial to exploit as much as possible the information at interannual periods, of particular interest given the two decades covered by satellite data.
Our magnetic models are consistent with previous models projected in time onto splines and regularized, such as CHAOS-7 [8], although exhibiting some higher frequency signals linked with the absence of smoothing. Low order Gauss coefficients, particularly sensitive to high latitude patterns, do not show obvious signs of high-frequency oscillations, as sometimes seen in link with external field leakage. Core flows are dominated by long periods, characterized by the eccentric westward gyre [9]. Transient motions at first order satisfy to the quasi-geostrophic constraint, and are more intense toward the equator. We recover interannual Magneto-Coriolis modes of period 7 years discovered recently from the analysis of Gauss coefficients data from CHAOS-7 [10]. Our approach calls for further investigation of core flow variations on shorter periods.
[1] Huder, Gillet & Thollard, pygeodyn 1.1. 0: a Python package for geomagnetic data assimilation. Geoscientific Model Development, 12(8), 3795-3803, https://doi.org/10.5194/gmd-12-3795-2019 (2019)
[2] Gillet, Huder & Aubert, A reduced stochastic model of core surface dynamics based on geodynamo simulations. Geophys. J. Int., 219(1), 522-539, https://doi.org/10.1093/gji/ggz313 (2019)
[3] Aubert & Gillet, The interplay of fast waves and slow convection in geodynamo simulations nearing Earth’s core conditions, Geophys. J. Int., 225, 1854–1873, https://doi.org/10.1093/gji/ggab054 (2021)
[4] Hammer, Finlay & Olsen, Applications for CryoSat-2 satellite magnetic data in studies of Earth's core field variations, Earth Planets Space 73, https://doi.org/10.1186/s40623-021-01365-9 (2021)
[5] Hammer, Finlay & Olsen, Secular Variation Signals in Magnetic Field Gradient Tensor Elements derived from satellite-based Geomagnetic Virtual Observatories, Geophys. J. Int. (in revision)
[6] Evensen, The ensemble Kalman filter: Theoretical formulation and practical implementation. Ocean dynamics, 53(4), 343-367.2003, https://doi.org/10.1007/s10236-003-0036-9 (2003)
[7] Friedman, Hastie & Tibshirani, Sparse inverse covariance estimation with the graphical lasso. Biostatistics, 9(3), 432-441, https://doi.org/10.1093/biostatistics/kxm045 (2008)
[8] Finlay et al, The chaos-7 geomagnetic field model and observed changes in the south atlantic anomaly, Earth Planets Space, 72(1), 156, https://doi.org/10.1186/s40623-020-01252-9 (2020)
[9] Pais & Jault, Quasi-geostrophic flows responsible for the secular variation of the Earth's magnetic field. Geophys. J. Int., 173(2), 421-443, https://doi.org/10.1111/j.1365-246X.2008.03741.x (2008)
[10] Gillet et al, Satellite magnetic data reveal interannual modes in Earth’s core, Proc. Nat. Acc. Sci. (in revision)

Models of glacial isostatic adjustment (GIA) simulate the viscoelastic response of the solid Earth to surface loading induced mainly by mass re-distribution between ice and ocean during the last glacial cycle. These models predict relative sea-level change and surface deformation. The GIA component of present-day uplift is responsible for crustal deformation at rates over 10 mm/year in areas such as Churchill (Canada) and Angermanland (Sweden). As the building blocks of such GIA models have a degree of uncertainty, the model output needs to be validated against observational data. Conventionally, such a validation has been done by employing paleo and recent observations such as sea level data, surface deformation data, and gravity data, which feature errors of their own. Since sea level observations sparsely cover coastal areas and GIA inference from satellite observations is challenging, in this work, we validate displacements predicted by a new GIA model, VILMA-3D, by utilizing GNSS-derived vertical land motion. VILMA-3D solves the sea-level equation accounting for the redistribution of water between ice and ocean, deformation and gravitational effects, rotational feedback, floating ice and moving coastlines. We have created a GIA model ensemble using geodynamically constrained 3D Earth structures derived from seismic tomography to consider more realistic lateral variations in the GIA response. For the validation of VILMA-3D, we employ a multi-analysis-centre ensemble of GNSS station and geocentre motion coordinate solutions that are currently assimilated into ITRF2020, where similar data processing strategies were adopted. In estimating GNSS-derived velocities, tectonic and weather signatures were reduced, the former by the estimation thereof and the latter by employing non-tidal geophysical loading displacement models developed at the GFZ. The trend signal is extracted from these GNSS time series with the STL method (seasonal-trend decomposition based on Loess). Vertical deformation rates from stations in regions where the GIA signal exhibits large gradients is crucial for the GIA model validation. To overcome the obstacle imposed by the lack of GNSS station clusters in such regions, we employ a weighting scheme that involves the network density and the cross-correlation of the station displacement time series. That is, a station with a low correlation to all other stations represents a more unique location and, thus, gets a higher weight. As measures of agreement, we employ weighted root mean square error and weighted mean absolute error. With this validation, we can determine which parameter is most suitable for modelling present-day uplift rates. Further, regional differences in the model performance can be evaluated.
The results of this GIA model validation study will be the basis for further investigations within the GRACE-FO project of the ESA Third Party Mission Program as well as within the German Climate Modeling initiative PalMod.

The Wilkes Subglacial Basin (WSB) host one of the largest marine-based sectors of the East Antarctic Ice Sheet (EAIS), which stretches for ca 1600 km from the George V Coast towards South Pole. This sector is potentially more susceptible to climate change due to the sub sea-level topography and is predicted to be one of the largest contributors to sea level rise in the future. However, understanding the long-term stability of the EASI in the WSB area requires better quantification of the solid Earth contribution to the cryosphere, in particular the geothermal heat flow and the influence heterogeneity in subglacial geology and the lithospheric architecture have on the thermal structure in the WSB region.

Here we present subglacial geothermal heat flow estimates, one of the least constrained parameters of the entire Antarctic continent, based on lithospheric scale geophysical models. We use a combination of airborne radar magnetic and gravity data compilations, crustal and lithosphere thickness estimates from both satellite and airborne gravity complemented with independent passive seismic constraints and Curie depth point estimates to develop new geophysical models for the WSB area. We model a major boundary along the eastern margin, separating the Ross Orogen from a composite Precambrian Wilkes Terrane. Precambrian basement is modelled as both shallower and of more felsic bulk composition. We suggest based on the similarity in aeromagnetic anomalies that the basement represents late Paleoproterozoic to Mesoproterozoic igneous basement as exposed in South Australia. The formerly contiguous basement of South Australia is associated with high geothermal heat flow (80-120 mW/m2) due to anomalously high intracrustal heat production within highly radiogenic Paleo to Meso Proterozoic granitoids.

First order differences in basement depth, bulk composition and metasediment/sediment cover coupled with spatial varying intracrustal heat production is likely to give rise to significant heterogeneity in geothermal heat flow beneath different parts of the WSB. A hypothesis we will test further with our newly developed thermal models for this area.

We present WINTERC-G-v2-beta, a model derived from Waveform tomography and Gravity INversion for the TEmpeRature and Composition of the crust and mantle at global scale. Like its predecessor, WINTERC-G, this preliminary second version is constrained in the crust and upper mantle by: (a) Rayleigh and Love dispersion fundamental mode curves, (b) filtered satellite gravity field, (c) surface elevation and (d) heat flow data. The updated model incorporates additional constraints to image the the transition zone (410-660 km) and the lower mantle: (e) two surface wave overtones, (f) travel times of P and S phases (from the ISC-EHB Bulletin) and P-to-S receiver functions (from the project GLImER, Rondenay et al., 2017; http://stephanerondenay.com/glimer-web.html), (g) full spectrum satellite gravity field data (geoid anomalies and gravity gradients from ESA’s GOCE mission), and (h) a geodynamic estimate of the Earth’s surface elevation and CMB topography compatible with the model. We jointly invert the full data set to estimate 3D variations of the Earth’s crust and mantle physical properties: compressional wavespeed (Vp), shear wavespeed (Vs), density (ρ) and attenuation (Qs) derived as a function of temperature, pressure and composition (a thermodynamically self-consistent framework). The crust is parameterized based on its thickness (depth of the Moho discontinuity) and lithology. The Mantle is separated into four sections: the lithospheric mantle, the asthenospheric or sublithospheric upper mantle, the transition zone, and the lower mantle. Each layer is initially assumed to have a homogeneous composition (in terms of the amount of Al2O3 and FeO major oxides), and a reference geotherm that is perturbed during inversion. Standard features observed in other global models (slabs, hot spots, etc) are present in the final product.

The Greenland ice sheet is one of the largest contributors to global sea level rise, and it is constantly monitored. However, still little is known about the heat flux at the glacier bedrock, and how it affects the dynamics of the major outlet glaciers in Greenland. Recent studies suggest the hotspot currently under Iceland was located beneath eastern Greenland at ~40 Ma BP and that the upwelling of hot material from the Iceland plume towards Greenland is ongoing. A warm upper mantle has a low viscosity, which in turn causes the solid Earth to rebound much faster to deglaciation.
In the area of the Kangerlussuaq glacier, a large GPS velocities residual after removing predicted purely elastic deformations caused by present-day ice loss suggests the possibility of such fast rebound to little ice age (LIA) deglaciation. Here we investigate the lithospheric thickness and the mantle viscosity structure beneath SE-Greenland by means of model predictions of solid Earth deformation driven by a low viscosity mantle excited by the LIA deglaciation to the present day. From the comparison of such modeled deformations with the GPS residual, we conclude that
1) a rather thick lithosphere is preferred, i.e. between 90 and 100 km 2) and the upper mantle most likely has a viscosity that changes with depth. Assuming a two layer upper mantle, our conclusion is that upper mantle viscosity has to be quite low, but it is not well constrained which part of the upper mantle has to be low, with a preference for low viscosity in the deeper upper mantle.
To understand such results we implemented forward modelling with more realistic earth models, relying on improvements in seismic models, petrology and gravity data. This yields 3D viscosity maps that can be compared to inferences based on the 1D model and forms the basis for 3D GIA models.

The conclusion based on the 1D model can be explained with 3D Earth models. In the area of the Kangerlussuaq glacier the seismic derived viscosities prefer a higher viscosity layer above a lower viscosity one. This stems from the slow decrease in viscosity with depth. The layer that is characterized as shallow upper mantle still contains shallow regions with low temperatures, while the deeper upper mantle reaches low viscosities. Generally, for GIA earth models the “higher above lower” viscosity layering is unusual. However, the analysis of the 1D model clearly shows this to be one of the preferred model regions, in combination with a large lithosphere thickness of 100 km. This is a notable result that draws attention to the importance of shallow layering in GIA models.

Geophysical observables (e.g. surface elevation, gravity field anomalies, seismic data, surface heat flow, etc.) are one of the main sources of information used to make inferences about the interior of the Earth. Obtaining consistent models requires combining simultaneously different observable datasets into joint inversions. Among geophysical data, gravity data from ESA’s GOCE satellite mission provides key information in properly constraining the Earth’s density distribution. WINTERC-G is a new global thermochemical model of the lithosphere and upper mantle (currently being extended into the transition zone and lower mantle) based on terrestrial and satellite gravity data (Fullea et al., 2021). The inversion procedure behind WINTERC-G is twofold. In step 1, 1D inversion of surface wave tomographic, surface elevation (isostasy) and heat flow data is carried out. Then, in step 2, the output model from step 1 is used as prior information for the inversion of GOCE’s gravity field data (geoid anomalies filtered and gravity gradients from GOCE at satellite height) to refine the 3D crustal density and upper mantle composition. The residual, non-isostatic surface topography originated in the model can be regarded as a proxy for dynamic topography although currently there is no explicit dynamic calculation consistently linking both topographies. Within a rigorous framework, residual topography and computed dynamic topography (i.e. solving the Stokes equation) should be consistently integrated into a joint inversion of the gravity field and the terrestrial observations with feedback from both the static and dynamic sides. This is currently missing in WINTERC-G and the goal of this study is to add a third step into the global inversion scheme that consistently integrates dynamic mantle flow, rheology, the deformation of density discontinuities (surface and Core Mantle Boundary, CMB) and the full geoid anomaly (including static and dynamic components). To do that, we will explicitly compute dynamic topography solving the associated Stokes equation fed by WINTERC-G 3D distributions of densities and viscosities obtained from the output of step 2.

Including an explicit computation of mantle flow into an inversion scheme at global scale requires solving a large number of forward large-scale 3D Stokes flow problems. Their high computational cost has been so far a bottleneck for global geophysical inversions accounting for mantle dynamic effects. The proposed methodology will use Model Order Reduction techniques, specifically the Reduced Basis (RB) method, as a tool to alleviate the burden of such expensive computations, by creating surrogate models that approximate the solutions at a much lower computational cost. An efficient implementation of the RB method at a global scale requires a high-fidelity solver as well as a good error estimator. PETSc libraries (https://petsc.org) will be used to build a Stokes high-fidelity and parallel solver which will allow us to create the very fast RB surrogate model essential for the inversion. In doing so, WINTERC-G will be able to simultaneously and consistently explain Earth’s surface elevation (both its isostatic and dynamic contributions) and CMB topography while matching at the same time the constraining satellite gravity data. Moreover, not only we will have a better understanding on how/whether to filter satellite gravity data, but also the feedback from the dynamic/isostatic topography balance will allow us to properly constrain crustal density and lithospheric mantle composition.

Fullea, J. Lebedev, S., Martinec, Z., Celli, N. L. (2021). WINTERC-G: mapping the upper mantle thermochemical heterogeneity from coupled geophysical-petrological inversion of seismic waveforms, heat flow, surface elevation and gravity satellite data, Geophysical Journal International, 226(1), 146–191.

The Earth's core field, a primary driver of the geomagnetic field, is shaped by multiple internal and external sources. Nowadays, information on this complex system is provided by ESA's three-satellite mission Swarm. Near the core-mantle boudary (CMB) there might be large mass transport involved. The satellite mission GRACE-FO (NASA/GFZ) has two intruments dedicated to tiny Earth gravity variations due to global mass transport – a microwave and laser ranging system. These variations might contain also a signal coming from the very deep regions like the CMB. In this contribution the time serious data from the two missions (satellite magnetometry and randing data) is compared to look for common patterns and correlation.

The Earth’s core is made of a solid inner core and a liquid outer core. In the Earth’s outer core, a dominant Coriolis force causes the fluid to form columnar structures aligned with the rotational axis. With a large-scale background magnetic field, a small disturbance may cause oscillations of these fluid columns, which then form torsional waves travelling perpendicular to the rotational axis. Inverse models have shown that the variation of the day length is correlated with the period of torsional oscillations due to the angular momentum transport. Since 1999, monitoring of the Earth’s magnetic field from satellites (e.g. CHAMP and the Swarm constellation) has provided us with a detailed description of the changes in the geomagnetic field. Over more than 20 years of satellite observations are now very much appropriate to the study of torsional oscillations as these waves have periods of about 6 years. An accurate description of torsional waves is therefore essential for the interpretation of geomagnetic data. Many past studies on torsional waves have assumed a simple 1D profile, and did not distinguish the region above and below the inner core. Recent observations and analytical models of the outer core have suggested that this is not necessarily the case. In this study, we take into account the realistic geometry of the outer core, and include the effects of conductivity and equatorial symmetry of the background field as well. We run 3D numerical simulations to quantify how much an initial Gaussian torsional wave pulse can be modified as it propagates, particularly at the tangent cylinder (a special location where the geostrophic cylinder intersects the inner core) where an inward propagating wave splits into two—above and below the inner core. We find that the width of the wave pulse is important in determining how the wave reflects. The wider the pulse, the more effects it feels from the presence of the inner core. The conductivity also plays a role: a conducting inner core generates more reflection compared with an insulating one. The combined effect is likely both geometrical and conductivity-dependent. There can also be a hemispherical difference in the wave speed inside the tangent cylinder if the background field is equatorially asymmetric. This is especially interesting from an observational perspective, since the difference in inverted wave speed can provide an estimate of the field asymmetry, which is otherwise unknown from deep inside the core. Given that the inverted Earth’s torsional wave signal has a relatively wide wavelength with respect to the core radius, and the inner core has a conductivity close to that of the outer core, we need to carefully consider these features when modelling the Earth’s core dynamics.

Greenland’s lithospheric structure underneath the up to 3 km thick ice sheet is still poorly known. Direct observations from the outcropping bedrock geology are limited to the ice-free coast and the seismic station coverage in Greenland is sparse. Models based on different geophysical methods often even lead to contradicting results, as for example deviating Moho depth estimates based on gravity inversion compared to receiver function analysis.
Our study aims to build a lithospheric model for Greenland which focuses on data integration and is consistent with multiple observables. In a self-consistent framework, we use petrological information of the mantle to model coherently seismic velocities, densities, and temperatures down to a depth of 400 km. During the modelling process, we adjust the lithospheric structure to reproduce velocities from the new regional tomography model NAT2021. The forward calculation of the corresponding lithospheric densities provides information on the gravity field and isostatic elevation conforming to the seismic structure. This link allows us to estimate which sources of the observed gravity field are as well represented in the tomography model. In a second step, we jointly invert the residual gravity field from the lithospheric background model with airborne magnetic data to obtain consistent crustal variations of densities and susceptibilities. Furthermore, we compare the modelled geoid and isostatic elevation with the observed values and the surface heat flow output of the model with the few heat flow measurement points in Greenland.
Our approach makes it possible to combine a wide range of datasets. The two-step modelling process allows us to model a higher resolution for the crustal part than for the lithospheric mantle, which is necessary to consider the varying data resolutions.

Surface deformations, e.g. soil compaction, soluble and swelling rock formations, groundwater extraction, natural gas extraction, mining, cavern storage operation and landslides. These processes can cause damages to houses and infrastructrure and can even lead to loss of life. Spaceborne interferometric SAR techniques, e.g. based on Copernicus Sentinel-1, are able to measure surface deformations for entire nations or even continents. In order to support the use of this technique the Ground Motion Service Germany (BBD) is operated, since 2019, by the Federal Institute of Geosciences and Natural Resources (BGR). The service provides mosaicked wide area Sentinel-1 Persistent Scatterer Interferometry (PSI) datasets of the entire nation (~ 360,000 km²). The PSI datasets are based on six ascending- and six descending- Sentinel-1 tracks. Line-of-Sight (LoS) datasets are provided, with an open data policy, via a WebGIS. In addition, since 2021, the service provides nationwide 2D-decomposed datasets (vertical and East-West velocities and time series). All PSI datasets are updated on an annual basis.
The presentation shows the characteristics of the recent update of the nationwide Sentinel-1 PSI datasets. The spatial measurement density and LoS, vertical and East-West velocity precision is reported. Based on selected case studies the information content and limitations of the wide area PSI dataset are discussed. Both, anthropogenic- and natural surface deformation processes are in the focus of the case studies. In order to show the capabilities of the dataset, the case studies include spatial scales from a few hundred of meters to more than 50 kilometers. Field observations, in-situ measurements and thematic data are used to assess the information content of the wide area PSI dataset. Finally, the presentation gives an overview of the BBD WebGIS, used for dissemination and visualization of the big PSI datasets. The WebGIS functionalities include, e.g. time series plots, queries, download, print and superposition with ancillary data.

Measurements of the gravity field can be used to estimate the density distribution inside the Earth at many different scales. Satellite gravty data are well suited for investigations at a continental scale and are sensitive to both crust and mantle density variations, although the resolution becomes worse with depth. However, such inversions are inherently non-unique, so additional constraints are needed. Most often these constraints come from seismic tomography models.

While tomographic models provide crucial information, there are some problems when connecting them to the gravity field. One problem is that tomographic models are affected by both lateral and vertical smoothing, which might also be different depending on the distribution of seismic stations. Furthermore, how to relate seismic velocities and density is the subject of an ongoing debate, especially in the presence of compositional variations.

To help alleviate these issues, we can use the Rayleigh wave phase velocity instead of the seismic tomography as constraints directly. The Rayleigh wave phase velocity at a geographical location depends only on the vertical distribution of seismic velocities and (weakly) density. Each Rayleigh wave frequency has distinct sensitivity to a specific depth range, with lower frequencies being more sensitive to deeper structure. Thus, the Rayleigh wave dispersion curve, which gives the relation between frequency and phase velocity, is - to some extent - an equivalent representation of the vertical velocity structure. The benefit of the dispersion curves is that they do not fix velocity anomalies at specific depths, so in a joint inversion of gravity and dispersion curves the different depth sensitivities can be exploited more fully. Finally, Rayleigh wave phase velocity is weakly sensitive to density as well, so some insights might be gained into the velocity-density relation.

We demonstrate the idea of this joint inversion in a transdimensional framework, where the Earth's mantle is parametrized in terms of a variable number of discrete anomalous volumes. Synthetic dispersion curves are calculated for recent seismic tomographic models and jointly inverted with synthetic gravity data simulated at 225 km height.

We present an approach to estimate local time series of the magnetic field gradient tensor elements at satellite altitude. These time series are derived from magnetic observations taken by the Swarm satellite constellation, and are computed using the Geomagnetic Virtual Observatory (GVO) method. The GVO time series are distributed in a worldwide network, which enables global investigations of spatial-temporal variations in the gradient tensor elements. We find that the Swarm constellation is beneficial for deriving the GVO gradient series compared with a single satellite mission. Our results support previous findings that gradient series could help to better resolve SV for harmonics above degree 6. We find evidence for a regional SV impulse (jerk) event in 2017 in the first time derivative of the gradient tensor elements. This event is located at low latitudes in the Pacific region, and its signature is independent of adopted data selection criteria. Striking patterns of localized
acceleration changes in global maps of the gradient tensor elements, are found to be associated with this event. The GVO gradient tensor series provides an effective way of compressing spatio-temporal information of geomagnetic field variations as gathered by the Swarm satellite mission. Such gradient tensor series may prove useful for studying geomagnetic jerks, core flows and in geodynamo data assimilation.

A sharp change in the trend of the secular variation of magnetic declination measured in Europe during 1969-1970 was the first identified instance of geomagnetic impulse (or jerk).  This 1969 geomagnetic impulse has effectively set the standard against which other such events are compared.   Recent studies of geomagnetic impulses during the era of global satellite observations, and in numerical geodynamo simulations, indicate that such geomagnetic impulse events are often associated with dipolar foci at the core-mantle boundary, where the radial field acceleration locally changes sign on an interannual timescale,  and that these foci typically occur at low latitudes.  This raises the question as to whether the 1969 event was fundamentally different from more recent events, and whether or not it is compatible with the hydromagnetic-wave driven mechanism seen in geodynamo simulations. 

In an effort to address these questions we here revisit the 1969 impulse, reprocessing the available ground observatory data and using measurements of field intensity, and along-track differences of field intensity, from the OGO-2, 4 and 6 missions to derive a geomagnetic field model spanning 1960-1980 using a field modelling technique similar to that employed for the CHAOS field model series (Olsen et al., 2006; Finlay et al., 2020).  We find that the 1969 impulse is closely associated with an intense change in the core-mantle boundary radial field secular acceleration with a dipolar structure, located under central and Northern America in 1969-1970.  This has a similar amplitude to recently observed secular acceleration changes and is longitudinally localised. However it extends meridionally up to mid-to-high latitudes under north America.  We find evidence for this feature in models build only using POGO satellite data  but it is most clearly seen in the ground observatory data, particularly from San Juan (vertical component) and Fredericksburg (northward component).  The famous V shaped features in the East-west component of secular variation in Europe also largely resulted from these changes taking place below North America, a consequence of how the Green's function for the east-west component of observatories in Europe samples the core-mantle boundary .  Based on our findings regarding its signature at the core-mantle boundary, it seems the 1969 impulse may be compatible with a mechanism similar to that which produced impulse events in recent years.  The idea that the 1969 impulse is fundamentally different to more modern events may in part be related to the uneven distribution of the available ground observatories, and due to the fact that in some prominent records it appears unusually well separated from other impulses.

Modern geomagnetic field models, derived using satellite and observatory magnetic data, are now covering more than twenty years. They provide information on the numerous external and internal sources generating the field and in particular on the core field. The presented work aims at providing reliable models that describe the short-term evolution of the core magnetic field and of the core surface flow. We use a sequential modelling approach – a Kalman filter, combined with a correlation-based modelling step, together with strong prior information on the spatial behaviour of the geomagnetic field and core flow contributions. A sequence of snapshot models has been built describing the core magnetic field and its secular variation together with their associated flow at the core-mantle boundary, allowing for a core field temporal resolution of the order of the year. Prior information on the flow and magnetic field models rely on the statistics of the Coupled Earth numerical dynamo model (Aubert et al., 2013). The resulting time series of core field and flow models present the general characteristics of the models based on more classic modelling techniques, thus supporting the reliability of this method. The flow models present strong quasi-geostrophic structure. New interesting features were found in both the core field and flow time series, especially at small spatial scales. The results obtained depend on the quality of the data inputs and particularly on the availability of satellite data. The model presented were derived from Oersted, Champ, Cryosat and Swarm satellite data. For recent years the modelling process rely exclusively on observatory and Swarm data, the latter being absolutely necessary to describe precisely the fast evolution of the core field. Recovering faster variations of the core field is a high priority research activity, that may be possible thanks to new generation of satellite constellations such as NanoMagSat.

The interaction of the oceanic tidal flows with the Earth's main magnetic field provides a powerful natural source of electromagnetic energy suitable for sub-oceanic upper-mantle electrical conductivity sounding. We have developed a new frequency-domain, spherical harmonic-finite element approach to the inverse problem of global electromagnetic (EM) induction. It is set up for an effective inversion of satellite-observed tidally-induced magnetic field (TIMF) in terms of three-dimensional (3-D) structure of the electrical conductivity in the sub-oceanic upper mantle.

In this contribution, two independent TIMF datasets, both derived from the continuous measurements of the geomagnetic field by the Swarm multi-satellite mission, are employed (Grayver & Olsen, 2019, and Sabaka et al., 2020). Using several regularization methods and a-priori conductivity models, we have obtained the 3-D conductivity structures below the global ocean in the 150-400 km depth range. Figure 1
shows the electrical conductivity in S/m at the depth of 230 km, obtained by the inversion of M2 TIMF from the first dataset.

In the next step, the recovered 3-D conductivity structures are locally compared with the 1-D conductivity profiles below selected coastal and island geomagnetic observatories. We also present a statistical verification of the recovered conductivity models using thermodynamically self-consistent compositional models coupled to the electrical-conductivity laboratory measurement of individual mantle constituents. Bayesian method based on the sampling of the laboratory conductivities and thermochemical parameters allows us to perform an error analysis.

Finally, we explore prospective benefits of improved spatio-temporal coverage of the planned NanoMagSat mission for TIMF recovery. For this purpose, we use the forward modelling of TIMF along synthetic NanoMagSat trajectories.

We present results of cluster analysis and geophysical modelling of the West and Central African rift system, where we integrate seismological and satellite data. For a description of lithospheric domains, two different methods based on seismic tomography and satellite gravity data have been used. First, the terracing method using the shape index has been applied to the gravity field (XGM2019) in order to enhance the signal of the large-scale tectonic units. In addition, the K-means cluster method (which is an unsupervised machine learning algorithm) has been applied to a tomography model over the study area.
Both models are compared and interpreted towards similarities and differences. The preliminary analysis based on K-means clustering of seismic tomography shows that the West and Central African rift system and its surroundings can be divided into at least three clear and distinct tectonic domains: The Northern part of the Congo craton, the Eastern part of the West African craton and a large area going from Cameroon’s littoral (Atlantic Ocean) to the Egypt’s Northern border (Red Sea). In addition, the preliminary analysis of the terracing of satellite gravity data, first confirms the localization of both, the Congo craton and the West African craton, but also splits the large area from Cameroon to Egypt into two known tectonic units: the Southern part of the Saharan meta-craton and the West and Central African rift system in the middle of the three others units.
The cluster analysis is also pointing to differences at crustal and upper mantle level and is the first step towards the evolution of a lithospheric scale model. In the model, we integrate our tectonic domain analysis with the existing seismic Moho depths estimate and other information. Magnetic data will be considered in addition to study link of the lithospheric architecture to the large and enigmatic Bangui anomaly.

Detailed study of the time-dependent internal magnetic field produced in the Earth’s core requires that one can appropriately describe and separate external magnetic fields generated by the electric current systems in the Earth’s ionosphere and magnetosphere. In particular, the magnetic field associated with the strong currents in the polar ionosphere, driven by the interaction of the solar wind and the magnetosphere, may contaminate core field models if not properly taken into account.
Earlier studies have developed several techniques to include the ionospheric field in geomagnetic field modelling. However, these approaches have only been able to account for specific periodicities in the field at non-polar latitudes or have used relatively little information on the underlying physical processes.
Here, we describe an approach for representing the ionospheric field during geomagnetic field modelling with a focus on the challenging polar regions. Magnetic apex coordinates are employed and the time-dependence of observed solar driving parameters is exploited. We include the new ionospheric field parameterization in the CHAOS framework for modelling the geomagnetic field and derive test models from satellite magnetic data to investigate the impact on the core field and to study the polar ionospheric current system under geomagnetically quiet conditions.
We find that the misfit in the polar regions is reduced, indicating success in accounting for previously unmodelled ionospheric signals. An ambiguity is however found in the zonal parts of the ionospheric and the internal field models. But this can be managed through model regularization. We also find that the divergence-free part of the horizontal ionospheric currents is weak compared to models of the ionospheric currents derived under more disturbed geomagnetic conditions than used here for the field modelling. We find that co-estimating the ionospheric field can results in an improvement in the high-degree and low-order components of the core field Secular Variation at the core-mantle boundary in the polar regions.
Our approach may allow us to relax the strong temporal smoothness constraints usually applied to the core field during geomagnetic field modelling and, thus, help to better resolve its time changes.

We present a post-obductional geodynamic model explaining the differentially elevated ~150-km-long passive margin of northeastern Oman based on multi-proxy analyses. The combination of methods such as multi-criteria analysis of the landforms, through the post-processing of very high resolution CartoSat-1 DSM, field work-samplings and laboratory XRF dating’s, together with semi-automated geomorphological mapping, differential GNSS measurements, and SAR interferometry (InSAR), are part of the proposed investigation. The results help understanding the geomorphological evolution of the Late Cretaceous to Quaternary relict landscape, where landforms are presented as indicators of the evolution of the area from Saih Hatat Dome (SAD) to Sur and define the processes that formed the current relief of northeastern coasts of Oman. At least six tectonic terraces were determined in shelffuls carbonate rock formations of Eocene age. This passive margin exhibits unusually high late Pleistocene uplift rates at an average of ~2.2 mm/a. InSAR-derived ground displacement documents that the uplift is still ongoing with an average rate of +1.3 mm/a. Geomorphological and stratigraphical analyses show that the area was heterogeneously uplifted following the Late Cretaceous. Thus, a fault-based tectonic model of differential and compartmentalized (“block-like”) uplift history of the studied passive margin of Oman is proposed.

Vast regions of North America were covered by ice during the Last Glacial Maximum (about 20,000 years ago) depressing the Earth. After the ice melted, the lithosphere started rebounding and this still ongoing uplift is measured, among others, by Global Navigation Satellite System (GNSS) observations and geological sea level indicators. While the vertical velocity (uplift) component is often used to constrain dedicated models of glacial isostatic adjustment (GIA), the horizontal motions as observed by GNSS have received little attention in such investigations in the past. However, horizontal motions are more sensitive to lateral lithospheric structures and viscosity variations in the mantle, which are known to vary significantly within the Earth

To realistically capture the correct horizontal motion due to GIA, numerical models need to be more complex by including lateral heterogeneity of the mantle’s viscosity structure, lithosphere variations and, most importantly, compressibility. Compressibility is shown to affect the results significantly. For example, a ten-thousand-year loading with a polar cap of 10 latitudinal degrees in extent and maximum 1500 meter in height, results in a horizontal displacement that in compressible models is about 20 meters lower as compared to that in incompressible models. This effect is significant as the maximum horizontal deflection in this scenario is only 40 meters. Using an improved Finite-Element model, called FEMIBSF, we can model the deformation of a compressible Earth due to glaciations, while also incorporating a laterally varying viscosity structure. Lateral variations in viscosity are calculated from 3D temperature fields of WINTERC-G, an upper mantle model that uses surface waveform tomography, surface heat flow, satellite gravity and topography to derive mantle temperature, among others. The temperature fields are converted to viscosities using an olivine flow law, varying the grain size and water content in order to create multiple options for the 3D viscosity fields.

In this study, we focus on the effect of lateral heterogeneities on the modelled horizontal displacements. Finally, we attempt to match modelled horizontal displacements with those observed by GNSS to constrain the mantle viscosity.

In times of global warming, the understanding of key processes for ice sheet melting and sea level change is of high relevance. The Antarctic ice sheet alone could potentially contribute about 60 m of sea-level rise. Glacial isostatic adjustment (GIA) – the viscoelastic response of the solid Earth to surface loading – is a key feedback mechanism for bedrock deformation, sea level, grounding-line position, ice sheet elevation and stability. Observational data reveal large viscosity contrasts in Antarctica characterized by an old, cold craton with a thick lithosphere and high viscosities in the east and a failed rift system with thin lithosphere and low viscosities in the west divided by the Transantarctic Mountains. The response time and bedrock topography evolution in GIA models are sensitive to the Earth structure and glacial history. Therefore, the consideration of independent data for the determination of the 3D Earth structure helps to improve GIA models. Hereby, the lithospheric structure, which controls the bending stiffness, and the mantle viscosity, which determines the timing of the adjustment process, are parameters of great interest. But in detail, the Earth structure for Antarctica is not uniquely defined and differs depending on the employed method, e.g., seismic tomography models, geodynamic models or transfer functions.

We develop 3D Earth structures under consideration of independent constraints and apply them to GIA models to investigate their influence on bedrock deformation (e.g., uplift rates) and sea level predictions. In a first step, we calculated a global 3D viscosity structure derived from seismic tomography models and constrained by the geoid, heat flux, the Haskell average and mineral physics. In a second step, we created a suite of 3D Earth structures as well as regionally adapted 1D Earth structures. All 3D structures are derived from the same seismic tomography models, but differ in the transfer functions from seismic velocity to viscosity. The ensemble is based on a combination of different reduction factors in the Arrhenius law and different viscosity contrasts between upper mantle and transition zone. In a third step, we are developing an improved parameterization for the Earth structure underneath Antarctica that combines the global geodynamically constrained Earth structure with a new Antarctic lithosphere model. This model provides lithospheric as well as crustal thickness, variations of temperature and composition, and the geothermal heat flow, and is based on an iterative combination of gravity and seismic tomography data.

The analysis of the suite of 3D Earth structures and GIA model output and a detailed view on the relationship between the different parameters of transfer functions and corresponding sea level predictions manifest the significant influence of lateral Earth structure variation on relative sea level predictions. Furthermore, the 1D models cannot reproduce the sea level predictions of the 3D models due to the strong lateral viscosity contrast. For Antarctica, the averaged 1D structure underestimates the viscosities in the east and overestimates them in the west. The influence of the improved parameterization for the 3D Earth structure underneath Antarctica will be investigated. In future studies, we aim to apply the obtained structures to coupled ice sheet - solid Earth models to also simulate the feedback of the Earth structure on ice sheet stability.

This study benefits from the German Climate modeling initiative PalMod and the SCAR subcommittee INSTANT-EIS, Earth – Ice – Sea level within the Scientific Research Programme INStabilities and Thresholds in ANTarctica and contributes to the GRACE-FO project of the ESA Third Party Mission Program.

In active subduction zones, the spatial and temporal variations of the processes accommodating the relative convergence of the tectonic plates are an active research topic. Over the last decade, observations from geodesy and seismology have provided new understanding about transient deformation phenomena, like Slow Slip Events (SSEs), and their relationship with seismic activity. Aseismic transient slips can have a large range of amplitude and duration (from sub-daily periods up to a few decades). However, this range is limited by the duration of geodetic observation and one can question the representativeness of the observed phenomena over several seismic cycles. To overcome those problems, homogeneous observations covering long portion of subduction zones can help to document different stages of longer processes or aseismic slip events with long recurrence time. It can also help to study the influence of properties that are variable in space along the subduction but stable at the time scale of several seismic cycles (like the geometry of the interface, the topography, the plate velocity, the geological and thermal structures of the subduction, etc.).
Here, we focus on the 1000-km long Mexican subduction zone by combining GNSS data with satellite SAR interferometry (InSAR) data that have the potential to measure consistently ground deformation at continental scale with a high spatial resolution (tens to hundreds of meters) and a temporal sampling of a few days or weeks. The Mexican subduction offers the possibility to observe by geodesy a large variety of slip processes from earthquakes, post-seismic afterslips and SSEs of different sizes and duration. The occurrence during the last 6 years of several subduction earthquakes with Mw larger than 7 associated with post-seismic displacement and the occurrence of large slow slip events with equivalent Magnitude larger than 7 makes the Mexican area a good target for InSAR analysis. However, the environmental conditions of the Mexican Pacific coastal range with dense vegetation, large topographic and atmospheric variations, are posing methodological challenges for InSAR. To face them, we use Sentinel-1 SAR data from 2016 to 2019 and propose two different approaches to separate the atmospheric component from the tectonic signals. The first one (parametric method) consists of a least squares linear inversion, imposing a functional form for each deformation signals and the atmospheric signal. The functional form of the atmospheric signal is constrained from GPS Tropospheric Zenital delay observations. The second methods uses independent component analysis of the InSAR time series.
The resulting geodetic time series of displacements are used to invert slip distribution at the subduction interface, in order to characterize the 2017-2018 SSE and to obtain a consistent coupling map covering a large part of the Mexican Pacific coast from Jalisco to Oaxaca. We demonstrate the capability of Sentinel-1 InSAR to fill the spatial gap between GPS stations in the challenging Mexican environment, which improves the coupling estimation where the stations are sparse. We show that The Jalisco and Michoacan area (longitude from -106◦E to -104◦E) has a relatively high and continuous coupling ratio up to 0.8. The spatial variation of coupling are consistent with the location of the 23 june 2020 Mw 7.1 Oaxaca and 9 September 2021 Mw 7.1 Acapulco subduction earthquakes. The effect of SSEs of a few months duration on the 3-year coupling estimate in the Guerrero and Oaxaca regions illustrates that the possibility of transient slip events occurring during or outside the period of observation should be considered (or at least not ruled out) when interpreting coupling map, for instance to infer cumulative slip deficit over hundreds of years.

The development of new technologies contributes to increasing the resilience of cities and infrastructures in highly vulnerable countries against natural and induced geological hazards. This work is based on the contribution of DInSAR (Differential Interferometric Synthetic Aperture Radar) technology addressing the seismic hazard in San José (Costa Rica) as a collaboration project between the startup Detektia Earth Surface Monitoring. SL and the Universidad Politécnica de Madrid, in the frame of the Kuk Ahpán project.
DInSAR is a technique based on the processing and analysis of a long series of Synthetic Aperture Radar (RADAR) images. This technology provides records (since 1992) and updated movements on any surface anywhere in the world without the need for ground instrumentation, with velocity accuracies of around 1 mm/year. In this context, satellite radar provides valuable information over very large areas that complements field work and on-site instrumentation. This work explores the contribution of this technology to manage seismic risk in highly populated metropolitan areas, including the local effect of faults. As a case study, this seismic risk assessment (including natural and anthropogenic hazards) will be applied in San José. This study area has been selected for its tectonics and the number of seismic structures in addition to the great data gap that exists. In the field of investigations of the deformations of the crust with GNSS techniques, there is a lack of data in the Central America area. It is worth mentioning the study carried out with geodetic, structural and paleomagnetic data, combined with GPS data from Guatemala, El Salvador, and southern Honduras, which shows the speeds of about 15 mm / year (Alvarado et al., 2011) of the antearcs of Nicaragua and El Salvador moving in the west-northwest direction referred to the Caribbean Plate. Apart from this work, no other studies have been found in the Gulf of Fonseca or in the area of the Chortís block in the Atlantic. The Kuk Ahpán project will cover this knowledge gap by complementing and validating the results of the measurements published in 2011. (Benito et al. 2019). The advantages that multitemporal DInSAR technology offers compared to traditional methods are: (i) current and historical millimeter precision, (ii) constant updates, (iii) no need for ground instrumentation and (iv) a high density of points of control. Currently, the control of infrastructures is carried out with traditional geotechnical and geodetic methods (strips, inclinometers, accelerometers, GNSS networks ...). This proposal will develop highly detailed monitoring systems in Central America at a very competitive economic cost for those facilities and infrastructures that do not have a control system. Likewise, it will allow the monitoring of those in which traditional methods are already installed. The methodological approach starts by integrating DInSAR data with geophysical and geodesical data such as bathymetry, geomagnetism, gravimetry, seismic profiles, GNSS series…The seismic models used provide a range of results: acceleration maps on the ground, coulomb forces, ..., the InSAR information could be used not only to contrast and validate the data, but also to contemplate the scenarios on a smaller scale and assess infrastructures or fragile areas, once the scenarios have been modeled and the place and intensity of the events are known. A long-term objective is proposed the design and establishment of a monitoring system that serves to provide early data in the management of events on populated areas and other affected infrastructures. The study will start with a performance of the main sismologica scenarios that can be found in the main populated areas of Central America. The models used in this section are obtained from a bibliographic review (Abrahamson, N.A et al, 2013), (Campbell, K.W. et al 2014), (Boore, D.M. et al 2014). Project researchers specialized in each study area contribute their knowledge throughout the project. Therefore, the models used can be updated in order to achieve more truthful results. The rupture models allow, knowing the geometry of the seismic structures (faults) and processing the spatial data in a GIS environment, to obtain a raster layer that shows the PGA (Peak Ground Acceleration) for a maximum magnitude earthquake (according to empirical cases). The resulting layer shows the acceleration over rock. Then, basing the decision on expert opinion, a weighting is performed on the different models to combine them and obtain a single result for failure. In order to finally obtain the PGA maps on the surface, information from different field campaigns will be considered: lithological maps, soil deep cartography and Vs30 registers. In a second stage, DInSAR technology is used to validate the data resulting from the previous phase. The following activities are proposed: Integration of available auscultation data (especially GPS / GNSS ), harmonization and refinement of the GPS / GNSS data series, obtaining the SAR image catalog corresponding to the available data series using current sensors (i.e. shipped in Sentinel 1 A / B-see table 1) or historical ones (ERS 1/2, ENVISAT, DInSAR processing of SAR images using technology developed by Detektia, validation of the displacements observed by the DInSAR technique with the auscultation database and estimation of the uncertainty of the series of DInSAR. The time series of the cartography used to detect movements, of a radar and optical nature, will be processed using Detektia technology (computers and codes). It will have a validation process with field data whose information will be obtained from the available GNSS / GPS networks and the new stations that the Kuk Ahpán project will install (network of 10 GPS stations on the Honduras edge of the fault) that they continuously monitor positions during the project's exercise time. A first SBAS (Small Baseline Subset) processing of the data will serve to get first results on a general view and a validation of the results on the previous stage. The idea is to update details for the fault mapping, coulomb stress maps or mechanism maps and other material used in the development of analysis in the field of seismic engineering. As a fact, the data gap for this study area will be filled. In addition, and on a building / infrastructure scale, a PSI (Persistent Scatterer Interferometry) processing will be used, to obtain detail in the deformations of the same. A long-term objective is proposed the design and establishment of a monitoring system that serves to provide early data in the management of events on populated areas and other affected infrastructures.

Savannas cover a large percent of the area of Africa. The savanna biome is a highly heterogenous and complex as it interacts with climate, topography, soils, geomorphology, herbivory and fires, with cover variations between open and closed vegetation cover.

Accurate vegetation cover products are essential for natural resources management and assessments, fire dynamics, habitat quality assessments, and climate models and in conservation at various spatial scales. However, current estimations for vegetation cover mainly use discrete mapping approaches for vegetation cover classes. These mapping approaches tend to be inadequate to capture the savanna heterogeneity.

Mixed pixel analysis quantifies the proportions of cover per pixel to establish pure tree, pure bush, pure grass and non-vegetation cover. Current global mixed pixel data are based on medium and low spatial remote sensing data such as Landsat, AVHRR and MODIS which lead to saturation of optical signal, phenological noise and confusion of trees with the herbaceous layer. This poster presents insights on the use of high spatial and temporal resolution remote sensing data to compute spatio-temporal dynamics of the savanna cover mentioned above using a mixed pixel analysis approach. Sentinel-1 and Sentinel-2 multi-temporal time series are utilized to analyse the dynamics of grass and woody vegetation covers at a pixel level in the savanna biome of Benfontein Nature Reserve and Mokala National Park in South Africa. These time series comprise of multi-temporal statistics such as the mean, median and standard deviation of VH and VV backscatter, and coherence from Sentinel-1 and indices such as Normalised Difference Vegetation Indices (NDVI), Soil-adjusted Vegetation Index (SAVI) and others from Sentinel-2. The time series are analysed at individual pixel level to visualise the time series of homogenous pixels (e.g 100% cover of shrub, grass and trees) and those pixels with more than one cover (e.g 30% shrub, 10% trees and 60% grass cover) to understand the spatio-temporal dynamics of those pixel which are mixed.
Keywords: mixed pixel, Sentinel-1, Sentinel-2, time series, savanna

Nearly half of the Earth’s land surface is classified as rangeland. The health and productivity of this land is essential for many rural communities, especially for those whose livelihoods depends directly on pastures such as pastoralists who graze their livestock on rangelands for food and income. However, in regions such as Central Asia, unsustainable management practices exacerbated by the impacts of climate change are increasing the degradation of rangelands, threatening the livelihoods of many rural people. Identifying where the most degraded areas are and implementing restoration and rehabilitation measures to invert the deterioration dynamic is crucial. Earth Observations (EOs) plays a key role in this process, however, to date, few efforts have effectively assessed the rangeland conditions. This can result in poorly planned investments, particularly in pastoralist areas, that may not strengthen the resilience of communities.

To respond to the need of monitoring and assessing rangelands condition, the Climate Resilience cluster of the ESA’s Earth Observation for Sustainable Development initiative developed a novel method to map rangeland condition changes and identify restoration opportunities. The cluster applied the methodology in two investment projects funded by the International Fund for Agricultural Development (IFAD), in Kyrgyzstan and Tajikistan. In these two cases the cluster analysed the changes occurred in the last two decades using data acquired from Landsat-5 (1984-2013), -7 (1999-present) and -8 (2013-present) and has integrated the local knowledge of the pastoral communities. Landsat imagery were considered for the historical analysis, but data from Copernicus Sentinel-2 mission, offering more detailed spatial and spectral resolutions, are expected to be used in future updates of the use cases.

The methodology applied relies on the computation of different well-established spectral indices as a proxy of the vegetation temporal evolution. All these indices are combined avoiding redundancy by excluding auto-correlated indices and weighted according to their significance when assessing status results against in-situ pasture measurements. The resulting rangeland dynamics are translated into qualitative classes following the guidelines of the Intergovernmental Panel on Climate Change (IPCC) for grasslands degradation. According to previous analyses, the majority of pastures in Central Asia correspond to grasslands. Key for the success of the study is the integration of the local grazing practices defined by the (i) specific grazing periods, (ii) seasonal-based altitudinal ranges and (iii) distance of pasturelands to villages for each administrative area. To ingest this information into the methodology the Shuttle Radar Terrain Mission Digital Elevation Model (STRM-DEM) at 30 m was used.

The results show a generalized degradation of rangelands in Kyrgyzstan and Tajikistan, which aligns with the local observations and measurements. This EO-based assessment of rangelands health informed IFAD project formulation, provided data to assess impacts of a closing project, informed pasture management maps at municipality level, and informed the update of climate commitments (reflected in the Nationally Distributed Contributions) of the Kyrgyz’s Government.

Grasslands cover about one third of the global ice-free land surface. In addition to food production, grasslands are also responsible for regulating services such as carbon sequestration and water storage and they embed a rich biodiversity. In land cover maps, natural or intensively cultivated permanent grasslands and other open vegetated landscapes are often embedded in a broad land cover class. However, all grasslands do not provide the same ecosystem services in the perspective of environmental and climate impact, biodiversity and habitat quality.
In a perspective of biodiversity, in the LifeWatch ERIC land cover layer (LW-LC) supporting the ecotope database, grasslands were classified in three classes, namely “monospecific grassland with graminoids”, “diversified grassland and shrubland” and “shrub or herbaceous (flooded)”. This classification was obtained semi-automatically, based on ortho-images, LIDAR data and Sentinel-2 time series and was consolidated through expert field observation and photo-interpretation.
The objective of this study is to produce a fully automated and thematically improved grassland map combining a management-based approach and an ecological approach. Both complementary approaches have been developed and demonstrated over Wallonia, the Southern region of Belgium, using Sentinel-1 and 2 time series and comprehensive in situ observations.
Research has repeatedly shown the impacts of grassland management intensity on their ecological value as habitats. The first approach therefore consists in characterising grasslands based on a combination of different factors of grassland management intensity. More specifically, grassland parcels are characterized (i) in terms of frequency and precocity of mowing events and (ii) in terms of fertilisation level. The frequency and precocity of mowing events are obtained using the Sen4CAP grassland mowing detection method based on Sentinel-1 and Sentinel-2 time series. This method reached a detection rate of 85% and a precision of 73% on hay meadows monitored during a field campaign in Wallonia (n=426). The fertilisation level can be estimated through the nitrogen nutrition index (NNI). The NNI indicates the level of nitrogen saturation in a vegetation cover. It is the ratio between the measured or estimated canopy nitrogen content (CNC) and the critical CNC (cCNC), which depends on the vegetation type and the above ground dry biomass (DM). During four growing seasons (2016, 2017, 2018 and 2020) field measurements of N content and DM were performed on grassland parcels across Wallonia. Our first results show that grasslands can be classified in two classes of NNI value (higher or lower than 0.6), using a random forest classifier and Sentinel-2 bands (4, 5, 6, 7, 8 and 8A) with an overall accuracy of 81%. This suggests that the fertilisation level of grasslands can be retrieved automatically.
The second approach aims at directly classifying grasslands in terms of ecological value, using the LW-LC layer as baseline. Several combinations of features derived from Sentinel-1 and Sentinel-2 time series are tested and compared. The expert-based detailed grassland map provided by the Département de l'Étude du milieu naturel et Agricole (DEMNA) is used in addition to the LW-LC layer to build a reference dataset. In this detailed map, grassland parcels are characterized according to the biotope typology of WalEUNIS , based on observed vegetation species.
Both approaches are eventually combined to provide a more detailed characterization of grasslands. The thematically detailed grassland map will be implemented in the ecotope database of LifeWatch to support biodiversity assessment and habitat monitoring. Furthermore, the developed approaches could be applied at a broader scale to map temperate meadows and pastures across Europe.

Understanding and monitoring ecological and socio-ecological dynamics in African rangelands

Background
African rangelands contribute to the livelihoods of hundreds of millions of pastoralists over large parts of Africa, providing livestock‐based food (meat, milk) and income from livestock sales. Rangelands also support livelihoods through other resources, including edible and medicinal wild plants and animals, firewood, building material (wood, thatch) and drinking water. However, African rangelands also supply additional ecosystem services that do not only benefit local communities, but also the wider African and global population. A prime example is the rich wildlife in protected rangelands that attracts millions of international visitors each year in a multibillion‐ dollar wildlife tourism industry. In addition to physical resources and revenue, African rangelands play a pivotal role in global carbon cycling and thus climate change mitigation and harbour unique bio‐ and cultural diversity. The capacity of African rangelands to sustain these functions over the coming decades will be affected by increasing anthropogenic pressures, including climate change, land conversion and degradation, pandemics and biodiversity losses. Effective adaptation to (and mitigation of) anthropogenic and climatic driven changes requires timely, relevant and accessible information on the extent, performance and resilience of rangeland ecosystems and resources. Here we propose such a monitoring system.

Aim
A functional monitoring system should recognise and address the complexities of “rangeland” as an umbrella term. First, rangelands support a wide variety of land‐uses, indicating a variety of demands on a monitoring system from different users and stakeholders. Second, rangelands span wide environmental and ecological gradients, indicating that a monitoring system should define variables that translate to useful products across the full breadth of these gradients. The primary gradients across African rangelands include [i] the annual total and temporal trajectories of primary productivity as determined by moisture availability, soil nutrient status and [ii] the ratio of herbaceous (grass) to woody plant biomass, which is strongly determined by the fire regime, herbivore pressure and human wood harvesting. Social‐ecological processes along these gradients determine the spatiotemporal availability, quantity, quality and resilience of rangeland resources, especially grass biomass that sustains grazing livestock and wildlife.

Objective
In the Rangeland monitoring for Africa using Earth Observation ‐ Continental Demonstrator (RAMONA) we will therefore identify rangeland archetypes based on social-ecological characteristics and dynamics and determine their spatial extent. Within each archetype we will quantify seasonal and annual dynamics of herbaceous productivity, the key resource in rangelands, and social-ecological performance reflecting trade-offs between rangeland productivity (e.g. livestock biomass) and rangeland integrity (e.g. climate change mitigation and biodiversity capacity). Since many African rangelands are highly heterogeneous at small spatial scales, we will perform all analyses at a high (10 m) spatial resolution.

Approach
Sentinel-1 and -2 will deliver the primary data, supplemented by Sentinel-3 for gap-filling where necessary, and historical EO data to determine baselines. Mapping of rangeland type and extent will also employ external land cover data. Herbaceous biomass productivity will be estimated using absorbed photosynthetically active radiation (APAR) and vegetation-type-specific light use efficiencies, which will refine current estimates using eddy covariance data. Rangeland and biomass phenology will be captured using TIMESAT (Eklundh & Jönsson). Rangeland performance will be calculated from available data on livestock and wildlife distribution and production, along with essential biodiversity and climate variables.

Results
Here we will present early results, detail expected outcomes and outline key stakeholders and use cases.


Eklundh, L. and P. Jönsson, 2015. TIMESAT: A Software Package for Time-Series Processing and Assessment of Vegetation Dynamics, in Remote Sensing Time Series, C. Kuenzer, et al., Editors.

The grasslands play a crucial role in global biogeochemical cycling, serve as an important habitat for animals and support livelihoods (Baumann et al., 2020; Squires et al., 2021). Over the 19th and 20th centuries, grassland biomes, including the Eurasian steppes, underwent drastic land‐cover changes, including conversion to croplands (Prishchepov et al., 2020, p. 700). Following the institutional transformation, for instance, the transition from state-command to the market-driven economy after the breakup of the Soviet Union and other socialistic satellites, land-use pressure on grassland ecosystems predominantly declined in Central and Eastern Europe, while grassland use intensification also occurred in some parts of Kazakhstan and Mongolia, rangelands of Iran (Dara et al., 2020; Mazloum et al., 2021). Yet, little is known about the state of grasslands, for instance, in the steppe biome of Europe, with a bulk of steppe grasslands in Russia and Ukraine. Satellite remote sensing provides an unprecedented source to monitor the state of the land cover land use, particularly with the advent of ESA’s Copernicus Sentinel monitoring program- 30-m SAR Sentinel-1 and optical 10-20-m Sentinel-2 satellites, and therefore, extending the availability of optical 30-m Landsat observations dating back to 1980s. However, the spatial resolution of Sentinel-1, Sentinel-2 and Landsat satellite data is insufficient to monitor detailed steppes changes, such as informal roads, oil and gas development, shrub encroachment, military use, etc. Therefore, very-high-resolution data like WorldView, IKONOS, and Planetscope constellations, ultimately, provide a valuable source of data to map detailed changes in steppes. Our major goal was to document the grassland dynamic in the dry steppe belt of European Russia and part of Eastern Ukraine from 1990 to 2018 with Landsat and Sentinel-1 and Sentinel-2 imagery. For that, we constructed seasonal cloud-free composites (Spring, Summer, Fall) for circa 1990 and 2018 in Google Earth Engine (GEE) (Pazur et al., 2021; Prishchepov et al., 2021). We used monthly Sentinel-2 composites to map mowing and tested SAR Sentinel-1 VV and VH polarization. By bringing the case study of Eastern Ukraine and Samara and Orenburg province of Russia, we revealed other disturbances in steppes, namely, oil and gas development, formal and informal roads, garbage dumps, and military land use (Eastern Ukraine). We tested a range of image classification techniques, such as the classification of land-cover change with a random forest classifier. We also tested U-Net Convolutional Neural Network architecture as well as crowdsource participatory approaches to map detailed disturbances in the steppes. Training and validation data have been collected with the use of very-high-resolution satellite imagery available via Google Earth, Planetscope constellations, dense Landsat, Sentinel-2 composites. Work has also been complemented by validation data collection during the field campaigns in 2018-2019. Our study showed the recovery of steppe grasslands at the expense of abandoned croplands from 1990 to 2018 particularly in Orenburg, Samara, Saratov, parts of Rostov provinces of Russia, and in insurgent parts Donetsk and Luhansk provinces in Eastern Ukraine. The results of broad-scale analysis correlated well with the change observed from the official agricultural statistics at the province level. The detailed zoom-in to Orenburg province in Russia also showed only 30% of grasslands were mowed in 2018. By looking at feature importance with random forest, we noticed SWIR bands and simple phenology metrics such as standard deviation calculated from May 1st to October 1st, contributed to improving the separability of thematic classes.
Despite steppe recovery due to widespread cropland abandonment from 1990 to 2018, the steppes, including the recovered steppe patches, underwent fragmentation due to other disturbances. Systematic mapping with very high-resolution images revealed fragmentation of steppes due to informal roads, oil and gas development, shrub encroachment, and garbage dumps across Eastern Ukraine and Samara and Orenburg of Russia. Also, the remaining steppe patches both in controlled and insurgent parts of Eastern Ukraine additionally experienced land degradation due to military land use. The mapped disturbances occurred primarily near settlements and roads, except for military use in Ukraine. Only 7% of analyzed areas for detailed disturbances were without such disturbances and were often found in age-environmentally marginal areas, such as the south and southeast corner of Orenburg province of Russia. Our experiments with U-Net CNN showed the great potential of this classification approach. Yet, the requirement of having a large training set and complexity in the parameterization of -Net CNN leaves room for random forest classifier or participatory crowdsourcing approaches we used too to document detailed disturbances.

Our study showed a great advantage of using Sentinel-1, Sentinel-2, Landsat time series to map the recovery of steppes. A multisensory approach, such as applying very-high-resolution imagery, is the way of mapping the detailed disturbances that are difficult to capture with Sentinel-1, Sentinel-2, Landsat time series. The presented study can be relevant beyond the Eurasian steppes, and the approach can be adapted to map disturbances of grasslands and other biomes and their degradation in other parts of the world.

Cited literature
Baumann, M., Kamp, J., Pötzschner, F., Bleyhl, B., Dara, A., Hankerson, B., Prishchepov, A.V., Schierhorn, F., Müller, D., Hölzel, N., Krämer, R., Urazaliyev, R., Kuemmerle, T., 2020. Declining human pressure and opportunities for rewilding in the steppes of Eurasia. Divers Distrib 1058–1070. https://doi.org/10.1111/ddi.13110
Dara, A., Baumann, M., Freitag, M., Hölzel, N., Hostert, P., Kamp, J., Müller, D., Prishchepov, A.V., Kuemmerle, T., 2020. Annual Landsat time series reveal post-Soviet changes in grazing pressure. Remote Sensing of Environment 239, 111667. https://doi.org/10.1016/j.rse.2020.111667
Mazloum, B., Pourmanafi, S., Soffianian, A., Salmanmahiny, A., Prishchepov, A.V., 2021. The fate of rangelands: Revealing past and predicting future land‐cover transitions from 1985 to 2036 in the drylands of Central Iran. Land Degrad Dev ldr.3865. https://doi.org/10.1002/ldr.3865
Pazur, R., Prishchepov, A.V., Myachina, K., Verburg, P.H., Levykin, S., Ponkina, E.V., Kazachkov, G., Yakovlev, I., Akhmetov, R., Rogova, N., Bürgi, M., 2021. Restoring steppe landscapes: patterns, drivers and implications in Russia’s steppes. Landscape Ecol 407–425. https://doi.org/10.1007/s10980-020-01174-7
Prishchepov, A.V., Myachina, K.V., Kamp, J., Smelansky, I., Dubrovskaya, S., Ryakhov, R., Grudinin, D., Yakovlev, I., Urazaliyev, R., 2021. Multiple trajectories of grassland fragmentation, degradation, and recovery in Russia’s steppes. Land Degrad Dev 32, 3220–3235. https://doi.org/10.1002/ldr.3976
Prishchepov, A.V., Schierhorn, F., Dronin, N., Ponkina, E.V., Müller, D., 2020. 800 Years of Agricultural Land-use Change in Asian (Eastern) Russia, in: Frühauf, M., Guggenberger, G., Meinel, T., Theesfeld, I., Lentz, S. (Eds.), KULUNDA: Climate Smart Agriculture. Springer International Publishing, Cham, pp. 67–87. https://doi.org/10.1007/978-3-030-15927-6_6
Squires, V.R., Dengler, J., Feng, H., Hua, L., 2021. Grasslands of the world: diversity, management and conservation.

Remote sensing applications are essential for monitoring and management of Asia’s high-mountain rangelands. However, low standing biomass amounts, scarce vegetation cover, ecological heterogeneity and soil variations make space-borne mapping approaches challenging in this environment. Additionally, poor meteorological infrastructure aggravates the identification of the main drivers of vegetation variations. Therefore, we analyzed the applicability of a multitude of remote sensing variables, mainly based on optical Sentinel-2 data, for the classification of land cover, and for the modeling of foliar vegetation fractions and biomass amounts in Afghanistan’s national parks Wakhan and Band-e-Amir. Furthermore, we assessed the influence of snow variables on vegetation as indirect meteorological parameters. We included field data of different years to test several models and variables under various conditions. Repeated spatial cross-validation showed good performance of the classification and the foliar vegetation models using a random forest algorithm, whereas biomass models indicated better results with a Lasso regression approach but with lower performance in general. Importance assessment results for optical variables showed that the short wave infrared bands were essential for modeling in addition to the near infrared domain in this region. Sentinel-2 derived snow variables did not improve modeling results, but showed some potential as vegetation indicators, especially at very low cover values. Furthermore, our temporal analysis showed that MODIS snow anomalies were correlated with vegetation anomalies. This indicates high potential for Sentinel-2 derived snow metrics in the future but at present, the time range of the Sentinel-2 sensor hampers respective analysis. Inclusion of Sentinel-1 backscatter and coherence metrics did not lead to improvements of the models, as the vegetation related signal might be too low in this environment. In summary, our research outlines several approaches for the retrieval of biophysical properties and vegetation monitoring in high-mountain drylands with extensive field data. We thereby promote the applicability of remote sensing methods in a challenging region and address potentials for future improvements.

In Wallonia (Belgium), permanent and temporary grasslands cover respectively around 43% and 4% of the total utilized agricultural area. This ecosystem provides a wide range of services by contributing to food, feed and fibre consumption, by providing climate related services (e.g. carbon/water storage) and by offering biodiversity and landscape amenity value.

Managing grasslands parcels is rather complex and implies from farmers constant decision-making during the growing season. These decisions rely notably on the assessment of grass availability over space and time. This information is usually not available, at best estimated from time-consuming and rather inaccurate in-situ field observations (e.g. with a plate meter).

Satellite remote sensing offers obvious opportunities for a large scale monitoring and quantification of grass production across and within fields. This study aims at assessing the potentialities of Sentinel data to monitor grassland growth. In situ observations (height, biomass) collected in grazing and mowing parcels in Wallonia (southern part of Belgium) have been related to biophysical variables and indices derived from Sentinel-1 and Sentinel-2.

Strong relationships have been observed with Leaf Area index (LAI) derived from Sen2Agri open source system (www.esa-sen2agri.org/) whereas vegetation indices, such as NDVI, tend to present saturation at high level of biomass.

Intra-field heterogeneity has been also assessed based on compressed sward heights (CSH) measured with a plate meter and through UAVs equipped with a Micasense RedEdge-M camera. Compressed sward height measurements have been precisely geolocalized with a RTK. UAV flights were performed closely to the date of Sentinel-2 acquisitions. Ground-based intra-field heterogeneity patterns assessed by CSH have been also observed from UAV and satellite.

The results acquired in the frame of this study represent a first step towards the development of grassland monitoring system in Wallonia. Through assimilation in grass growth models, it allows to estimate and predict grass growth and quality in Wallonia on a daily step. This system would encourage farmers to adopt more easily grass-based systems and would provide relevant data for feeding decision support systems. The information would be also relevant for regional authorities and insurance companies (CAP monitoring, agricultural calamities…).

Intensive field campaigns in grassland parcels planned in 2022, 2023 and 2024 with a view to develop such a decision support system that will constitute the reference grasslands growth observatory for Wallonia.

Grasslands fulfill important ecosystem services, like carbon storage, water purification and the provision of habitats. In addition, managed grasslands provide fodder for livestock and play an important economic role. In Germany, 13% of the landscape is covered by grassland, which accounts for about one-third of the agriculturally used area. The management type (grazing, mowing or a combination of both) and intensity (stocking rate, mowing frequency, fertilizer application) varies in space and time and strongly affects species composition and ecosystem services. In Germany, most of the managed grasslands are regularly mown. The timing and frequency of mowing events are usually unknown on a regional scale, but would be required e.g. for an assessment of ecosystem services or ecosystem matter fluxes like for nitrogen.
Time series of remote sensing data can be used to analyze the temporal and spatial dynamics of grasslands. Depending on the chosen data source, analyses at high spatial resolution and with large coverage are possible. Previous studies showed that vegetation indices based on optical remote sensing data achieve good relationships to grassland mowing dynamics. However, also major limiting factors, namely cloud-gaps and the availability of comprehensive reference datasets, which are needed for calibration and validation, were identified. In contrast, data from active sensors like Sentinel-1 are available independent of cloud coverage. But previous studies showed a less consistent and less explicit relationship of Synthetic Aperture Radar (SAR)-based parameters to grassland mowing events. Within a study testing grassland parcels in 3 regions in Germany, a combination of optical and SAR-based parameters revealed improved results compared to the model only including optical data in a classification approach. However, the feasibility of optical and SAR-based parameters, and their combination, to detect grassland mowing events has still to be evaluated with a comprehensive reference dataset and for a large region of interest, consisting of heterogeneously used grasslands.
This study presents a grassland mowing event detection approach for entire Germany for the year 2019. Time series of several parameters of the Copernicus Sentinel-2 (S2) and Seninel-1 (S1) missions were analyzed regarding their potential to detect mowing events, including the Enhanced Vegetation Index (EVI), radar backscatter VV and VH, the polarimetric decomposition parameters (PolSAR) entropy, K0 and K1, and interferometric (InSAR) coherence VV and VH. Mowing event detections of these were calibrated and validated with a reference dataset of daily RGB camera information, revealing mowing events and other management actions. The reference dataset included information of 162 grassland parcels in Germany which were mown one to six times in 2019, resulting in 491 observed mowing events in total.
After pre-processing and filtering the S2 data with the MAJA algorithm (Hagolle et al. 2017), time series of all parameters were interpolated and smoothed and local minima/maxima were located. These were classified as mowing events according to a calibrated thresholding mechanism. The approach using only optical data resulted in 60.3% of successful mowing event detections for entire Germany. In general, the detection based on EVI (S2) was successful when dense, cloud-free data was available. From the SAR-based parameters, only PolSAR entropy and, slightly less, InSAR coherence VV and VH showed effects (increases) after grassland mowing events. However, these effects were not always present at each mowing event. The PolSAR entropy was therefore analyzed during long cloud-gaps of the optical data regarding its potential to detect mowing events for a focus region in southern Germany, which is characterized by high cloud cover, as well as a high heterogeneity of grassland use intensities and climatic gradients. By including the PolSAR entropy, the number of successful mowing detections increased from 64.6% to 73.8% for the focus region. However, also the number of falsely detected events increased, resulting in a decreasing F1-Score of 0.61 for the combined S2+S1 detection approach compared to 0.65 for the approach only using S2.
According to our results using the S2-based detection approach, 13% of grasslands in Germany are not mown at all and a majority is only mown one (38%) to two times (33%) and is probably also used for grazing. Only 3% of all grasslands are mown four to six times and are therefore used intensively. Areas with many intensively used grasslands are in southern and south-eastern Germany and the very north.
Future studies should test if the inclusion of additional potential drivers of the SAR-based parameters, like precipitation, improve the mowing event detection during cloud-affected periods, for example by integrating it into a deep learning model.

Livestock farming is an important part of the Armenian agricultural development strategy. The agricultural sector employs more than one third of Armenia’s labor force and accounts for 13% of GDP, hence threats to livestock and pastures can significantly impact livelihoods. For sustaining and developing that sector, fodder provision from grasslands is a key factor. Grasslands constitute 39% of the total territory of Armenia and 57% of the agricultural lands. Apart from resources for livestock, they provide important areas for biodiversity and ecosystem services. The condition of natural pastures and grasslands, however, is being deteriorated due to anthropogenic pressure and unsustainable management practices, leading to overgrazing and erosion. These risks are potentially further aggravated through climatic changes such as more frequent droughts, heat waves, and lack of snow cover. Hence, the setup of an integrated management approach for local decision-making becomes important, emphasizing the need of robust and up-to-date spatial data.
In the context of the “GrassAM” project conducted by DLR and GIZ, we aimed at mapping grassland extent, grasslands types, grassland above ground biomass (AGB) and livestock carrying capacities at 10 m spatial resolution in the entire country of Armenia in the year 2020. In order to create a grassland mask for Armenia, a land use and land cover (LULC) classification was carried out using in situ data together with Sentinel-1, Sentinel-2, and digital elevation (DEM) data in a random forest classification approach implemented on Google Earth Engine. 400 sample points of 7 classes (“pasture”, “meadow”, “other grasslands”, “annual arable land”, “perennial arable land”, “bushland”, “bare soil”) were collected by the partner organization ICARE during summer 2020, distributed over all districts and ecological zones. To complement the classification, additional points were sampled on screen for the class “water”. Urban and forest areas were masked using DLR’s World Settlement Footprint 2015 [1] at 10 m resolution as well as the Hansen Global Forest Change maps [2] at 30 m resolution, respectively. The resulting classification achieved an overall accuracy of 80%, while the grassland area was slightly overestimated with 79% user’s accuracy and 92% producer’s accuracy.
Of the 400 in situ sites, 147 pasture and meadow points also included wet and dry AGB samples. These measurements have been collected in one 30 x 30 cm plot per field (mowing at 2 cm height), which was assumed to be representative for the surrounding 30 x 30 m. The fresh plant mass was placed in a paper container, labeled and weighed with precision of 0.1 grams. Samples were then dried at room temperature for 48 - 72 hours and weighed again. The measured green AGB ranges from 1.733 – 27.800 kg/ha, with a mean of 12.367 kg/ha, and dry AGB ranges from 1.011 – 14.300 kg/ha, with a mean of 5.416 kg/ha. The AGB measurements were split 60/40 in training and validation data. To create a spatially balanced training data, the selection of training samples was based on allocation of points in hexagon tessellation (1 point per grid cell; 2 points if there are more than four samples are available per cell). Biomass was modeled in a next step using a random forest regression model. The training samples have been used to test a set of 730 different geospatial features (monthly statistics and bi-weekly interpolated features of B2 - B12 Sentinel-2 bands and of eight vegetation indices, elevation, slope, monthly mean temperature at 2 m, monthly precipitation sums) as predictors using a Sequential Forward Feature Selection. Six features (Sentienl-2 mid-June and mid-July NDVI, Band 12 median, May precipitation, June temperature, elevation) were selected and achieved a R-square of 0.66 with an RMSE of 4.013 kg/ha for green AGB. The country-wide biomass maps are the basis to model grassland carrying capacity, i.e. the maximum number of cattle equivalent animals that can be sustained in a given grassland area in a season. AGB was multiplied with a proper use factor of 0.65 as it was suggested by [3] to estimate the available fodder. This amount is divided by the daily requirement of fodder per animal unit (equivalent of 400 kg live weight of cows) multiplied by pasture season length. For both quantities, landscape-zone specific assumptions have been made, resulting in an optimal stocking density of 1- 3 animals per hectare.
Test for improving biomass and carrying capacity models as well as the input data sets are still ongoing. The resulting maps, that characterize the allowable grazing pressure on a country-wide scale, could be used to improve grassland management and to increase the resilience of grassland ecosystems to future climate conditions.

[1] Marconcini, M., Metz-Marconcini, A., Üreyen, S., Palacios-Lopez, D., Hanke, W., Bachofer, F., Zeidler, J., Esch, T., Gorelick, N., Kakarla, A., & Strano, E. (2020): Outlining Where Humans Live –The World Settlements Footprint 2015. Scientific Data7(242). doi.org/10.1038/s41597-020-00580-5.

[2] Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend (2013): High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 342, 850–53.

[3] de Leeuw, J. Rizayeva, A., Namazov, E., Bayramov, E., Marshall, M. T., Etzold, J., Neudert,R. (2019): Application of the MODIS MOD 17 Net Primary Production product in grassland carrying capacity assessment, International Journal of Applied Earth Observation and Geoinformation 78, 66-76, https://doi.org/10.1016/j.jag.2018.09.014.

Grasslands cover 26% of the global land area, and 70% of the agricultural surfaces worldwide. They regulate carbon and water cycles, constitute a biodiversity reservoir and an important food source for livestock. In mountains, they are key constituents of the natural, economical and cultural heritage and part of the traditional transhumance systems. They are however threatened by climate and land use change, that can be faced only by means of effective adaptation and mitigation strategies. A pivotal condition for effective strategies to take place is the accurate quantification of mountain pastoral resources across space and time. This effort can nowadays take advantage of the invaluable contribution of Sentinel-derived data streams.

In this presentation, we will illustrate a random forest approach used to predict the distribution and the productivity of grassland surfaces in a topographically complex protected area (Gran Paradiso National Park, 710 km2) in north-western Italy. We built on a massive terrain vegetation survey as ground truth, and on remote-sensing-derived, climatic and topographic layers as predictors. We had two objectives: a) modeling the distribution of mountain grasslands across the entire park at a 20-meters spatial resolution, and b) modeling the productivity of these areas.

We classified grassland presence/absence with high accuracy (up to 90%). A comparison of our results to the standard Copernicus European Grassland Product revealed the presence of extensive high altitude grassland areas potentially available for wild herbivores. Also grassland productivity was modeled with high accuracy both according to three broad productivity classes (90% accuracy) and to a more detailed classification into thirteen pastoral categories (80% accuracy). Productivity estimates agree well with satellite-derived leaf area index maps and with area-averaged NDVI seasonal patterns.
We developed a method that could be deployed in similar landscape settings, and demonstrated that accurate field campaigns and high-resolution remote sensing allows for robust prediction of grassland distribution and productivity even in complex terrains. This information can contribute to improve the management of pastoral resources, and promote effective adaptation strategies.

Grasslands cover ca. 7% (2,100,000 km2) of the African continent. They provide a wide range of ecosystem services (e.g., forage, water, recreational spaces, carbon sequestration), and host large wildlife communities. Despite their importance, African grasslands are reported to be suffering from degradation and, perhaps more worryingly, have received little consideration within international policies (e.g., United Nations Sustainable Development Goals). One of the key issues at present is the gradual encroachment of woody plants, which is shifting African grassland from a grassy- to a (less palatable) woody-dominated biome. However, the way climatic and non-climatic disturbances interact with the woody vegetation layer is still poorly understood, particularly at large spatiotemporal scales. Addressing this gap would shed new light on grassland ecosystem dynamics and feed the prevention of potential biome shifts. Here we identified grasslands in sub-Saharan Africa according to the ESA Climate Change Initiative (CCI) land cover product and use annual minimum vegetation optical depth (VOD) from passive microwave observations as a proxy for woody vegetation. We then use independent climatic (precipitation, wet and dry season metrics and accumulated rainfall metrics derived from CHIRPS, and ESA CCI soil moisture) and non-climatic (GFED4s with small fires and Sentinel-2 MSI derived burned areas, and SEDAC population density) data within a structural equation modelling (SEM) framework to assess how both spatiotemporal variations and interactions between climatic and non-climatic drivers affected VOD during 1997-2016. Preliminary results based on four terrestrial ecoregions (Sahel Acacia Savannas, Greater Karoo and Kalahari drylands, Southeast African subtropical grasslands, Madagascar Island) reveal strong collinearity between different accumulated rainfall metrics, meaning that changing the accumulation period may be irrelevant to predict VOD (and burned areas). We also observed distinct fire regimes. A decrease in burned areas across the Sahel and Kalahari may promote the proliferation of woody plant species, while steady fire activities in South African and Malagasy grasslands leave room for interpretation including other disturbances (e.g., grazing) or more complicated interactions between climatic and non-climatic forces.

Land cover and land use change monitoring is fundamental for ecosystem services, global biodiversity, food security, and climate change analyses. To provide management relevant information, land cover and land use change analyses need to be carried out at suitable spatial and temporal scales. Remote sensing data are an outstanding source of information on the Earth surface and moderate-to-high spatial resolution observations are routinely used for land cover and land use change monitoring. Ever-increasing archives of satellite data allow to obtain information at the ‘field level’ going back to, as early as, the 1980s. Among others, Copernicus Sentinel-2 missions and NASA/USGS Landsat program provide the most widely accessible satellite observations. While the Sentinel-2 mission offers high revisit time and 10- to 20 m resolution, its observation record is still relatively short. Landsat satellites have been acquiring data every 16 days at 30-m resolution ensuring an uninterrupted data record since the early 1980s. Convergence between technical specifications of both satellite systems therefore creates a unique opportunity for data integration for land cover and land use change analyses, which has been implemented, among others, in the Harmonized Landsat Sentinel (HLS) project and the Sen2like framework.
Here we analyzed potential gains and tradeoffs of synergetic use of Sentinel-2 and Landsat collection 2 time series for long-term grassland monitoring. Specifically, i) we quantified the increase in density of clear-sky observations when supplementing 2015-2021 Landsat observations with Sentinel-2 time series; ii) we evaluated whether integrating 2015 2021 Sentinel 2 data into the 1984-2021 Landsat time series enhances quality of the resulting time series for long-term analyses. We implemented our analysis for test sites across Europe characterized by different climate and hence different cloud cover probability. We used two most popular variants of data aggregation to construct annual time series: i) a single annual data point per-pixel representing land cover conditions during a selected phenology period, here the summer, and derived as a maximum value composite; ii) annual sums of monthly composites cumulated over the pixel-specific growing seasons, corresponding to GPP.
We compared the utility of Landsat – Sentinel-2 integration for long-term grassland monitoring using all available clear-sky pixels in 2015-2021 Sentinel-2 and 1984-12021 Landsat observations. We compared the frequency of the clear sky acquisitions calculated based on 2015 2021 Landsat time series and combined Landsat – Sentinel-2 data. We integrated Landsat and Sentinel-2 time series for long-term trend analysis through ground cover fraction estimates. We used the Land Use/Cover Area frame Survey (LUCAS) to identify compatible Landsat- and Sentinel-2-specific image endmembers characteristic for temperate grassland ecosystems (i.e., green vegetation, non photosynthetic vegetation and soil) and ran Spectral Mixture Analyses on each available pixel using the corresponding image endmembers. We consolidated both time series, and derived the maximum summer composites and monthly composites used for annual sums prioritizing Sentinel-2 data and pixels with the lowest unmixing RMSE. When needed, we predicted missing monthly data (Lewińska et al., 2021) before calculating annual sums of ground cover fractions, i.e., Cumulative Endmember Fractions (Lewińska et al., 2021, 2020). To limit interference of variance in the length of snow cover, we calculated Cumulative Endmember Fractions only for photosynthetically active periods defined on a per pixel basis and unified across the time series. We compared the resulting 2015-2021 time series by measuring the absolute difference between summer maximum value composites and Cumulative Endmember Fractions (for all three ground cover fractions) calculated for Landsat only and integrated Landsat – Sentinel 2 time series. Finally, we analyzed differences in 1984 2021 long term trends in grassland ground covers calculated using summer maximum value composite and Cumulative Endmember Fractions, based on Landsat time series and integrated Landsat – Sentinel-2 time series. We analyzed trends in all three ground cover fractions excluding the effect of temporal autocorrelation (Ives et al., 2021).
Our work evaluates the benefits of Landsat and Sentinel-2 integration for long-term grassland analyses requiring different time series length and data availability. Our findings allow a deeper understanding of the tradeoffs between, on the one hand, enhanced data quality and observation frequency and, on the other hand, efforts of integrating multi-source data from Landsat and Sentinel-2. Moreover, we present here an alternative workflow for integration of Sentinel-2 and Landsat archives for grassland monitoring based on time series of ground cover fractions.

References:
Ives, A.R., Zhu, L., Wang, F., Zhu, J., Morrow, C.J., Radeloff, C., 2021. Statistical inference for trends in spatiotemporal data. Remote Sens. Environ. 266, 112678. https://doi.org/10.1016/j.rse.2021.112678
Lewińska, K.E., Buchner, J., Bleyhl, B., Hostert, P., 2021. Changes in the grasslands of the Caucasus based on Cumulative Endmember Fractions from the full 1987 – 2019 Landsat record. Sci. Remote Sens. 4. https://doi.org/10.1016/j.srs.2021.100035
Lewińska, K.E., Hostert, P., Buchner, J., Bleyhl, B., Radeloff, V.C., 2020. Short-term vegetation loss versus decadal degradation of grasslands in the Caucasus based on Cumulative Endmember Fractions. Remote Sens. Environ. 248. https://doi.org/10.1016/j.rse.2020.111969

Normalized difference (ND) metrics are essentially scaled estimates of the spectral slope, analogous to a 1st derivative. The metrics are related to spectral derivatives in their sensitivity to spectral shape and their insensitivity to the magnitude of the reflectance. This can be effective in capturing the differences in spectral shape between bare soil (BS) and non-photosynthetic vegetation (NPV).
Consider the finite difference approximation of a 1st derivative and the associated ND metric:

∂ρ/∂λ = [ ρ(λ2 )−ρ(λ1) ] / (λ2−λ1) (Eq. 1)

ND(ρ1, ρ2 ) = (ρ2 -- ρ1) / (ρ2 + ρ1 ) (Eq. 2)

where λ represents the wavelength, and ρ indicates reflectance. Both are responsive to the difference in reflectance at specific wavelengths (or wavebands). The derivative is scaled by the wavelength difference and will respond to changes in that difference; the same magnitude difference over a shorter wavelength difference represents a steeper slope. The wavelength difference is not very meaningful in multispectral image analysis where the wavelengths (and wavebands) are fixed. However, brightness can change drastically within an image, and between images collected at different times, making it difficult to compare spectral features. Scaling by the magnitude of the reflectance, as is done with the ND metrics, removes the brightness variations and creates a dimensionless, scaled slope that allows for more uniform comparison of color differences.
This concept can be extended to a representation of spectral curvature, analogous to the second derivative. As with slope, the ND normalization would use the magnitude of the reflectance where the second derivative uses wavelength differences. It will also be convenient if the metric scales from [-1, 1] and is simple to compute. Following the general form of Eq. (2), the corresponding normalized curvature (NC) formula is:

NC(ρ1,ρ2,ρ3 ) = (ρ3 − 2ρ2 + ρ1) / (ρ3 + 2ρ2 + ρ1 ) (Eq. 3)

This expression maintains the convenient scaling [-1,1], with -1 representing the maximum negative curvature (opening downward), 0 representing a straight line, and +1 representing the maximum positive curvature (opening upward).
Given an appropriate choice of spectral bands, the ND and NC metrics can provide a simple, but effective characterization of the spectral shape critical to discriminating specific targets. This approach is examined using a selection of data from the ECOSTRESS spectral library to mimic the four 10-m, VNIR spectral bands available on Sentinel-2 to distinguish between BS and NPV. Discrimination between BS and NPC is poor if the full spectral width of each band is modeled, but improves markedly as the bandwidth narrows around the same central wavelengths.

Extensive rangelands in the arid and semi-arid lands (ASALs) on the Horn of Africa play a critical role for the food security of local populations, whose welfare is largely dependent on livestock herding. Satellite Earth Observation (EO) with medium to coarse resolution data (e.g., MODIS) is widely used in the region for operational monitoring of rangelands dynamics, often in connection to drought early warning systems and climate risk financing initiatives.
Despite the fact that semi-arid rangelands are characterized by a large heterogeneity, efforts towards understanding the fine-level spatial variability and rapid vegetation dynamics of rangeland have been limited. However, such understanding is a prerequisite for predicting how changing weather and grazing patterns may affect the future productivity of these systems. As such, it is fundamental to enhance EO systems, both ground and satellite-based, aimed at monitoring pastoral production systems and their stressors at fine scale.
This study aims at providing an overview of the research conducted at the Kapiti Research Station of the International Livestock Research Institute (ILRI) with the objective of investigating the complex spatial and temporal patterns of typical rangeland vegetation in the Kenyan ASALs. To that end, a variety of experimental activities and in-situ measurements have been performed in recent years in this rangeland site of 128-km2. In terms of in-situ EO observations, since 2019, the research station has been equipped with a Fluorescence Box (FloX) field hyperspectral spectrometer (JB hyperspectral devices, Düsseldorf, Germany), acquiring continuous visible and near-infrared reflectance and sun-induced fluorescence (SIF) measurements. We have also installed an eddy covariance tower and several digital cameras for phenological observations and livestock/wildlife detection. These measurements together provide a unique ground reference dataset in the African continent for assessing quantitatively how the continuous stream of high-resolution EO data, nowadays available from sensors such as Sentinel-1, Sentinel-2 and PlanetScope, can inform the development of next-generation of rangeland monitoring tools. While such tools may also incorporate coarser-resolution data for regional analyses, for example with respect to drought, the fine-scale monitoring could provide direct support for decision making by land managers and pastoralists with respect to vegetation management and herd mobility.
Due to climatic shifts in East Africa, the last three years have displayed an unprecedented weather variability in the region, including the highest seasonal rainfall on records, several pulse events (out-of-season rainfall events and dry spells), and severe drought. Therefore, key questions that we are investigating using multiple EO datasets include a) how the retrieval of critical phenological stages (e.g. start of season, peak season and end of season) is affected by persistent cloud cover, b) the extent to which alternative vegetation indices can accurately describe quick growth patterns associated to weather variability (dry spells or rainfall pulses) or intensive grazing events, c) the potential of SIF as a physiological indicator for early detect drought stress and for assessing rapid changes in gross primary productivity, and d) the assessment of fine-scale spatial heterogeneity in phenology and productivity that are associated to different grass communities, life forms, and grazing patterns.

The Australian Government, through the National Landcare Program, invests in farm-scale interventions aimed at improving the condition of soils and biodiversity to benefit the environment and communities. One of the challenges of this Program is to adequately measure the impact of the interventions to assess their effectiveness. Examples of these intervention include fencing to reduce or exclude stocking rates; culling of feral animals (such as goats); application of lime or water ponding earthworks to reduce soil salinity, among others.
We have developed a method, in consultation with the end users, to monitor the effects of these interventions. The method uses vegetation cover detected using satellite imagery for frequent national and regional reporting. Vegetation cover protects the soil, making it a good indicator of the risk of soil erosion by wind and water. In addition to reducing soil erosion, vegetation cover increases water infiltration into soil, reduces evaporation, drives soil carbon sequestration, maintains or improves soil condition, and enhances biodiversity and agricultural production.
Monthly national vegetation cover data with a spatial resolution of 10 to 500 metres is delivered via the Rangelands and Pasture Productivity (RaPP) Map online tool (https://map.geo-rapp.org/). The vegetation cover data is derived from the MODIS, Landsat and Sentinel 2 satellites and is produced using a spectral linear unmixing technique underpinned by more than 4000 field observations across Australia (Guerschman et al 2015, 2018, Scarth et al 2010). The RaPP Map tool allows users to visualise maps of vegetation cover, extract time series data for any location of interest and compare locations across space and time. The recipients of government funding can then track changes in vegetation cover where interventions have been made, compare with control areas, assess the spatial variability in the changes in vegetation cover and include all these outputs in their annual reporting.
One of the challenges of monitoring vegetation change is to properly identify areas where the changes are a result of the on-the-ground interventions vs. those changes due exclusively to natural rainfall variability. We are addressing this problem in 3 complementary ways. First, where possible, a control area is established at the beginning of the project, or a similar area to the one that received the treatment is identified post facto. The comparison between the intervention vs the control area provides the grounds for identifying the specific effects of the interventions. Second, we are modelling the long-term (2000-2018) relationship between vegetation cover and the accumulated antecedent rainfall in each location to produce an estimate of the expected amount of vegetation cover in a given month into the future. Then we compare with the satellite-derived vegetation cover estimate and if the difference is significant (using a statistical approach) we conclude that the intervention is having a positive effect. Third, we are implementing a spatial benchmarking approach, where all the locations similar to the intervention area in vegetation type, position in the landscape and land use are compared. If changes in the relative ranking of the intervention area to the similar areas identified change over time, we also conclude that those changes are due to the interventions made.
In addition to the Landcare Program, the RaPP Map tool and datasets are being used by many users in Australia and internationally to, for example, maximise support for farmers to capture soil carbon; estimate fodder availability and detect anomalies from average values, provide regional assessments of vegetation cover condition and estimate soil management practices in agricultural setting and identify sources and produce forecasts of atmospheric dust (Klose et al 2021). We illustrate the use of the satellite-derived vegetation cover information and the RaPP Map tool through several specific examples.

Guerschman JP et al. 2015. Remote Sens. Environ. 161, 12–26. https://doi.org/10.1016/j.rse.2015.01.021
Guerschman JP and Hill MJ. 2018.Remote Sens. Lett. 9, 696–705. https://doi.org/10.1080/2150704X.2018.1465611
Klose, M. et al. 2021. Geosci. Model Dev. Discuss. 1–59. https://doi.org/10.5194/GMD-2021-32
Scarth P et al. 2010. https://doi.org/doi:10.6084/m9.figshare.94250

The overall goal of the DESTAM (Dense Satellite Time Series for Agricultural Monitoring) project is to develop innovative AI-based methods to condense optical satellite image time series using SAR satellite imagery. The increased information gain through improved time series is necessary, among other things, to solve the problem of high pixel- or field-specific variability of crop systems in smallholder agriculture in Sub-Saharan Africa. In particular, monitoring at key phenological times (beginning, peak, and end of season) is fundamental, and thus close-meshed time series are needed at these times, which often do not exist at present. A continuous monitoring by means of optical satellite remote sensing is, however, not possible, despite the high temporal resolution of modern satellite missions such as Landsat or Sentinel-2. It is only possible to a limited extent because of the frequent cloud coverage, as the cultivation period usually falls in the rainy season (rainfed agriculture). Our study area includes two sites in Africa, specifically in Burkina Faso and in Kenya. Burkina Faso in West Africa is one of the poorest countries in the world and has a high rate of chronic malnutrition. Generally, unimodal agriculture is practiced with main crops (monoculture) being sorghum, millet, maize, rice, groundnuts, and sesame. Kenya in East Africa is among the lower middle-income countries and practices unimodal and bimodal farming of maize, beans, and sorghum, with both monoculture and intercropping. Extensive field surveys are available for 2018 and 2020 (Burkina Faso), and additional surveys have been recorded in both sites in 2021. The recorded information includes, among other things, exact field boundaries, crop types, monoculture or intercropping, unimodal or bimodal cultivation and information on crop rotation. These parameters are linked to freely available 10 m spatial resolution satellite imagery from the Copernicus mission (Sentinel-2 and Sentinel-1) to analyse the impact of vegetation index time series densification on field crop type identification. Although Sentinel-2 has a high revisit time, cloud coverage impedes continuous time series. For this reason, AI-based methods will be developed to condense the Sentinel-2 vegetation index time series using cloud independent Sentinel-1 SAR imagery. The focus here is to analyse different vegetation indices such as NDVI, GNDVI, NDRE and NDWI and the level of densification that is needed to characterise different crop types on the basis of its phenological curve. The phenological curve is analysed over time and the different crops show a different peak and harvest time and therefore drop of the vegetation index, which allows to characterize the different crops. The different vegetation index time series are linked to the in-situ crop type data to predict crop types on household field level. In the results, we show the influence of Sentinel-2 time series densification on the accuracy of field crop type classification. Furthermore, we discuss the different vegetation indices and their suitability for crop type differentiation. Experience has shown that chronic malnutrition is a consequence of lower yields and/or lack of diversity in crops and our findings might contribute to initiate preventive measures. Decision-makers in crisis prevention and insurance companies can use these data to assess productivity bottlenecks and market needs at an early stage.

Surface reflectance data acquired in red and near-infrared spectra by remote sensing sensors are traditionally applied to construct various vegetation indices (VIs), which are related to vegetation biophysical parameters. Most VIs use pre-defined weights (usually equal to 1) for the red and NIR reflectance values, therefore constraining particular weights for red and NIR during the VI design phase, and potentially limiting capabilities of the VI to explain an independent variable. In this paper, we propose an approach to estimate biophysical variables, such as Leaf Area Index (LAI), Canopy Chlorophyll Content (CCC) and Fraction of Photosynthetically Active Radiation (FPAR) absorbed by green vegetation, represented as linear combinations of the red and NIR reflectances with weights determined empirically from observations and radiative transfer model (PROSAIL) simulations. The proof of concept is first tested on available close-range observations over maize and soybean crops in Nebraska, USA. The empirical results compare well with those from PROSAIL model simulations. The proposed LAI model is then used with data from Landsat 8, Sentinel-2 and Planet/Dove, and the results are validated with in situ LAI measurements in Ukraine. We show that the weights on red and NIR reflectances are vegetation-specific and stable in time. The approach is further tested on crops and forests in the conterminous USA and on a global scale using MODIS LAI and FPAR products as proxies for “ground observations”. It is encouraging to see that the derived maps of coefficients/weights exhibit regular patterns over the globe compatible with those of vegetation classes and crop types. Tedious and thorough work on compiling available in situ measurements on various crop types needs to be accomplished prior to large-scale applications, and the method needs to be further tested and proven that it works at a large scale.
The proposed parameterization may be attractive for global studies of various sub-classes of vegetation, once the parameter coefficients are established, validated, tabulated and their stability verified. Ultimately, this approach may provide quantification of vegetation traits for the past decades and be a useful asset for climate models that include satellite-derived land cover classifications and vegetation variables for simulating surface fluxes. This is a conceptual work describing a paradigm, which could ultimately be useful in global models.

Sustainable agricultural growth is vital to ensure food security in the Africa’s Sahel region in meeting the demand of rapidly expanding population, those already suffering from rising challenges of climate change and conflicts. Due to the low soil fertility in this arid and semi-arid region, traditionally, crop rotation practice is commonly adopted to maintain soil fertility as cropped field will be left as fallow for a number of consecutive years and to serve as the feed land for livestock. The areal difference in cropped and fallow field can serve as an indicator of cropland expansion (fallow field/natural savannahs as main land use) and intensification of agriculture (where cropped field as main land use). Accurately mapping the extent of cropped and fallow fields is thus indispensable for a better understanding of the status of Sahelian farmlands. We advanced our previous work on mapping these two land use/cover types of farmlands by using the new generation optical imagery of Planet Basemap provided via Norwegian International Climate and Forest Initiative program (NICFI) and state of art convolutional neural network image segmentation algorithm. We found that cropland expansion occurs in the frontiers of farmlands and natural savanna, while intensified farmlands are featured by a higher woody cover in the cropped field. Given limited availability of resources in this region, a well conservation of trees in the field is crucial for sustainable land use intensification and our map may be used to identify the hotspots of unsustainable development of farmlands.

The spatial modeling has become a dominant means to estimate evapotranspiration (ET) fluxes over regional and continental areas. One of the most widely used ET spatial models is the land surface temperature-based approach as the land surface temperature (LST) is potentially a signature of both ET and the soil water availability via the surface energy balance. In recent decade, many efforts have been devoted to extract the LST from remote sensing data. Nevertheless, the spatial and temporal resolution of thermal-based remote sensing are coarse for the hydro-agriculture purposes. Therefore, there is a crucial need for LST data at higher spatial/ temporal resolution for monitoring the plant water status at the field scale. Consequently, downscaling techniques have been proposed to improve the spatial resolution of the LST data available at a high temporal frequency. The aim of this work is to develop a simple method to disaggregate land surface temperature and assess its impact on three LST-based energy balance models. To do so, LST derived from MODIS (MODerate resolution Imaging Spectroradiometer) at 1 km resolution and Sentinel-2 optical bands at 20 m resolution have been selected. Multiple regression equation between the 6 optical bands and the LST has been applied to disaggregate the LST from 1 km to 20 m resolution. The disaggregated LST have been used to feed the TSEB (Two source energy balance model), the PM (Penman Monteith model) and the METRIC (Mapping EvapoTranspiration at high Resolution with Internalized Calibration model). The predicted ET using the three energy balance models have been compared to Eddy covariance measurements installed over rainfed and irrigated wheat fields in the Tensift basin, central Morocco. The outcomes of the disaggregation approach compared to in-situ LST, and the results of the estimated ET compared to EC system will be presented during the conference. This targeted research is very promising and responds to the needs of managing agencies by providing spatialized ET over different canopies and scales.

The pressure on biodiversity and ecosystems due to the loss of natural and semi-natural vegetated land is more and more problematic. To tackle this issue, land cover change monitoring techniques are key. Also, in the particular context of the Common Agricultural Policy, the detection of new non-eligible areas – such as buildings and infrastructures – inside agricultural parcels is of great importance. As part of the Area Monitoring System, it allows to keep accurate and up-to-date agricultural field boundaries maps, which is one of the main conditions for identifying crop types, agricultural practices, etc.
Such detections can be done using image recognition algorithms on Very High Resolution images. However, this kind of data is not available everywhere. Furthermore, when they exist, these images are usually not frequently updated. On the contrary, although having a lower spatial resolution, Sentinel-2 (S2) images have a global coverage at a high temporal resolution. Although the spatial resolution is reduced, this allows for a more frequent detection of some changes.
The present work explores the possibility of using S2 data to regularly detect new non-eligible elements in agricultural parcels in an automated way. The primary motivation is to understand the advantages and limitations of such approach in the operational context of the CAP monitoring.
This study is conducted in the Walloon Region, the southern portion of Belgium, on an area of interest of 20 x 20 km². It is covered by a single Sentinel-2 tile, from which time series have been derived for the 2020 agricultural season. The Land Parcel and Identification System is also used. It contains polygons of agricultural blocks that are surrounded by physical non-agricultural elements (road, forest…) and polygons of fields parcels with information about their crop type.
The algorithm is based on the simple idea that the dynamics of some vegetation indices, such as the NDVI, of a urban area is very flat compared to an agricultural area. Several additional tricks are implemented to lower the number of false detections. They include a filter based on the knowledge about the crop type inside the parcels, or the application of a negative buffer to remove border effects. Additional geometric techniques are also needed to mitigate the effect of the imperfect alignment of S2 images from date to date which breaks the time series of single pixels.
In order to test the detection algorithm, the 2019 and 2020 orthophotos (25 cm resolution) are used to create a validation dataset containing 46 urban elements and about 4000 non-urban polygons. On this validation dataset, the algorithm correctly detects 78 % of agricultural polygons including urban elements. The 22 % that are missed mainly concern parcels for which the non-eligible element is not contained in the polygon after the buffer application. On the other side, one should note that the algorithm still has a high number of false positives (~ 10%). Depending on the cause of the false detection, several options could be considered to lower that figure: a maximum threshold on the number of flagged pixels inside a parcel, imposing a change condition on the time series instead of just a threshold (especially for grasslands), …
These results indicates that Sentinel-2 data can be relevant for the detection of new non-eligible elements inside agricultural parcels in an automated way. In the context where information about the agricultural practices is available, it allows for more frequent detections than using VHR imagery. This is especially relevant in the operational context of the CAP. However, further exploration is needed in order to lower the number of false positives.

Accurate agricultural yield prediction is a fundamental tool for sustainable agricultural planning and for ensuring food security in regions critically affected by climate change and extreme weather events. Existing regression-based approaches of crop yield estimation typically focus on a specific set of predictor variables, and have not been compared systematically. This paper demonstrates and compares the utilization and the combinatorial use of three different sets of object-based predictors from Earth Observation data of the Sentinel-2 satellites for sugarcane yield estimation. The study area is located in an irrigated landscape in the south-east of Baska Lake in Metahara, Ethiopia, which is located 180km east of the state capital Addis Ababa. It encompasses an area of circa 10.000 hectares of irrigated area. In-situ data of sugarcane yield and sugar quantity in quintals per hectare was collected for the growing season between the years 2018 and 2020 for 547 parcel objects. Multitemporal Sentinel-2 Level-2A data was used for the time period between ratooning (April 2018) and harvesting (December 2019), capturing the growing season for sugarcane for the years 2018-2020. We compare several regression models using a range of predictor variables, such as (i) multi-temporal variables (i.e., parcel-based vegetation index time series), (ii) time series descriptors (i.e., phenological metrics) and (iii) spatio-temporal variables (i.e., cultivated area fraction) - defined as the ratio of actual area under crops and the declared parcel size - development as sub-parcel information at pixel-level. The overarching goal of this study is to compare three different sets of predictors with two different supervised regression algorithms and to evaluate alternative predictor variables from remote sensing data for sugarcane yield estimation. Regression accuracy metrics R² and RMSE are compared for linear and Random Forest regression models. The results revealed the best R² scores of 0.84 for the estimation of sugarcane yield in quintals per hectare and 0.82 for the estimation of sugar quantity through Random Forest regression based on the combinatorial use of all predictor variables. The experiments have also proven that the use of dimensionality-independent phenological metrics achieves good yield estimation results which could be a very useful variable set for model transfer and domain adaptation.

Temporal information contained in SAR time-series, combined with new deep learning methodologies, opens the path to a new generation of applications for agriculture monitoring. Indeed, while the majority of the temporal analysis of agricultural SAR time-series is done using polarimetry [1] or interferometry [2] information, new works involving deep learning with the σ0 temporal signal are being published [3]. However, as demonstrated in [4], the use of labels in an agricultural context is often limiting. Indeed, the notion of intra-class variance, i.e. variance in the distribution of data samples belonging to the same class, is not modeled by supervised methodology. Patterns testifying of variations in agricultural practices are then over-looked by such algorithms. Unsupervised learning, and more particularly Autoencoders [4, 5],appear to be a great asset to explicit these intra-class patterns.

This study presents an unsupervised methodology to extract variations in agricultural rice prac-tices using year-long SAR time-series of σ0 backscatter value in both VV and VH. We apply a Convolutional Autoencoder to visualize the multi-temporal stack of rice fields while clarifying existing subgroups of rice fields. A Convolutional Autoencoder is a deep learning architecture made up of two parts: a Convolutional Encoder, trained to extract temporal features and to project them onto a space of lower dimension, i.e. the embedding space and a Decoder, tasked with recreating the original time series, given the embedding values. This bottlenecking design of the encoding process, coupled with the reconstruction task, allows for generating a sparse and separable embedding space.

The explanation behind the separation in these groups is illustrated using a modified Grad-CAM [6] methodology, applied to the embedding layer of the autoencoder to find which key dates are crucial for which embedding space (Red corresponds to high contributions, while blue corresponds to low contributions). The Grad-CAM methodology will measure the contribution of each convolution kernel to the partial derivative of the embedding space and weight them with the activation of these same kernels on each date of the input time-series. This way, we can pinpoint top-contributing patterns within the time-series, for the generation of each embedding dimension.

With this employed strategy, we can show that both the red and green colors of the RGB embedding image appear to result from convolutions applied to the flooding and harvest sea-sons. In contrast, the blue color focuses on the field preparation and growing seasons. Thus,a combination of the variation of all 4 of these descriptors, as analyzed by the Convolutional Autoencoder strategy, results in the image shown in fig. 1. While Autoencoders, as other Deep Learning algorithms, remain hard to interpret, its combined use with Grad-CAM methodologies paves the way for better interpretability of results, as seen in this study.

[1] Alberto Alonso-Gonz ́alez, Carlos Lopez-Mart inez, Konstantinos P. Papathanassiou, andIrena Hajnsek, “Polarimetric sar time series change analysis over agricultural areas,”IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 10, pp. 7317–7330, 2020.

[2] Alejandro Mestre-Quereda, Juan M. Lopez-Sanchez, Fernando Vicente-Guijalba, Alexan-der W. Jacob, and Marcus E. Engdahl, “Time-series of sentinel-1 interferometric coherence and backscatter for crop-type mapping,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 13, pp. 4070–4084, 2020.

[3] Lingbo Yang, Ran Huang, Jingfeng Huang, Tao Lin, Limin Wang, Ruzemaimaiti Mijiti,Pengliang Wei, Chao Tang, Jie Shao, Qiangzi Li, and Xin Du, “Semantic segmentation based on temporal features: Learning of temporal-spatial information from time-series sar images for paddy rice mapping,”IEEE Transactions on Geoscience and Remote Sensing,pp. 1–16, 2021.

[4] Thomas Di Martino, Régis Guinvarc’h, Laetitia Thirion-Lefevre, and Elise Colin Koeniguer, "Beets or cotton? Blind extraction of fine agricultural classes using a convolutional autoencoder applied to temporal SAR signatures,” IEEE Transactions on Geoscience and Remote Sensing, pp. 1–18, 2021.

[5] Thomas Di Martino, Regis Guinvarc’h, Laetitia Thirion-Lefevre, and Elise Colin Koeniguer,“Convolutional autoencoder for unsupervised representation learning of PolSAR time-series,” in 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, 2021, pp. 3506–3509.

[6] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, DeviParikh, and Dhruv Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision,2017, pp. 618–626.

The development of algorithms for simulating surface level biosphere exchanges combined with the increasing spatiotemporal resolution of remote sensing data provide an opportunity to better model and understand agroecosystems and identify sustainable management practices for conserving freshwater resources. The water use efficiency (WUE) allows an evaluation of the relationship between plant productivity and its consumption of water resources and can be defined as the amount of carbon assimilated as biomass or grain produced per unit of water used by crops. In this study, WUE was computed using two method: i) as a ratio between the gross primary productivity (GPP) and the evapotranspiration (ET) and ii) linear regression models using data derived from ERA5-Land meteorological data and vegetation indices from Sentinel-2. While the first method can be applied over different landscapes without the need for calibration, it is very sensitive to the accuracy of GPP and ET. Here, GPP and ET were derived based on the Vegetation Photosynthesis Model and the Priestley-Taylor Jet Propulsion Laboratory model, respectively. The second approach can provide accurate WUE estimation, but the model requires calibration over different environmental conditions. Using a dense olive tree plantation in Saudi Arabia as a test case, both of these approaches were evaluated against ground based water flux measurements obtained from an eddy covariance system. It was determined that a normalized difference vegetation index-based linear regression model produced the best results, with a root mean square error and coefficient of determination of 1.06 g C/Kg H2O and 0.67, respectively. The approach offers the capacity for broad-scale application using commonly available remote sensing and meteorological forcing data sets.

The Czech drought monitor was developed between 2012 and 2014 and has since been operating as an online platform (https://www.intersucho.cz/en). It uses an operational system that consists of four pillars: (1) the daily SoilClim soil moisture model, which runs based on high-density network data from the Czech Hydrometeorological Institute with a 55-year baseline period; (2) weekly reports of surface layer soil moisture and drought impacts on crops provided by dozens of experts; (3) the drought forecast (+9 days) released daily based on an ensemble of five numerical weather prediction models combined with a weekly drought outlook (+2 months); and finally (4) Earth observation (EO) products providing information about soil moisture, state of vegetation condition and water stress in the broader European context. Our contribution will summarize the most important functionalities of the Czech drought monitor with a special focus on EO products and their ability to capture the drought events which caused significant threats in Central Europe and beyond over the last few years. The EO data used within the Czech drought monitor include (1) soil water index (SWI) based spaceborne Advanced Scatterometer (ASCAT) sensor measurements; (2) anomaly in two-band enhanced vegetation index (EVI2) derived from surface reflectances measured by Moderate Resolution Imaging Spectroradiometer (MODIS) onboard of TERRA satellite; and (3) evaporative stress index (ESI) obtained from land surface temperature (MODIS) driven by diagnostic atmosphere land exchange inverse (ALEXI) model. The most important benefits of the implementation of EO methods into the Czech drought monitor include no or limited assumptions about the studied surfaces, no dependence on ground measurements, complete spatial context, and relatively long datasets (2001 till present) with already climatological value. Each of the selected EO methods is describing drought (and its impacts) from a different angle and has a different response time. This, together with the above-mentioned benefits, enables us to evaluate the drought in a more comprehensive way than by using one single method and hence deepens our understanding of the processes related to drought and its impacts and finally increases the ability to design more complex adaptation measures.

Since information is the ultimate enabler, agricultural practices have also evolved and are now highly dependent on Information & Communication Technologies (ICT). The burning of crop fields is of high interest for farmers since the crops are irreversibly damaged and the soil is degraded. Moreover, the burning of crop fields is an important source of pollution. In this context, we propose an ICT algorithm for the mapping of burned crop fields.
As data source we chose Sentinel-2 top of the atmosphere corrected L2A products since they offer a wide variety of spectral bands, high acquisition cadence, open near-real-time availability and good resolution. For test, development, and validation, we chose sites in Argentina, Botswana, Bulgaria, Greece, India, Italy, Republic of Moldova, Romania, South Africa and USA.
The proposed Burned Crops Mapping algorithm (BCMA) uses a time-series of L2A to produce a 10-meter resolution burned croplands map and is designed as a modular multi-step waterfall. The 1st step computes multiple burned area detection indexes (BDI). The 2nd step implements the technique of multi-temporal detection of changes; hence, for each pair of consecutive L2A products, we compute the delta variant of each selected BDI. Next, we train and employ a classification algorithm to classify each pixel based on the values of the respective delta BDIs; i.e. 3rd step. The 4th step corrects the classification where necessary. The last process, the 5th, generates a map of burned pixels for the entire time interval of interest.
Through literature, visual and statistical assessments, we selected the following BDIs: NDVI, BAI, NBR, NBR2 and MIRBI. Moreover, we included the SWIR reflectance value captured by L2A band 8a, as it is mentioned in literature to be a good discriminant of burned areas. By using multiple fold cross-validation and pair corrected T-Test, we evaluated a multitude of classification algorithms, e.g., rule-based, Bayesian-based, tree-based, logistic-based and support vector machines. The chosen algorithm was Naïve-Bayes, which was the best performing algorithm with a percentage of correct classification surpassing the 95% mark.
As advocates of Naive-Bayes, we mention its computational simplicity and demand for a low number of training data. For classification correction, we employed the scene clasification layer (SCL) provided with each L2A. Thus, only pixels SCL classified as a dark area, vegetation, non-vegetated or unclassified can be mapped as burned. To create the final burned area map, we aggregate intermediary maps generated between each pair of consecutive L2A.
For validation purposes, we employed comparisons among BCMA maps, L2A RGB composites and third-party maps. Aadditionally, we compared burned maps generated by BCMA and respective MODIS MCD64A1 burned area maps. Overall, we found the algorithm to produce highly accurate maps of burned croplands. The algorithm can be used with any optical sensor and, thanks to its modular design, it can be easily extended or modified. Moreover, BCMA has linear scalability and low computational complexity.

Soil, as the top layer of the earth, provides the foundation of life. Soil type and quality are closely intertwined with the potential of agricultural use, as different soils provide nutrients and water to crops in varying quantities. Therefore, it is of immediate interest to farmers to determine what soils their lands offer -- to learn for which crops they are particularly suited and where soil improvement measures are needed most. Further, it is important to have knowledge about small-scale spatial differences in the soil in order to coordinate fertilization and till it appropriately. The obvious way to obtain information about soils would be to have samples of them analyzed in the laboratory, which can be costly and time-consuming. Further, remote sensing data can be used to derive soil properties from its spectral information. This direct approach requires bare soil and it can be strongly distorted by factors like soil moisture. To partly overcome this limitation, another approach is to look at the growth behavior of crops and indirectly derive information about spatial disparities in soil from this. The question addressed in this work is whether inferences about differences in soil composition within a delimited area can be drawn based on plant parameters from either multispectral, multi-temporal UAV imagery or a ground-based wireless sensor network. The data is thereby complemented by in-situ measurements. For the experiment presented here, a winter wheat field was observed over the period of one northern temperate zone vegetation season. The selected field was particularly suitable for the experiment because, according to the official soil survey map, two different main soil types - plaggic antrosol and gley - are present here in a direct proximity. Approximately every ten days UAV images were taken with a DJI Phantom 4 multispectral. In addition, the plant height, the chlorophyll content of the top leaf and the soil moisture at a depth of ten centimeters were determined on fifteen plots, 1 ⨉ 1 m each. The ripening phase of the wheat was also continuously recorded with a network of ten multispectral sensors from June 15th to August 3rd. The data collected in this way was then examined to determine whether different expressions of plant characteristics indicated differences in soil and how suitable each of the record methods was for this task. Additionally it was assessed whether either data set would benefit from the integration of the respective other. For evaluation, soil samples were analysed and a multispectral image of the uncovered soil after harvest was recorded. Through the analysis of the data, it was shown that the wheat plants exhibited differences in growth behavior on the slightly different soil types. Wheat plants in the different soil areas differed particularly in growth height, which resulted in fairly varying ripening time. The methods used during the experiment showed advantages and disadvantages; the suitability of both record methods and sensor types used was discussed.

The spatial distribution of different growth patterns of winter triticale (× Triticosecale), winter rye (Secale cereale) and winter barley (Hordeum vulgare) were investigated on an arable field in north-eastern Germany with heterogeneous soil properties. Remote sensing data were used to monitor the extent to which the spatial distribution of growth differences changes over the years from 2012 to 2020. The "Normalized Differentiated Vegetation Index" (NDVI) served as vegetation index to determine those changes. Atmospherically corrected RapidEye and LandSat images with a resolution of 5x5 and 3x3 m, respectively, were used here. Since the sandy soil had spatial heterogeneous properties such as water holding capacity, as well as a within field altitude difference of 6 meters, a large influence of these parameters on plant growth was assumed if the spatial NDVI distribution remained constant and correlated with the soil characteristics. Since different winter cereal species were planted, an influence on the spatial growth distribution was also assumed here, as well as an influence of the climate, in particular an influence of drought in the areas with low water holding capacity. The observed years had all a fairly similar climatic water balance during the growing season (April - June), which was initially positive, but became negative as the year progressed. The NDVI values varied between 0.51 and 0.87. A significant relationship between elevation and NDVI as well as soil classes and NDVI was found in most cases. The lowest NDVI values, and thus the worst growth characteristics, were generally found in the depressions with the lowest water holding capacity. This was true for all crops, as well as for all observation periods. However, rye showed a more heterogeneous distribution of NDVI values, especially in April and May. A climatic influence on the spatial distribution of the NDVI could neither be confirmed nor excluded, due to the similar climatic water balances in all years. The species-specific differences could not be clearly proven due to the small number of observation years, although the strong influence of soil parameters on plant growth was evident. In particular, areas with poor growth characteristics generally remained spatially uniform, especially in the pre-harvest period. Up to 80 % of the grid cells with the lowest relative NDVI values remained spatially consistent over time. The results of this study provide an opportunity to adapted precision agriculture practices in order to optimize crop management and yield.

Accurate mapping of oil palm is important for understanding its past and future impact on the environment. We propose to map and count oil palms by estimating tree densities per pixel for large-scale analysis. This allows for fine-grained analysis, for example regarding different planting patterns. To that end, we propose a new, active deep learning method to estimate oil palm density at large scale from Sentinel-2 satellite images, and apply it to generate complete maps for Malaysia and Indonesia. What makes the regression of oil palm density challenging is the need for representative reference data that covers all relevant geographical conditions across a large territory. Specifically for density estimation, generating reference data involves counting individual trees. To keep the associated labelling effort low we propose an active learning (AL) approach that automatically chooses the most relevant samples to be labelled. Our method relies on estimates of the epistemic model uncertainty and of the diversity among samples, making it possible to retrieve an entire batch of relevant samples in a single iteration. Moreover, our algorithm has linear computational complexity and is easily parallelisable to cover large areas. We use our method to compute the first oil palm density map with 10 m Ground Sampling Distance (GSD), for all of Indonesia and Malaysia and for two different years, 2017 and 2019. The maps have a mean absolute error of ± 7.3 trees/ha, estimated from an independent validation set. We also analyse density variations between different states within a country and compare them to official estimates. According to our estimates there are, in total, >1.2 billion oil palms in Indonesia covering >15 million ha, and > 0.5 billion oil palms in Malaysia covering >6 million ha.

Desert locust (Schistocerca gregaria sp.) invasion has frequently caused huge economic loss, threatened food security and population’s livelihood in the North African countries (Gómez, D. et. al, 2019). For instance, the locust plague that took place in 2004 had damaged 80% to 100% of the crop production in Burkina Faso, Mali, and Mauritania, leading to acute needs for external aids. As a result, an accurate early warning system is a critical challenge that can largely benefit from the development of open earth observation data. The current literature largely focuses on ecological niche modelling by considering potential habitat as a static state or monitoring locust plague with temporal dynamics of vegetation indices. The performance of the available swarm predictions, nonetheless, is limited by insufficient considerations of land use information, as well as the low representation of favourable conditions in the training data (Klein, I., Oppelt, N., & Kuenzer, C., 2021). For example, it is pointed out that vegetation indices such as NDVI cannot well represent other relevant factors such as land-use type, phenology and seasonal variations which play important roles in behavioral ecology of the locust (Bryceson, K.P, 1989, Ceccato, P., 2005, Waldner, F. et. al., 2015). This study aims to find out if the integration of Landsat spectral reflectance and FLDAS which utilizes different land models, rainfall, and meteorological inputs into a CNN model that can overcome these limitations and improve desert locust prediction in Ethiopia. To integrate more relevant variables into training the model, the present project applies SRTM DEM, all spectral bands of Landsat imagery, as well as Famine Early Warning Systems Network Land Data Assimilation System (FLDAS), which is a model making use of the available hydroclimatic observations and designated for food security assessment.

First, presence data samples of desert locust swarms from 1998 to 2005 are downloaded from FAO Locust Hub and resampled. After that pseudo-absence points are generated by creating a buffer for the locust plague that occurred in a specific year, ending up with over 20,000 presence and 30,000 absence data points over North Africa. Then, Landsat time series and FLDAS bands including evapotranspiration, surface pressure, specific humidity, soil heat flux, latent heat net flux, total precipitation rate, soil moisture, soil temperature at different depths, near-surface air temperature, and near-surface wind speed are retrieved at the month before and during the locust plague using cloud computing with Google Earth Engine. Additionally, DEM and long-term temperature data from WorldClim are retrieved for training. It is followed by the development of a CNN model using TensorFlow and the splitting of data points for training the machine learning model, as well as validation which applies only to a subset of data in Ethiopia. Finally, the results from the CNN model are compared to other machine learning models, such as SVM and random forest classifier.

The results of the CNN model achieved fair results with 75% accuracy, whereas the accuracy using random forest classifier is over 90%. Meanwhile, superior performance is observed using Landsat imagery instead of vegetation indices. This study explored the high potential of the FLDAS global model which leverages existing land surface models based on multiple land surface models and meteorological inputs in locust monitoring applications. The application of full spectral reflectance and FLDAS dataset has shown satisfactory performance in predicting locust swarms in Ethiopia. Hence, in addition to the conventionally used dataset such as WorldClim and NASA SMAP, there is a high potential of using full spectral reflectance and FLDAS model for locust forecasting, meanwhile, machine learning models such as random forest might be sufficient for precise prediction.

References

Bryceson, K.P. The Use of Landsat MSS Data to Determine the Locust Eggbeds of Locust Eggbeds in the Riverina Region of New South Wales, Australia. Int. J. Remote Sens. 1989, 10, 1749–1762.
Ceccato, P. Operational Early Warning System Using Spot- Vegetation and Terra-Modis To Predict Desert Locust Outbreaks. In Proceedings of the 2nd International VEGETATION User Conference, Antwerp, Belgium, 24–26 March 2005; pp. 33–41.
Gómez, D., Salvador, P., Sanz, J., Casanova, C., Taratiel, D., & Casanova, J. L. (2019). Desert locust detection using Earth observation satellite data in Mauritania. Journal of Arid Environments, 164, 29-37.
Klein, I., Oppelt, N., & Kuenzer, C. (2021). Application of remote sensing data for locust research and management—A review. Insects, 12(3), 233.
Waldner, F.; Babah Ebbe, M.A.; Cressman, K.; Defourny, P. Operational Monitoring of the Desert Locust Habitat with Earth Observation: An Assessment. ISPRS Int. J. Geo-Inf. 2015, 4, 2379–2400

Satellite imagery is now a widely used source of information in the concept of precision agriculture aimed at maximizing yields along with increasing the efficiency of crop production management and optimizing farmland inputs. Most precision agriculture applications based on EO data processing are mainly focused on monitoring crop health and its local variability. In addition to these uses, EO data also have great potential for assessing agri-environmental risks and their impacts. Here, they benefit from the fact that high-resolution satellite imagery covers long time periods (e.g. Landsats) with a relatively high temporal frequency of data collection (e.g. about 5 days in the case of Sentinel-2).

The main focus of our poster presentation is to present the developed system for monitoring several types of risks that can have a major impact on crop production. This includes risks of crop damage due to adverse conditions (e.g. increasing number of droughts, heat stress, pests and diseases, etc.), but also addresses risks related to soil management, such as a decrease in soil organic carbon content or contamination of groundwater sources with mineral nitrogen due to possible overfertilization of agricultural soils. The risk assessment is based on dynamic analysis of crop conditions using Sentinel-2 and Landsat data together with crop condition records provided by a network of automatic weather stations and IoT sensors.

The first thematic area focuses on the risks associated with the occurrence of adverse vegetation conditions that can lead to crop damage and reduced yields. Yield data were collected for about 3000 agricultural plots in the Czech Republic for the period 2016-2020. Time series of several vegetation indices obtained from Sentinel-2 data were analysed for their sensitivity for yield prediction and mapping of its local variability. Together with records of corresponding meteorological conditions (including parameters such as air temperature, precipitation, etc.) were used to train an ANN-based model for yield prediction. The trained algorithm is then applied to data from the current growing season using current Sentinel-2 scenes and data from our network of weather stations.

The second part of the project focused on long-term changes in soil organic carbon (SOC) content. During 2020, intensive in-situ soil sampling was conducted to obtain representative examples of different soil types exhibiting different levels of SOC. Laboratory measurements of SOC content and soil spectra were used to develop a simple ANN model for estimating SOC, which was further applied to bare soil composites (BSE). BSE is a specific multitemporal aggregation of Landsat and Sentinel-2 data showing the surface of agricultural land without any vegetation cover. These BSEs were calculated for the period 1985-2020 with a five-year frequency. Finally, the developed model applied to this BSE sequence is used to detect lands at risk of significant SOC loss.

The last topic of the demonstrated workflow concerns the level of nitrogen uptake by crops. It is based on the estimation of the nitrogen nutrition index (NNI) from Sentinel-2 data. The NNI is a useful indicator of nitrogen balance, which is further used to identify nitrogen surpluses and over-running nitrogen inputs to agricultural soils. These data, along with runoff regimes, are ultimately used to assess the risk of nitrogen contamination of local groundwater sources.

In addition to the algorithms for processing and analysing the satellite images, serious attention is also paid to the dissemination of the system outputs to its target users, which is carried out by a dedicated application integrated into the CleverFarm software platform. We hope that the presented approach can serve as a valuable source of information regarding agri-environmental risks not only for farmers, but also for other stakeholders such as state nature conservation authorities, etc.

Whereas precision agriculture is nowadays a consolidated practice in traditional crop growing, organic crops did not receive as much attention in terms of automated monitoring and decision support. Scientific state of the art in satellite-based monitoring of farmland, for example, reports very few cases -if any- of techniques developed specifically for organic crops, which possess different needs and show different patterns with respect to traditional agriculture. Although still a small fraction of total farmland, organic is set to grow significantly in the future, and this calls for specific research to address its peculiarities.
In this framework, we have jointly set up a test site designed to experiment simultaneous remote and proximal monitoring of organic crops. It consists of an organic vineyard located in the central Po valley in Northern Italy and intended for organic Lambrusco wine production. The vineyard extends over 1.2 hectares and is equipped with a sensor network counting 8 devices homogeneously distributed plus a gateway collecting measures via a LoRa protocol. Sensors enable real-time acquisition of soil humidity and temperature at a depth of 25-30 cm every 5 minutes. All data converge to a data center, and are accessible e through a dedicated web interface at the following URL: http://wine.easyiot.it/
In addition to environmental parameters, received signal strength indicator (RSSI) from each station is also recorded. The value of RSSI is linked to vegetation parameters such as leaf density and water content, that are relevant in a crop monitoring context. These values are published via the GitHub repository at https://github.com/emanueleg/lora-rssi and discussed in [1]. In addition to sensor data, pictures of the vineyards were taken along the entire growing season at regular intervals of one to two weeks, to identify the changes in phenological phases of the monitored vine. On the side of spaceborne data, dense time series of Sentinel-1 and -2, COSMO/SkyMed, SAOCOM data were built on the test site thanks also to an ongoing project from the Italian Space Agency (ASI).
As shown in previous work [2], dense satellite time series can be used for example to estimate features of the vineyard output; in our case, we focused on relationships among the crop status and spaceborne and in situ parameters.
In the experiment to be presented at LPS 2022, we compared local and spaceborne parameters, investigating mutual links and correlations. At the conference we will present some interesting connections found in the parameter set, that may be useful to set up novel monitoring techniques specially designed for organic crops.

[1] E. Goldoni, L. Prando, A. Vizziello, P. Savazzi, P. Gamba, Experimental data set analysis of RSSI-based indoor and outdoor localization in LoRa networks, Internet Technology Letters 2 (1) (2019) e75, https://doi.org/10.1002/itl2.75.
[2] F Dell’Acqua, D De Vecchi, GC Iannelli: “The ESA ARTES Demonstration Project "Saturnalia": forecasting wine quality from space seemed impossible, it is now an operational downstream service”, European Space Agency Phi-Week 2020 Frascati, Rome, Italy, 28 Sept - 2 Oct 2020.

The aim of this paper is to map agricultural crops by classifying satellite image time series. Domain experts in agriculture work with crop type labels that are organised in a hierarchical tree structure, where coarse classes (like orchards) are subdivided into finer ones (like apples, pears, vines, etc.). We develop a crop classification method that exploits this expert knowledge and significantly improves the mapping of rare crop types. The three-level label hierarchy is encoded in a convolutional, recurrent neural network (convRNN), such that for each pixel the model predicts three labels at different level of granularity. This end-to-end trainable, hierarchical network architecture allows the model to learn joint feature representations of rare classes (e.g., apples, pears) at a coarser level (e.g., orchard), thereby boosting classification performance at the fine-grained level. Additionally, labelling at different granularity also makes it possible to adjust the output according to the classification scores; as coarser labels with high confidence are sometimes more useful for agricultural practice than fine-grained but very uncertain labels. We validate the proposed method on a new, large dataset that we make public. ZueriCrop covers an area of 50 km x 48 km in the Swiss cantons of Zurich and Thurgau with a total of 116'000 individual fields spanning 48 crop classes, and 28,000 (multi-temporal) image patches from Sentinel-2. We compare our proposed hierarchical convRNN model with several baselines, including methods designed for imbalanced class distributions. The hierarchical approach performs superior by at least 9.9 percentage points in F1-score.

In the Mediterranean regions, the succession of drought periods threatens the water resource and food security. In this study,1) we analyze the potential of different drought indices to identify drought periods for two regions with different climates: Kairouan in Tunisia, and Catalonia in Spain, and 2) we evaluate the effect of these periods in the vegetative cycle of cereals, and we identify the indices that give more accuracy for cereal yield prediction.
To attend to the objectives of this study, satellite data was used, MODIS NDVI, MODIS LST, MODIS ETP, and SMOS. In this context, spatial resolution improvement algorithms have been applied such as DISaggregation based on Physical And Theoretical scale Change (DISPATCh), to solve the spatial resolution problem of SMOS from 40 km to 1 km.
In this study, we analyze the potential of two droughts indices:1) Soil Moisture Anomaly Index (SMAI) calculated from SM DISPATCh data, which gives an idea of the soil water status and 2) Evapotranspiration Anomaly Index (EAI) calculated from MOD16A2, which reflects the vegetative activity and indirectly the health status of the plant.
We identified drought periods from the 2010/2011 to 2019/2020 agricultural year for EAI and SMAI. We found the existence of water stress periods detected by SMAI and corresponding to wet periods in EAI. In this case, we talk about the distinction between vegetative drought and soil moisture drought. With the SMAI results, we note more dry periods which is normal because of the influence of solar radiation on the vegetation cover, we will mostly have positive anomalies and less stressful periods. For the study area in Tunisia, the strong correlation obtained between yield and SMAI is in December (R² = 0.88), this results highlights the importance of water during this period. To estimate the yield, it can be predicted only from January, when the crop cover has started to appear. The correlation between yield and SMAI decreases slightly in January (R²=0.4), February (R²=0.54), and March (R²=0.3). A stronger correlation, but later, with EAI in March (R²=0.58) and April (R²=0.64). For the second study area in Catalonia, Lleida, we find a different behavior. For wheat, the strong correlation between yield and EAI occurred in February (R²= 0.68) followed by a correlation equal to 0.96 with the SMAI in March. Similarly, for barley but with a stronger correlation with EAI (R² =0.8) in February.
In this study, we have shown the advantage of improving the spatial resolution of SM to better study drought and especially its effects on yield. As a conclusion, the study of the effect of drought on cereals yields is very useful for anticipating the measures to be taken by the state to supply the market and avoid food insecurity, and by the farmer to improve productivity.

In Hungary, cereals are typically grown on about 2.5 million hectares each year, with the main crops being maize and winter wheat. Due to climate change and basin-specific meteorological characteristics, extreme events related to weather and water balance are becoming increasingly frequent and have a significant impact on crop production. In the present work, we investigate drought impacts on maize and winter wheat by joint analysis of satellite vegetation indices, meteorological factors and soil properties using time-series and multivariate statistical methods for the period of 2014-2021. Crop maps, produced annually by Lechner Knowledge Centre based on Sentinel-1, Sentinel-2 and Landsat 8 images by applying machine learning methods, were used to define the sample areas for the study period. For the phenological studies, NDVI, EVI2, TNDVI and kNDVI vegetation indices were derived from the Terra MODIS sensor 8-day integrated MOD09Q1 surface reflectance time series, aggregated to a 0.05° grid for data continuity. In this phase of the study, MODIS data were chosen because of the unparallelled, consistent and long-term archive. The vegetation index time series were used to investigate phenological variation among sample areas in the respective years. To study the influence of meteorological factors, cumulative precipitation sum, cumulative heat sum and cumulative soil moisture sum curves were derived for the growing season of the respective crop type using hourly and daily data from the ERA5-Land meteorological reanalysis database. The vegetation index curves were compared with county-level (NUTS-3) average yield values reported annually by the Hungarian Central Statistical Office and with the drought damage claims made by farmers. Preliminary results suggest that in some study years the variability of yield values among study areas can be well explained by differences in the time series of vegetation indices and meteorological factors, while in some years, such as in 2020 in case of maize, these relationships are less significant. To better understand the underlying causes, we plan to build multivariate models and to integrate terrain and soil parameters. Mid-term development plans include the extension of the study period further in the past, depending on the availability of crop maps. In the longer term, we plan to investigate the usage of Sentinel (2 and 3) imagery, possibly combined with MODIS and Landsat. Some of these additional results are also planned to be presented at LPS 2022.

Over the next few decades, climate change and socio-economic developments will induce to a high pressure on already inadequate water resources, mainly in arid and semi-arid regions, where water demand is growing high especially in the agricultural sector. Remote sensing is well suited to help the sustainable use of agricultural resources and water in particular. To this end, more efforts have to been made to get quantitative estimation of key variables, such water stress, evapotranspiration, or soil moisture.
Recent studies have highlighted the sensitivity of the radar backscatter for vegetation water stress detection, due to the direct backscatter of the canopy, and the bidirectional attenuation of the signal as it passes through the vegetation layer. The interaction between radar wave and vegetation depends mainly on the vegetation dielectric constant, which itself is mainly influenced by vegetation water content. Several studies have shown the sensitivity of radar signal to diurnal water cycle of vegetation over tropical Forest with spaceborne scatterometres, by observing differences between morning and evening acquisitions. On the other hand, in situ radar measurements experiments have allowed to analyze the radar response with high temporal frequency over tropical forest or maize crops.
A similar experiment has been developed over an olive orchard, drip irrigated, located in the semi-arid Mediterranean Chichaoua region, in central Morocco. It allows for the first time to get over crop field in situ C band radar measurements which are continuously acquired since January 2020 with a time-step of 15 min. This study analyzes the results obtained.
The radar device consists of 7 C-band antennas at the top of a 20 m high tower, that collect measurements at 4 polarizations (VV, VH, HV and HH). The olive orchard has been equipped by an Eddy-covariance system for measuring the convective fluxes, soil moisture and temperature profiles, Sapflow and micrometric variations of the olives trunks to monitor the physiological functioning of olives trees. The study site being visible by three different passes, six Sentinel-1 acquisitions are available every 10 days, allowing the comparison with in-situ radar measurements
Both backscattering coefficient and interferometric coherence temporal profiles are analyzed from in situ measurements. On a daily scale, the interferometric coherence |ρ| between two consecutive acquisitions (i.e. 15 min. baseline) shows a daily cycle similar to the one observed over tropical forest: Coherence values shows a daily amplitude up to 0.6, whatever the polarization configuration: they are high during the night, when wind and vegetation activity is low. They begin to decrease at 7 AM, to reach a minimum value (~ 0.4) at 7 PM, due to wind and vegetation activity. Then they increase to recover high values (~ 0.9). It is worth noticing that the beginning of the coherence decrease, between 7 and 9 AM, is not due to wind (which begins at 9 AM) but coincides with beginning of the plant activity and water flow transfers, which is confirmed by sapflows measurements Furthermore, coherence values are still significant (~ 0.4) for six days baselines when morning acquisitions are concerned. This result is promising for Sentinel-1 data. Future work will make it possible to clarify whether these can be useful for the detection of water stress.
The backscattering coefficient σ0 shows a similar diurnal cycle, although in opposition of phase, which is correlated to the daily evolution of the evapotranspiration, which is in turn linked to the Sapflow and to the micrometric variation of the olive’s trunk... σ0 is low and constant during the night, then increase after 7 AM to reach a maximum around 1 PM, and then decrease continuously until midnight. The daily amplitude depends on the season, from 2.5 dB in January to 1 dB in May.
on a seasonal scale, the temporal signatures of the interferometric coherence and backscattering coefficient, between the in-situ radar measurement and Sentinel-1 present a good overall correspondence, whit a high drop during the summer which corresponds to an increase in vegetation activity such us vegetation growth and high-water demand, and a slight decrease from winter to spring which corresponds to a decrease in vegetation activity due to its resting period.
These results will clarify the potential of C-band radar data, and especially Sentinel-1, for the monitoring water stress over tree crops.

With the launch of CubeSat constellations, it is now possible to acquire near-daily image coverage at high spatial resolution. The high spatiotemporal resolution of CubeSats is providing unprecedented capabilities for the provision of actionable information suited for treatment, irrigation, and harvest scheduling of crops. However, establishing a link between the image data and crop phenology is key to understanding the requirements of crops in terms of treatment amounts and timing for yield optimization. Despite the near-daily image availability of PlanetScope CubeSat data, cloud cover and radiometric inconsistencies between sensors on different platforms preclude direct derivation of biophysical parameters such as leaf area index (LAI). In this research, we used CubeSat data harmonized based on Landsat-8 and Sentinel-2 data to at surface reflectance, cloud-masked based on temporally driven cloud and cloud shadow detection, and gap-filled based on multi-sensor observation data acquired both before and after the prediction date. From this time-series of daily gap-filled and radiometrically robust CubeSat data of 3 m pixels, a hybrid inversion method using machine-learning and physically-based understanding was employed to produce daily cloud-free LAI images throughout the growing season of one rain-fed and two irrigated maize fields located in Nebraska, USA. Here, we demonstrate how daily CubeSat-derived LAI measurements and maize plant phenology are related and showcase how CubeSat-based LAI measurements can be used to detect within-field variability and facilitate treatment and irrigation scheduling based on associated crop phenological characteristics. The image-derived results showed a high correlation against the field measurements of LAI data collected for all three fields throughout the growing season (coefficient of determination = 0.94 - 0.97; root mean square error = 0.42 - 0.73 of LAI). Intra-field patches with low LAI values could be recognized from early plant growth. The rainfed field was found to lack about four days behind the two irrigated fields in terms of LAI values and reached maximum LAI values that were slightly lower than those of the two irrigated fields. The vegetative stage of the maize plants was associated with increasing LAI values, which reached a maximum level around the time of silking and the beginning of kernel formation. As deficits of nutrients and water can have major effects on yield during and before silking, the daily LAI values derived from the CubeSat imagery may be used for treatment and irrigation scheduling. LAI values remained at a high level for most of the reproductive stage, i.e. the kernel development stage, but started to decrease towards the time of physiological maturity and maximum biomass accumulation. Senescence and an associated decrease in LAI values were identified for the last two reproductive stages. Maize used for cattle feed is generally not harvested until the plant material has fully dried and is suitable for silo storage. Hence, the time-series of LAI values was also found suitable for harvest scheduling. The high spatiotemporal resolution of the CubeSat data can provide unique insights into the phenology of crops and deliver actionable intelligence to facilitate food production.

Timely monitoring of agricultural production and early yield predictions are essential for food security. Remotely sensed time-series allow to capture the variability in crop growth and yield across cropping systems. However, most studies leveraging remote sensing data for yield estimation use moderate resolution data which has been aggregated over large areas and is therefore not suitable to be used at the field management level. Yield estimation at the field level using remote sensing or crop modelling has received less attention and few studies have combined these two approaches to produce yield maps over large areas at the field level. Our study uses remotely sensed time series from different sources to estimate yield using both, regression-based approaches and dynamic crop simulations, to gain a better understanding of how field-level yield mapping can be improved. We focus on assessing the performance of multiple satellite data streams, namely PlanetScope, Sentinel-2, and Landsat 8, to investigate their spatial, spectral (surface reflectance and vegetation indices), and temporal characteristics to estimate yield for maize, winter wheat, and winter barley in Germany. Reference data was collected by farmers during the years 2017 to 2021 and is available at the field-level (weighbridge) and within-field-level (grain yield monitor tracked by GPS). In addition, field boundaries, information about crop types, and crop management were collected. Our field level analysis used data from different years to train a Random Forest regression model, which was tested on a year previously unseen by the model. The within-field analysis uses data from the same year for training and testing. The MOdel for Nitrogen and Carbon in Agroecosystems (MONICA) was used to simulate leaf area index and yield time series. Scalable Crop Yield Mapper (SCYM) approach was used to simulate yield at pixel and field-scale. The leaf area index and yield output from crop model simulations using MONICA were used to train regression models. The most important predictors for explaining within-field yield variability when using PlanetScope data were the bands green and NIR, whereas the red-edge, as well as near and shortwave infrared bands and VIs were ranked higher in importance when using Sentinel-2 and Landsat 8 data.

Tillage monitoring in the Netherlands using Sentinel-1 and Sentinel-2 observations
Abstract - A3.04 Agriculture - Methods and Algorithms, Science, Applications and Policy

Monitoring tillage practices has become of great importance in agricultural and environmental sciences due to its impact on the soil properties, soil erosion, evaporation processes, infiltration, carbon storage, nutrient uptake, biodiversity and crop production itself.
Tillage is a common soil management technique and is used to break the soil to decrease the bulk density in order to support root development. Still, conventional tillage, which includes intensive and deep tillage, has shown negative impact on the crop production and soil properties compared to shallow or no tillage, so called conservation tillage.
Hence, monitoring the intensity, depth and frequency of tillage is of vast important to optimize and secure crop development, soil erosion and soil health. Especially in countries such as the Netherlands, which is the second largest exporter of food and agriculture, it is necessary to understand the past and current farming practices to maintain and improve soil health and resilience to extreme weather conditions. Further derivatives of this monitoring system, for example improved carbon sequestration assessment, can be used to support governmental sustainability goals and policies set by the European commission and common agricultural policy (CAP).
Considering large-scale mapping of tillage and crop management practices, it is not feasible to rely on in-situ measurements. For example, the average amount of individual crop parcels within the Netherlands covers more than 770.000 fields, which makes it a serious challenge to communicate and manage the individual crop management information. Thus, the suggested solutions is based on remote sensing observations, which allow temporal and spatially consistent monitoring.
In this study, the goal is to investigate a potential monitoring system for tillage practices of farmers. The hypotheses of this effort is the sensitivity of interferometric coherence to structural and biomass changes. SAR Interferometric Coherence is a product derived from the Single-Look-Complex (SLC) data describing the similarity between two complex consecutive SAR images. The SLC data is a product from the Sentinel 1 SAR mission, with a temporal resolution of 6 days. This high temporal resolution allows monitoring crop dynamics and individual farmer practices. Many previous studies show and highlight the potential use and sensitivity of the interferometric coherence signal for crop-type mapping, growing season estimations, soil moisture estimations and crop phenology monitoring.
Still, none of these studies has created a novel approach to detect tillage over a wide range of different crop types in a potentially operational manner. The suggested methodology is based on a combination of multiple channels and products of Sentinel 1 and Sentinel 2. Sentinel 1 and Sentinel 2 are both orbiting in a constellation of two satellites, which allows a high temporal resolution of at least 6 days covering the microwave and optical range of the electromagnetic spectrum. The independence of SAR data to cloud coverage and sun illumination additionally guarantees a consistent monitoring product. The data used for this project was derived from the Agricultural SandboxNL and further expanded by Sentinel 2 and interferometric coherence signatures.
First results show a great potential to identify tillage dates over a range of different crop types. S1-derived interferometric coherence allowed identifying different crop stages, whereas Sentinel 1 GRD and Sentinel 2 signatures are further used to separate and classify the identified events. Eventually, the work can be further expanded to create yearly tillage occurrence maps over large areas of interest.

Weather and climate, among other factors, impact yield and production of crops. Farmers managing crops need to anticipate the consequences and interplay of these factors across a growing season. However, if this fails, it can lead to large fluctuations in production and a loss of resilience with societal consequences. Crop diversification could be a viable strategy to increase crop production resilience, with beneficial effects for biodiversity and farmers’ incomes, but also for food security on larger scales. It is not well known how diversification and resilience are connected across spatial scales. Diversification might be achieved by considering different crops in spatially disconnected landscapes (γ-diversity). This could improve the larger scale production stability of for instance cereal crops. However, crop diversity in the same landscape (α-diversity) might be effective to improve resilience at the farm level. Here, a more diverse set of crops could slow the spread of pests and diseases, thus lowering the need for agri-chemical use, while enhancing agro-ecosystem services and farm income. As a first step towards understanding these factors, we take advantage of a new 2018 European Union wide high-resolution (10 m) land-use dataset based on LUCAS in-situ data and Sentinel-1 Synthetic Aperture Radar observations with specific information on crop types. We analyse crop diversity computed at different incremental spatial scales ranging from 1 km to 100 km. This analysis suggests that complementary information is derived by using two statistical metrics, one for the overall crop diversity computed at the larger spatial scale and one characterizing the scale at which a significant portion of diversity is diagnosed. These two metrics can be potentially used to quantify these complementary aspects of diversification across the Member States of the European Union. The availability of frequent and high resolution Earth Observations opens avenues for innovative data-driven indicators with relevance in the context of the Common Agriculture Policy.

ForestMind: Actionable insight for a sustainable forest-commodity future

ForestMind, enacted at the request of members of the UK food supply chain, is a two year £4 million project funded by the European and UK Space Agencies. Led by Satellite Applications Catapult and partnered with a consortium of 10 UK industry, innovation and research partners, we are focused establishing a new neutral and trusted entity that delivers actionable insight and support to validate and change sourcing decisions, and guide supplier development to reduce deforestation. We do this through leveraging specialisms in AI, remote sensing, economics, traceability, enforcement and sustainability from across the consortium.

We are first focused on enabling commodity sourcing organisations to understand and act on deforestation risks in supply chains globally and plan to address broader environmental and social risks beyond the project for all parties within the supply chain from farmer to customer and covering:
- Enacting legislative compliance
- Demonstrating commitment progress
- Oversight of supply chains
- Managing reputational risk
- Verifying supplier claims
- Supporting environmental protection

ForestMind is not about re-creating, but integrating and scaling propositions in a user driven approach. Our project partners bring reputation, expertise and capabilities essential to delivering the actionable insights that validate and enable supply chain decisions and we have the support of leading supply chain organisations. We want to work with customers who want to deliver a future prosperous and biodiverse global forest commodity ecosystem and want ForestMind to enable users across supply chains to have information and education to make profit AND make positive impacts on the ecosystem

We want to go far beyond the ESA funding, and rapidly become multi-commodity and multi-geography and are first focused on enabling commodity sourcing organisations to understand and act on deforestation risks in supply chain, and work with their suppliers to deliver benefits and will offer broader environmental and social assessments for all parties within the supply chain from farmer to customer.

Rice is one of the major commodities traded in the international food market and a basic food in the diet of more than half of the world’s population, especially in underdeveloped or developing countries. Rice is a crop of economic importance in Spain, that contributes to 28% of the European rice production, being the second major rice producing country in Europe. Therefore, an accurate and timely crop yield forecasts are necessary for making viable agricultural investments, developing proper agricultural planning, increasing market efficiency and stability, as well as manage food shortages. On the other hand, fertilizers and pesticides are routinely used in rice cultivation to maintain optimal yield and to protect plants from diseases. However, the overuse of fertilizers and pesticides, has adverse effects on the environment and human health, that has led to the regulation of the use on nitrogen in agriculture to minimize its impact. Fertilization management in modern agriculture aims to supply just the fertilizer needed to maximize yields while avoiding their excesses to ensure environmental sustainability.
Precision agriculture is defined as the management of crops based on the knowledge of the spatial and temporal variability in an agricultural holding, to improve the economic return and minimize the environmental impact. Earth Observation data provides timely, objective and accurate information that is critical for precision agriculture applications. In this work we study two main applications of precision agriculture in rice crops based on Sentinel-2 and drone data. Particularly, this study explores the use of remote sensing data to monitor rice crop yield and nitrogen content in the plant, allowing more efficient, sustainable and profitable rice management practices.
First, we develop a rice yield model at field scale based on yield measurements over 70ha/year around the Albufera in Valencia (Spain) from 2017 to 2020 seasons. The dataset covers the two main rice varieties, that is Bomba and JSendra, and provides average yields at field scale. Training and validation follows a structure of k-fold cross validation. Preliminary results show that linear regression models that relate Sentinel-2 spectral bands to the yield data at field scale provide Root Mean Square Errors (RMSE) of 0.27 t/ha (6.2%) in Bomba and 0.54 t/ha (5.1%) in JSendra.
The second objective of this work is to analyse rice crops with different nitrogen level treatments using Sentinel 2 and drone data carrying a multispectral camera (RGB and NIR bands) and a thermal camera (monochannel 7.5-13.5 µm). This is conducted over an experimental field close to the Albufera that was split into three different areas characterized by different nitrogen treatments: low, medium and high. Field data on the within-field yield was also collected by harvesting machines. This study will determine the best band/combination of bands that can monitor the rice development over the different nitrogen treatments and how they correlate with the final yields.

The main objective of SYKE's Valumavesi-project is to produce scientific knowledge, methods and tools to support sustainable water management in agriculture and forestry in changing environment. As a part of the project, remote sensing and modelling methods are studied for estimating drainage status and flood risk of agricultural fields. This study presents the preliminary results for using Sentinel-2 data to estimate soil moisture, which is an essential parameter to evaluate the drainage status of the fields. The short-term aim of the KUTI-work package is to produce an estimate of soil moisture for pilot drainage basin of Loviisanjoki River, and long-term aim is to produce a nationwide process for operational soil moisture monitoring.
Two different sets of in-situ measurements of soil moisture were used:
• Finnish Field Drainage Association (https://www.salaojayhdistys.fi/) has test area at Sievi, Western Finland. There are sensors measuring soil moisture at depths 15, 40 and 70 cm at two different locations (two sensors at both locations) in same agricultural field plot. Time period of Sentinel-2 observations was 26.6.-10.9.2021. Due to lack of cloud free satellite imagery, there were 62 samples in this dataset.
• Agricultural fields of Carbon Action Field Observatory (CAFO) program (https://www.fieldobservatory.org/). 30 agricultural fields in Southern Finland were selected for analysis, each having two soil moisture sensors with measurement depth 7.5 cm. So far, data from 12 fields with large number of cloud free satellite images have been used resulting 140 samples. Time period of Sentinel-2 observations was 9.5.-13.8.2021.
Atmospheric correction was done to Sentinel-2 images using Sen2cor-software. Cloud mask was done using Idepix-algorithm (https://github.com/bcdev/snap-idepix) with locally programmed Shadowmasker-software which finds the cloud shadows. These tools were implemented to Calvalus-processing cluster (https://github.com/bcdev/calvalus2) at Finnish National Satellite Data Centre. This processing cluster was also used to extract the reflectance values from images for the Sievi test points and to compute average reflectance values for fields of CAFO-program.
So far, in-situ measurements of soil moisture have been compared to Sentinel-2 bands and different band indices by computing correlation between them. In the future, empirical models for soil moisture estimation will be created using the features with best correlations. Also, the use of OPTRAM-model will be tested.
Soil moisture measurements were collected automatically, in 0.5 – 1 hour interval depending on sensor. Soil moisture data were generalized by computing the average values between 8 and 12 o’clock for each day.
Correlations were computed between average soil moisture and Sentinel-2 bands and indices computed from these bands. In the case of CAFO-data, the correlations were determined using all observations or observations divided according to NDVI-value to see the influence of vegetation. NDVI-groups were A. NDVI ≤ 0.4, B. 0.4 < NDVI < 0.6 and C. NDVI ≥ 0.6. When correlations were computed with Sentinel-2 bands, the largest correlations (CC) were acquired with middle infrared bands B12 and B11:
• All observations: CC 0.61, Band 12
• NDVI A: CC 0.77, Band 11
• NDVI B: CC 0.68 Band 12
• NDVI C: CC 0.40 Band 12
Different kinds of image indices (approximately 150 of them) were collected from literature and tested. E.g. all Simple Ratio and Normalized Difference Indices were computed and tested. The largest correlations were acquitted with following indices:
• All observations: CC -0.67, Visible and Shortwave-infrared Drought Index VSDI_B12 = 1 - ( B12 + B4 - 2*B2 )
• NDVI A: CC 0.83, Normalized Difference Index NDI_B11_B5 = ( B11 - B5 ) / ( B11 + B5 )
• NDVI B: CC 0.68, Modified Shortwave-infrared Perpendicular Drought Index, MSPDI_B12 = (( B12 + B4 ) + M * ( B12 – B4 )) / SQRT( 1 + M^2 )
• NDVI C: CC 0.58, Simple Ratio SR_B8_B7 = B8 / B7
In Sievi, there were some differences between measurement set-up to study impacts of different actions. Firstly, there are only two observation sites, but with different measurement depths and it is known that mowing of timothy grass took place during early August. Also, the average soil moisture measurements of different days were correlated with Sentinel-2. Concerning depth, the correlations were highest with 15 cm measurement depth and lowest with 40 cm depth. The correlations were much higher before mowing of timothy grass, and lower afterwards. This indicates that soil moisture estimation should be done before any mowing, cutting or ploughing operation. Before mowing, there is no noticeable difference between in-situ measurement time and Sentinel-2 acquisition time correlations, in other words correlations are quite same if time difference is one or three days, but after mowing this difference is noticeable in a sense that there is much more variation in values of correlation.
The tendency towards higher correlations prior to mowing seems to suggest higher applicability of this method to fields with plant cover. This could be due to the slower responses of the plants to changes in soil moisture, compared to those of bare soil, or soils with threshing waste. This discovery is promising to agricultural applicability since the changes in soil moisture during the growing season are those with the most agronomic relevance.

The EU Green Deal, Europe’s ambitious strategy to become the first climate-neutral continent by 2050, is working among others to transform the agriculture sector from a carbon emitter to a promising carbon sink. Under the EU Green Deal the practice of carbon farming, in particular, has been singled-out to play a key role in achieving these EU climate goals.
Earth Observation data is recognized as an important source of information to monitor the impact of agricultural management practices over time and support quantifying its effects on the reduction of greenhouse gas emissions. Today’s high spatial and temporal resolution satellites offer the possibility to monitor carbon farming processes over vast areas and scale up information which is collected at field level. This can be used both to identify gaps and drive the decision-making process in the agricultural sector towards reaching emission reduction targets.
In this context, the EU-funded Horizon 2020 project AgriCapture (https://agricaptureco2.eu/) is currently developing a systematic, robust, and flexible platform for quantifying, verifying, and promoting soil carbon capture. To do so, it uses multiple data streams from Earth Observation, machine learning, and field-level soil sampling to explore and quantify how such practices can be measured and verified for regenerative agriculture and carbon credit certification schemes. Besides open Sentinel data, the project explores the potential of high-resolution, high-cadence imagery from Planet’s commercial satellites. More specifically, in this project we are evaluating the use of a next generation analysis-ready imagery product, called Planet Fusion, for mapping and monitoring crops and the tracking the application of regenerative agricultural management practices. Planet Fusion synergizes inputs from PlanetScope and public sensors, including Sentinel-2, to produce daily cloud-free, high resolution, temporally consistent, radiometrically robust, harmonized and sensor-agnostic surface reflectance feeds.
In this session, we present the published results of initial experiments with Planet Fusion data. In our first set of experiments, we aimed to observe the impact of this new type of data on crop classification. For this, we selected a variety of crop fields around Europe and analyzed the biophysical parameters for multiple years with daily cadence. These analyses allowed us to understand the inter-annual and cross-field similarity of the crops and to develop a machine learning (ML) approach to identify these crops. We evaluated the performance of the proposed ML approach by using in-situ data from the project’s pilot farms.

The growing status and condition of sugarbeets during growing season is of outstanding economic importance for sugarbeet processing industry to plan production. Therefore, operational monitoring on parcel-level is necessary to provide sugar producers with accurate information about production-relevant events like date of sowing, stands closure or harvest. The status quo are still costly and labor-intensive field inspections to record the important phenological phases. This is partly done by reporting from the farmers or the sugar producers send their own staff into the fields.

Sentinel-1 (S1) and -2 (S2) data play an important role here, as they can provide high-resolution information on vegetation stocks in terms of time, space and spectral resolution. Recent remote sensing studies (Khabbazan et al., 2019; Kavats et al., 2019; Shang et al., 2020; Amherdt et al., 2021; Meroni et al., 2021) are mainly using S1 products in particular to extract important cultivation dates. It’s common practice to use time series of different polarizations, backscatter coefficients (e.g., σ^0 ) and/or coherence products to detect abrupt changes in the signal and relate them to the target agricultural management events. Simple rule-based classification approaches are used to extract dates like emergence, row closure, and harvest (Khabbazan et al., 2019). In this study, important cultivation events, like seed bed preparation, sowing, and harvesting as well as the row closure of the stands were derived, by coupling S1 and S2 time series with phenological information and metrics using a model-based clustering approach. The entire process chain is automated on the German national processing cloud platform CODE-DE (Storch et al., 2019) using Python and R.

The S1 data were processed to backscatter γ^0 for both polarizations VV and VH, using CODE-DEs CARD-BS processor (CODE-DE, 2021). S2 time series were cloud masked based on L2A data (Sen2Cor). All satellite data are resampled to 10 m GSD. Phenological phase data (Gerstmann et al., 2016; Möller et al., 2020) are stored in data cubes on an in-house JKI-DataCube and are retrieved on-the-fly via WCS.

In the following, we outline our approach on parcel-level (Fig. 1): A sugar beet field is selected and passed to the processing chain. The corresponding pixels are extracted from the S1 and S2 time series and the median is calculated for each available acquisition date. The SAVI (Huete, 1988) is calculated from the S2 spectral bands and the time series is filtered using the BISE approach (Viovy et al., 1992). In the next step, the time series (backscatter from S1 and SAVI from S2) are interpolated to provide data for each day of the year. Phenological phase data are retrieved to get information about the growing season and phenological phases for sugar beets within the geographical region.

Figure 1: Workflow for the derivation of production dates for sugar beet. (see uploaded image: Figure_Workflow-BeetScan.png)

The actual detection of the production dates is based on the model-based clustering approach mclust (Scrucca et al., 2016), with which significant changes in parameter time series can be detected. The cluster analysis is performed within dynamic temporal windows defined with the help of the phenological phases. The selection of most appropriate phenological windows, BISE tuning parameters and explaining S1 and S2 parameters is the result of an optimization procedure validated by observed production dates. The optimization result can be considered as a query using relative cluster minimum or maximum values as well dynamic phenological windows, which ensures the transferability of the approach. The resulting query enables the detection of year- and parcel-specific production dates, which is exemplary shown for sugar beet parcels in Lower Saxony, Germany and is implemented as web service. Compared to classical phenological metrics based on vegetation index times series (Gao & Zhang, 2021), the suggested methodology enables the mutual usage of optical and radar-based parameters and leads to a more accurate detection of sugar beet production dates.

References
Khabbazan, S., Vermunt, P., Steele-Dunne, S., Ratering Arntz, L., Marinetti, C., van der Valk, D., Iannini, L., Molijn, R., Westerdijk, K., & van der Sande, C. (2019). Crop Monitoring Using Sentinel-1 Data: A Case Study from The Netherlands. Remote Sensing, 11 (16), 1887. https://doi.org/10.3390/rs11161887

Kavats, O., Khramov, D., Sergieieva, K., & Vasyliev, V. (2019). Monitoring Harvesting by Time Series of Sentinel-1 SAR Data. Remote Sensing, 11 (21), 2496. https://doi.org/10.3390/rs11212496

Shang, J., Liu, J., Poncos, V., Geng, X., Qian, B., Chen, Q., Dong, T., Macdonald, D., Martin, T., Kovacs, J., & Walters, D. (2020). Detection of Crop Seeding and Harvest through Analysis of Time-Series Sentinel-1 Interferometric SAR Data. Remote Sensing, 12 (10), 1551. https://doi.org/10.3390/rs12101551

Amherdt, S., Di Leo, N. C., Balbarani, S., Pereira, A., Cornero, C., & Pacino, M. C. (2021). Exploiting Sentinel-1 data time-series for crop classification and harvest date detection. International Journal of Remote Sensing, 42 (19), 7313–7331. https://doi.org/10.1080/01431161.2021.1957176

Meroni, M., D’Andrimont, R., Vrieling, A., Fasbender, D., Lemoine, G., Rembold, F., Seguini, L., & Verhegghen, A. (2021). Comparing land surface phenology of major European crops as derived from SAR and multispectral data of Sentinel-1 and -2. Remote Sensing of Environment, 253, 112232. https://doi.org/10.1016/J.RSE.2020.112232

Storch, T., Reck, C., Holzwarth, S., Wiegers, B., Mandery, N., Raape, U., Strobl, C., Volkmann, R., Böttcher, M., Hirner, A., Senft, J., Plesia, N., Kukuk, T., Meissl, S., Felske, J.-R., Heege, T., Keuck, V., Schmidt, M., & Staudenrausch, H. (2019). Insights into CODE-DE – Germany’s Copernicus data and exploitation platform. Big Earth Data, 3 (4), 338–361. https://doi.org/10.1080/20964471.2019.1692297

CODE-DE. (2021). Sentinel-1 l2 backscatter (bs). Retrieved November 26, 2021, from https://code-de.org/de/portfolio/25?q=Sentinel-1

Gerstmann, H., Doktor, D., Gläßer, C., & Möller, M. (2016). PHASE: A geostatistical model for the Kriging-based spatial prediction of crop phenology using public phenological and climatological observations. Computers and Electronics in Agriculture, 127, 726–738. https://doi.org/10.1016/j.compag.2016.07.032

Möller, M., Boutarfa, L., & Strassemeyer, J. (2020). PhenoWin – An R Shiny application for visualization and extraction of phenological windows in Germany. Computers and Electronics in Agriculture, 175, 105534. https://doi.org/10.1016/j.compag.2020.105534

Huete, A. (1988). A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment, 25 (3), 295–309. https://doi.org/10.1016/0034-4257(88)90106-X

Viovy, N., Arino, O., & Belward, A. (1992). The Best Index Slope Extraction ( BISE): A method for reducing noise in NDVI time-series. International Journal of Remote Sensing, 13 (8), 1585–1590. https://doi.org/10.1080/01431169208904212

Scrucca, L., Fop, M., Murphy, T. B., & Raftery, A. E. (2016). mclust 5: Clustering, Classification and Density Estimation Using Gaussian Finite Mixture Models. The journal, 8 (1), 289–317. http://www.ncbi.nlm.nih.gov/pubmed/27818791%20http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC5096736

Gao, F., & Zhang, X. (2021). Mapping Crop Phenology in Near Real-Time Using Satellite Remote Sensing: Challenges and Opportunities. Journal of Remote Sensing, 2021, 1–14. https://doi.org/10.34133/2021/8379391

Crop calendars translate the crops' phenology to a format we can easily understand, such as the Start of Season (SOS) and the End of Season (EOS). They represent essential information needed for crop type detection, crop monitoring, and crop yield estimation. Existing crop calendar products are only available at national and subnational scales and do not offer spatially explicit information on the variability in crop phenology in function of environmental conditions. WorldCereal is a European Space Agency (ESA) funded project whose main objective is generating algorithms for global classification of cropland, crop irrigation, and maize and wheat crop types at a resolution of 10 m. In the framework of this project, crop calendars are required first to develop robust models that can be implemented at global scale and second to trigger the algorithms within one month after the EOS. However, existing crop calendar products do not meet the requirements needed which should rely on continuous crop calendar data at a finer spatial resolution and at global scale. In this study we created new Baseline SOS and EOS crop calendars covering summer and winter cereals that integrate the four main existing crop calendar products: the Group on Earth Observations' Global Agricultural Monitoring (GEOGLAM) Crop Monitor, the United States Department of Agricultural Foreign Agricultural Service (USDA-FAS), the Food and Agriculture Organization (FAO), and the Joint Research Center's Anomaly hot Spots of Agricultural Production (ASAP). Using as input these baseline maps, we trained a random forest algorithm with climatological data (i.e. seasonal average and standard deviation of the air temperature, dew-point temperature, and precipitation during the period 1989-2019); geographic information such as latitude, longitude and elevation; and the distances between the areas with a the greatest percentage of crop in each baseline unit. The results were validated at polygon level using some baseline data set aside for validation purposes providing a R² for the SOS (EOS) of 0.87 (0.92) and RMSE of 27 (26) days for winter cereals and R² of 0.88 (0.63) and RMSE of 24 (26) days for summer cereals. Additionally, remote sensing data was used to evaluate the proposed maps. To this end, we computed Land Surface Phenology (LSP) metrics (i.e. SOS and EOS) over the training database from WorldCereal, which includes Sentinel-2 and Landsat-8 observations over reference data where the crop type is known. The LSP validation shows a SOS (EOS) R² of 0.68 (0.88) and RMSE of 27 (18) days for winter cereals and a R² of 0.7 (0.82) and RMSE of 27 (25) days for summer cereals. The crop calendars generated explain the phenological variability of winter and summer cereals on a global scale successfully. Hence, they will be used in WorldCereal project to define Agro Ecological Zones, areas where the SOS and EOS of each seasonal crop are expected to be similar, and to normalize the seasonal response of the crops throughout the world.

Maize is the dominant feed crop in Dutch dairy farming, due to its high protein content and metabolisable energy value. Monitoring crop development throughout the season can provide crucial information for maize yield estimation, and its relation to weather-related stress. Maize development can show substantial spatial and interannual variability within the Netherlands due to variable weather conditions. In addition, management decisions and legislation affect sowing and harvest times; for example, since 2019 maize cultivated on sandy and loess soils needs to be harvested by 1 October to ensure timely planting of a winter cover crop for reducing nitrogen leaching. Despite a general idea of broad crop calendars, and the recent progress in making high-resolution large-scale phenology products, at country and continental-scales still little information exists on how sowing and harvest dates for specific crops vary in time and space. The objective of this study is to assess the spatial and inter-annual variability of maize phenology for the years 2017-2020 and to evaluate to what extent the variability is driven by weather conditions, management, and legislation.

We used surface reflectance products from Sentinel-2 to derive vegetation index time series using the Google Earth Engine (GEE) platform; a separate series was made for each field known to be cultivated with maize using a quality-controlled national open dataset on farmer-reported crops. For each time series, we first used the weighted Whittaker method to smooth seasonal signals of Enhanced Vegetation Index (EVI) and Normalized Difference Vegetation Index (NDVI) and divide the year into growing seasons from the smoothed signals. Subsequently we fitted piecewise logistic functions to the time series of each maize growing period. Following the function fitting, we estimated sowing- and harvest-date (SD and HD) for each field using a threshold approach. The Sentinel-2 estimated SD and HD compared reasonably for 12 locations across the Netherlands for which ground reference data was collected in the three growing seasons between 2018 and 2020 (RMSE[SD] = 9.9 days and RMSE[HD] = 14.8 days). We then aggregated the estimated SD and HD of maize fields into 10 x10 km grid cells by calculating their weighted mean and standard deviation. This grid size was used to allow direct linking to gridded meteorological data and assess weather drivers of spatial and temporal variability across the Netherlands. We found that SD and HD differed significantly between years; for example, the estimated HD in 2018 shifted forward significantly (on average 24.4 days) compared with HD in the rest of study years because of the extreme summer drought. Our presentation will provide further evidence on how farmer’s decisions on agricultural practices is influenced by weather conditions and legislation.

This study shows that Sentinel-2 time series allow to extract crop-specific sowing and harvest dates for large areas; while here the focus was on a relatively small country with detailed knowledge on maize field locations, progress in Europe-wide crop type mapping and monitoring would allow to generate detailed crop calendars at continental-scale. Moreover, mapping the spatial and temporal variability of crop-specific phenometrics allows for a better understanding of climate effects on agriculture.

Results from the study “Impact of COVID-19 on Harvest of Row Crops” will be presented. Within this study, it was quantified when harvest of row crops occurred in 2020 in 19 European countries and whether there were significant delays which might be due to the COVID-19 pandemic, since the pandemic with its lockdowns disrupted supply chains and usual agricultural worker migration between countries.

For this purpose, nearly 200 000 samples of summer crops, winter cereals and winter rapeseed were analysed. They were distributed across the agricultural areas of nineteen different European countries. The selection was built out of approximately 900 000 pre-selected points, which were part of VISTA’s YPSILON® service chain for yield prediction. These points were classified from Sentinel-2 data and thinned out for computational reasons. Each selected field has a size of at least 1 ha to get a large enough pixel-sample to account for the spatial resolution of the Sentinel-1 scenes and their speckle effects, while at the same time including only the relevant crop type over the included area to avoid mixed signals.

The algorithm developed during this project derives harvest dates by the interpretation of the two basic SAR parameters backscatter intensity and interferometric coherence by using Sentinel-1 imagery. The interferometric coherence is a measure for the degree of correlation between two co-registered complex SAR images. Once harvest has occurred and bare soil is visible, the resulting more stable ground conditions (compared to the instable conditions when the ground is covered by vegetation) should increase the coherence significantly. Since it is possible that the coherence time series is biased by e.g. dynamically changing environmental conditions, the backscatter intensity signal is added as second indicator. A drop in the backscatter intensity signal can be due to a loss in multiple volume scattering and therefore can indicate a harvest event. By defining absolute and relative thresholds on the one hand and start dates for the observation period for each crop type on the other hand the algorithm determines harvest dates in a complex iterative way. Thus, backscatter and coherence are variables that can detect changes of the land surface and thus harvest events with an accuracy between one and five days.

For about 800 maize fields in Germany the derived harvest dates were compared against in situ data, and for about 1600 maize fields in the three countries (Germany, France and Czech Republic) the harvest was validated with harvest dates derived from optical satellite data (Sentinel-2). Due to this extensive analysis, the robustness of the algorithm was deemed satisfying.

To find out whether any countries or regions were affected by a delay of harvest, the cumulative harvest progress curves derived by the radar algorithm were compared with the one modelled by YPSILON® both at national and regional level. This extensive big data analysis framework YPSILON® is based on the advanced physically-based crop growth model PROMET, which takes both current weather conditions as well as the biomass growth as seen from Sentinel-2 into account, since the model assimilates LAI time series from optical imagery. Thus, the model predicts a harvest date that takes into account many measured variables and current information, such as satellite data (Sentinel-2), weather data (current and forecasts but also historical) and information about soil conditions and topography. Depending on the modelled phenological development of the crop and its maturity stage, an expected harvest date depending on crop maturity is estimated and can therefore be taken as the reference. The results from the YPSILON® service was therefore used to cross-check the expected harvest dates according to the crop conditions in 2020 with the ones derived from the radar harvest detection algorithm. Because YPSILON® describes the ‘perfect’ time of harvest based on the observed phenological stage of the crop, a later harvest date or longer lasting harvest period derived with the radar algorithm means a delay in harvesting.

Out of the sample of the nineteen European countries, especially Spain, Great Britain and Hungary stood out as having delayed harvests. A research on Covid-19 cases, lockdowns and national restrictions revealed that all three countries showed high incidence values and strong restrictions at different points in time, but borders in general remained open for workers. Hence, a clear case for Covid-19 impact on harvest cannot be made.

The results examined within this study are already available via three different online channels:
ESA’s RACE Platform (https://race.esa.int/?indicator=E10a10), the YPSILON® Portal (https://ypsilon.services/news/covid-harvest) and the Food Security TEP (https://foodsecurity-tep.net/app/#/explorer), because this is a fast and efficient way of distributing data to a wider public and independent of the persistent restrictions imposed due to COVID-19.

This study was funded by ESA under Contract No. 4000131553/20/I-DT.

Radar data have been used to retrieve and monitor the surface soil moisture (SM) changes in various conditions. However, the calibration of radar models whether empirically or physically-based, is still subject to large uncertainties especially at high-spatial resolution. To help calibrate radar-based retrieval approaches to supervising SM at high resolution, this paper presents an innovative synergistic method combining Sentinel-1 (S1) microwave and Landsat-7/8 (L7/8) thermal data. First, the S1 backscatter coefficient was normalized by its maximum and minimum values obtained during the 2015-2016 agriculture season. Second, the normalized S1 backscatter coefficient was calibrated from reference points provided by a thermal-derived SM proxy named soil evaporative efficiency (SEE, defined as the ratio of actual to potential soil evaporation). SEE was estimated as the radiometric soil temperature normalized by its minimum and maximum values reached in a water-saturated and dry soil, respectively. We estimated both soil temperature endmembers by using a soil energy balance model forced by available meteorological forcing. The proposed approach was evaluated against in situ SM measurements collected over three bare soil fields in a semi-arid region in Morocco and we compared it against a classical approach based on radar data only. The two polarizations VV (vertical transmit and receive) and VH (vertical transmit and horizontal receive) of the S1 data available over the area are tested to analyze the sensitivity of radar signal to SM at high incidence angles (39°-43°). We found that the VV polarization was better correlated to SM than the VH polarization with a determination coefficient of 0.47 and 0.28, respectively. By combining S1 (VV) and L7/8 data, we reduced the root mean square difference between satellite and in situ SM to 0.03 m3 m-3, which is far smaller than 0.16 m3 m-3 when using S1 (VV) only.

Abstract

Large scale land acquisitions (LSLAs), often referred as “land grabbing”, are highly dynamic and complex land use systems that are rapidly transforming ecosystems and societies in many low-income countries of the world, bringing on one hand sustainability challenges and, on the other hand, undermining the right of peoples to self-determination over natural resources. As such, monitoring of those large-scale agricultural expansions has appeared to be of paramount importance. In response to that need, the Land Matrix international initiative has emerged to promote the creation of an open access database on world land transactions. This open tool enables the collection and visualization of data on land deals based on publicly available sources (i.e. from governments, corporations, medias, citizen). However, because information on those acquisitions is often opaque and scarce, systems allowing near real-time LSLAs detection, characterization and monitoring are needed.
In this context, the increasing availability of free-of-cost global satellite data products has shown great potential for providing insights into land dynamics, particularly of large and remote areas. While LSLAs are not directly observable from remote sensing images (no one-to-one relation between land cover and functionality), they may be inferred from observable land cover and spatio-temporal characteristics at different scales, and structural elements in the landscape. At a pixel level, land use and land cover (LULC) changes are often detected using change detection algorithms applied on temporally-dense satellite image time series (SITS) of vegetation indices. So far, most of the LULC change studies have focused on forested land covers where significant deviations (anomalies) from the mean are relatively easy to detect. However, LULC changes, and in particular human-driven ones such as those induced by LSLAs, often imply a change in (seasonal) interannual patterns (not always with significant shifts from the mean), that are less well detected by change detection algorithms. Accurate and automatic detection of those type of changes would thus pave the way for the development of generic and unsupervised approaches to LULC change detection.
This study deals with the detection of agricultural LSLAs under different environmental conditions. Focus is given on Senegal for which we have ground-truth data. In addition, its strong north-south gradient of rainfall from dry to semi-humid climate, and relatively small sizes of its deals make Senegal an interesting and difficult study case study for the detection of LSLAs. The detection method proposed here is based on a two-step approach: 1- the detection of (if any) breakpoints in dense MODIS 2000-2020 Vegetation Index (NDVI) time series using the very fast BFAST monitor algorithm. Because BFAST monitor algorithm is subject to a high false positive rate, we implemented a second step to select the breakpoint most likely related to the desired land use change (biggest pattern change), 2- the selection, for each pixel, of the breakpoint associated to the biggest phenological change, based on a time series distance computed between the subsamples before and after each breakpoint.
Results consist of change-intensity maps, date-of-change maps, and a comparison of the change detection maps obtained using our method vs. using the biggest BFAST-magnitude change detected. Areas potentially related to agricultural LSLAs are identified and qualitatively/quantitatively characterized (e.g. year of change, spatial expansion) and evaluated against field data (when available) and high spatial resolution spatial imagery (Landsat/ Google Earth). The method was also tested over different more humid and forested specific areas found in the literature (e.g. Laos and Mozambique), where agricultural LSLAs have been reported and characterized. For these areas we produced maps of deforested areas, with associated date-of-change, that could be assessed qualitatively. The results indicate that our method has a high potential for detecting LSLAs even in humid regions, and thus for mapping the extent and dynamics of deforestation driven by different types of commodities. Future efforts will focus on a finer assessment of the driving factor of the detected LULC changes (e.g. fire, forest management, commodities cropping etc.) through the application of image analysis techniques (clustering and object-based image analysis).

The desert locust is a well-known plague affecting a large area from North Africa to India that compromise the food security and the economy of fragile countries. In general, the desert locust breeds extensively in semi-arid zones extending from West Africa through the Middle East to Southwest Asia, threatening livelihoods of the population in over 65 countries. Between 2019–2020, unprecedented locust breeding was observed in Eritrea, Somalia, and Yemen due to unusually heavy rainfall in the horn of Africa between October to mid-November 2019, more than 400% above average.
The main goal of this work, carried out in the framework of the European Space Agency (ESA) project «Earth Observation for Yemen» (EO4Yemen)), is to develop a geospatial risk outbreak model for the timely location of desert locust invasion and mapping desert locust risk zones in Eastern Africa region. An estimation of the impact, detection and measurement on the potential destruction of the locust and then its mapping to provide updated information on its damage status to the government/decision makers but also to farmers.
Here we present a case study where Earth Observation data and climate data have been processed and analyzed with Artificial Intelligence techniques with the twofold scope of assessing the impact of locust presence and to predict the locust swarm migration by locating the suitable breading areas. In this case study, the Earth Observation Change detection techniques have been applied to Sentinel 1 and Sentinel 2 for measuring the impact of locust presence in terms of crop and vegetation damages and observation of migration processes. Climate data have been analyzed in order to create a prediction climate model of locust habitat and related environmental conditions for locust development. Different climate indicators have been compared to locust presence occurrences with Machine Learning techniques to identify climate and temporal dependencies and correlations for locust risk prediction.
Earth Observation data for change detection makes use of both Sentinel-1 and Sentinel-2 data. NDVI from Sentinel-2 is used with clear sky conditions, while a machine learning method has been applied to compute vegetation index on Sentinel- 1 data and analyze in time the evolution of vegetation in case of cloud coverage. Change detection methods applied to the vegetation indexes to identify those areas affected by locust presence.
The locust prediction part models the correlation between locusts´ occurrences and climate data and indicators. Four climate indicators are used: temperature, precipitation, soil moisture and vegetation tenure. The model is trained over the 10 days before the event occurrence, and provides a warning for the length of the forecast data used to run the model . Four neural network approaches have been implemented: Fully Convolutional Neura Network (FCNN), Long short-term memory (LSTM), Convolutional LSTM (ConvLSTM), and Support Vector Machine (SVM). ConvLSTM provides best performances among the different experiments performed (accuracy: 0.96, macro average: 0.49). The main issue faced is the unbalanced database: while the FAO locust watch database provides locusts detections, a dedicated strategy has been implemented to obtain a similar amount of points with “no detection”.
Results of the application of the development methodologies over Yemen and East Africa will be presented.

Water Resources, especially in intensively cultivated areas, are under high pressure and reciprocally influence agriculture. Nitrogen and other fertilizer and nutrient excesses affect water quality into the deepest layers of groundwater. Withdrawals for agricultural irrigation in drought periods affect water quantity and thus water availability for agriculture as well as for all other uses. Agriculture’s impact on water resources cannot be described by one factor though but it results from the sum of all actions, developments and conditions in an area. Environmental and regulatory conditions build the static framework, plant and land surface development as well as the management practices and activities of all agricultural participants (i.e. farmers) and water management authorities shape the dynamic processes of an agricultural area.

More than any other German federal state, Bavaria is clearly dominated by agricultural landscape1). Agricultural vegetation with an area of 3.1 Mio. ha (1) counts for about 46% of the total land use (2). At the same time with an average field size of < 2 ha (~1.74 ha in 2014 (3)) and a total of nearly 2 Mio. single fields e.g. in 2019 (4) Bavarian agriculture, similar to many other regions in the world, is very small-structured and thus a challenging analysis and management puzzle with millions of tiny little intertwining pieces.

To support this challenge addressing among others the Sustainable Development Goals (SDGs) 6 and 12, as a crucial factor in terms of agricultural water usage and nutrient consumption the current crop biomass development can be used. In the research project VieWBay (Virtueller Wasserraum Bayern = Bavarian Virtual Water Space) (5) from ESA COPERNICUS Sentinel-2 data we generate area-wide high-resolution green leaf area [m²/m²] data sets on field-scale for all of Bavaria for the years 2017 to 2020/2021. This is realized with a specially developed automated processing chain that consists of satellite data pre-processing, point sampling based on annual federal field geometry and crop cultivation data (4), plant parameter retrieval by inversion of the radiative transfer model SLC (Soil-Leaf-Canopy) (6) and finally the creation of a consistent daily/10-daily time series of leaf area development for each of the ~2 Mio. fields per year.

A crop type specific inversion of the SLC model to derive plant parameters uses look-up-tables (LUTs) consisting of a huge amount of reflectance spectra that are compared and matched to measured satellite reflectance spectra. Doing this for all agricultural pixels in 10 m spatial resolution in Bavaria for several years is not feasible regarding computational power and duration, which are linked to energy consumption and according CO2 emissions. To reduce the computing time for the plant parameter retrieval in Bavaria different approaches were carried out.

The use of Artificial Neural Networks (ANN) using LUTs and corresponding plant parameters from the SLC model as training data promised a strong decrease of computing time and in most cases showed good accuracies in plant parameter retrieval. However, in several cases the retrieved parameters differed greatly from the results of the normal LUT inversion approach.

Furthermore, the use of pre-processed LUTs for varying sensors and acquisition geometry (e.g. observer zenith, sun angle) instead of on-the-fly-computed LUTs strongly reduced the computing time.

As another method, a KMEANS clustering approach of the LUT also showed good results in reducing the time needed to find the best match between the observed spectrum and the SLC modelled spectrum within the LUT by calculating RMSEs and finding the spectrum with the smallest error.

A decrease of the computing time is also possible by lowering the number of pixels which are used within the LUT-Inversion approach. Therefore, instead of a fully-spatial processing a random sample approach was developed and carried out. For this approach within every agricultural field in Bavaria three pixels were selected as representative sample for the whole field. Since the overall aim of the VieWBay project is not to derive in-field varieties of plant parameters but calculating the regional water balance, this sample approach is representative for larger regions and therefore sufficient.

The different approaches to reduce computing time will be shown in detail and the impact on reducing energy consumption and therefore CO2 emissions will be quantified.

As part of an emerging innovative tool for decision-makers and stakeholders, the resulting data sets on green leaf area development and distribution will be used: firstly for the calibration of the sophisticated physically-based crop growth model PROMET (7) which will be used for the simulation of the water balance and the water and substance flows in Bavaria; secondly the data sets can be used for a better estimation of agricultural processes like nitrogen and carbon allocation as well as water consumption through crop growth; and thirdly they build the basis for an approach of a model-based algorithm to detect irrigated fields throughout Bavaria – a basis information for the assessment of irrigation water consumption and requirement.

The work presented here is part of the research project VieWBay, funded by the Bavarian State Ministry for the Environment and Consumer Protection and supported by the Bavarian State Office for the Environment.

(1) https://www.bmel-statistik.de/fileadmin/daten/MBT-0101160-0000.xlsx
(2) https://www.agrarbericht-2020.bayern.de/landwirtschaft-laendliche-entwicklung/landnutzung.html
(3) https://www.lfl.bayern.de/mam/cms07/publikationen/daten/informationen/agrarsturkturentwicklung-bayern_lfl-information.pdf
(4) InVeKos, correspond. IACS = integrated administration and control system
(5) https://viewbay.geographie-muenchen.de
(6) VERHOEF, W. & BACH, H. 2007. Coupled soil–leaf-canopy and atmosphere radiative transfer modeling to simulate hyperspectral multi-angular surface reflectance and TOA radiance data. Remote Sensing of Environment, 109, 166-182.
(7) MAUSER, W. & BACH, H. 2009. PROMET - Large scale distributed hydrological modelling to study the impact of climate change on the water flows of mountain watersheds. Journal of Hydrology, 376, 362-377.

Covering large and diverse areas with an important agronomic variety, grasslands account for almost a quarter of the European Union surface. The amount of grassland management practices strongly influences the quality of their ecosystem services. Information about large scale management intensities is then a crucial factor for sustainable decision-making in landscape policy and planning.

The high resolutions of the European Copernicus Sentinel satellites offer new possibilities to monitor vegetation by providing synoptic and regular observations. Especially, the high temporal resolution opens the door to the development of new approaches for grassland mowing detection. In addition, the availability of synergistic and complementary optical and SAR may overcome recurrent cloud cover in optical data or statistical fluctuations in SAR data improving mowing detection.

Mowing events can be defined as an abrupt change in time series. Most of the works feature the use of vegetation indexes such as the NDVI which allows describing the vegetation’s phenology. Threshold approaches are mainly employed to detect changes in local trends of the NDVI time series [1,2,3]. While achieving accurate detection, such threshold approaches could not be operationally deployed due to several factors [4]. Unfortunately, one of the main limitations is the observed temporal resolution of optical images such as Sentinel-2, hampered by cloud cover. Yet, some intensively exploited grasslands are characterized by very short phenological cycles, with rapid regrowth following mowing. The time interval allowing the detection of the mowing can be relatively short, of the order of one week. As a result, mowing events occur very fast with respect to the temporal observed sampling period of the optical satellite measurements.

The temporal resolution being decisive, the use of SAR time series has been proposed in the literature. Thresholds on backscattering coefficient or interferometric coherence have for example been used. However, complex interpretations and inconsistencies in SAR responses to mowing events have been reported. Indeed, important impacts of exogenous factors such as soil moisture or acquisition geometry are associated to SAR images are limiting factors to their use. Satisfactory results have nevertheless been obtained, although mostly exploiting very high resolution SAR images or full polarization sensors, for which data are difficult to obtain.

Few works have proposed the use of reference data describing mowing events. Supervised detection techniques have subsequently been employed. Unfortunately, large scale reference data describing management practices over grasslands are not available. As a result these works are using only a few tens of parcels as a representation of grasslands. This reference data scarcity is also noticeable in the validation methodologies with small areas being assessed leading to the definition of area-dependent criteria.

A further limitation encountered in the literature lies in their analysis scale. National Land Parcel Identification Systems are commonly used to define object level approaches based on parcels boundaries. Nevertheless, grasslands have the specificity to be exploited at scales finer than the administrative declaration scale of the parcel, with intra-parcel rotational strategies employed by farmers. A grassland parcel thus may encompass different vegetation states (e.g., one part mowed, the other growing) within its administrative boundaries.

In this work, we propose to exploit the synergy of Sentinel-1 and Sentinel-2 image time series to detect mowing events at a super-pixel scale which defines homogeneously managed sub-parcel areas. The proposed methodology is based on three contributions: (i) a super-pixel strategy to segment the grasslands parcels; (ii) applying the Sentinels Regression for Vegetation Monitoring approach [5] to construct dense NDVI time series at the super-pixel level in order to deal with the lack of optical data; (iii) the study of different detection strategies to identify mowing over grasslands.

Considering the dense temporal sampling of the resulting NDVI time series, different change point detection strategies have been considered. The well-known BFAST method mainly proposed to exploit low or medium resolution optical Earth Observation time series strategies has been explored. The obtained results have been compared and evaluated with other methodologies from simple thresholding methods to segmentation strategies such as the bottom-up or PELT methods.

A large validation dataset is constructed containing over 2,000 abrupt changes over grasslands observed on two large scale areas covering more than 3,000 km².

The obtained results highlight the difficulty of detecting mowing over grasslands by further analyzing time series pattern. NDVI stability before mowing or the different magnitudes of mowing-related drops are investigated. Results have corroborated that simple strategies, combined with NDVI time series obtained by the SenRVM method lead to satisfactory results.

[1] S. Asam & al., 2015, “Estimation of grassland use intensities based on high spatial resolution LAI time series.” ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 285-291.
[2] S. Ester & al., 2018, “Combining satellite data and agricultural statistics to map grassland management intensity in Europe. “, Environ. Res. Lett. 13,074020, doi: 10.1088/1748-9326/aacc7a
[3] S. Reinermann & al., 2021, “ Detection of grassland mowing events with optical satellite time series data.”, 21th EGF Symposium: Sensing – new insights into grassland science and practice. Online. Volume 26.
[4] A. Garioud & al., 2019, "Challenges in Grassland Mowing Event Detection with Multimodal Sentinel Images," 2019 10th International Workshop on the Analysis of Multitemporal Remote Sensing Images (MultiTemp), pp. 1-4, doi: 10.1109/Multi-Temp.2019.8866914.
[5] A. Garioud et al., 2021, “Recurrent-based regression of sentinel time series forcontinuous vegetation monitoring.”, Remote Sens. Environ., 263, 112419. doi:1016/j.rse.2021.112419

Identification of crops over large areas is necessary for monitoring agricultural production, establishing food security policies or controlling compliance with the Common Agricultural Policy. Data acquired by Sentinel-1 and Sentinel-2 satellite constellations with a combination of high temporal and spatial resolutions has already demonstrated, separately or fused, its ability to classify crops. However, the usage of raw data at national or continental scale requires processing of huge amounts of data. This involves complex gap filling procedures in time and space (e.g. due to cloud cover) in order to get a clear description of crop development each season.

To overcome these deficiencies we performed crop classification by using High Resolution Vegetation Phenology and Productivity (HR-VPP) product and its phenology metrics. The HR-VPP is a new Copernicus 10-meter resolution product covering the period 2017-2020 and obtained based on the Plant Phenology Index (PPI) derived from Sentinel-2. It provides phenological information such as: start and end of vegetation season (day-of-year and PPI value), length of season, minimum and maximum of season (day-of-year and PPI value), seasonal amplitude, slope of the gathering up and senescent period, as well as seasonal and total productivity.

We evaluated HR-VPP for crop type classification over a selected test area in southern Poland covering one tile of Sentinel-2, including over 200 thousand parcels sown with 23 crop types. Reference data for training and validation of the algorithms were prepared from in situ data for the year 2020.

A number of different methodological approaches that have been explored will be presented. Firstly, we used several machine learning methods (such as Random Forest, Extreme Gradient Boosting, Support Vector Machine, KNN, Neural Network, Adaboost). Secondly, we used soil data (SoilGrids) to stratify the area, so that each area was classified separately. Third, we looked at the differences in classifying all crops at once as opposed to a hierarchical crop-by-crop approach. Finally, we considered approaches on how to derive classification results probabilistically.

This study revealed that the HR-VPP phenological metrics allow for regular (several times per season) and timely crop classifications. Our methodology features a great potential to be implemented not only at a country scale (Poland) but also across Europe.

Modelling crop yield based on remote sensing data holds great potential for a multitude of applications including precision farming, research, agro-insurances and policymaking. For many of these – especially for precision farming – data is needed on the field level. On a large spatial scale, such modelling approaches have been performed by mainly using data obtained from the Landsat satellites. Yet, with the high spatial resolution of the Sentinel-2 (S2) satellites, such studies become feasible for smaller fields and thus for more multi-faceted landscapes such as found in Switzerland, southern Germany, and other central European regions. However, this is also related to a set of new challenges such as how to treat field borders, in-field-heterogeneity and how to derive corresponding ground truth data such as yield data – notably on the field and sub-field levels.
In collaboration with a farm in Switzerland, georeferenced combine harvester yield data on field level was obtained. The data set spans five years from 2017 to 2021 with 77 fields in total. Most crops are cereals such as wheat and barley. The average field size of 9.6 ha is small compared to field sizes in the international context, but relatively large compared to the average field size of 1.6 ha found in Switzerland, making the data set eligible for training a yield prediction model on the pixel level.
In a first step, multiple spectral index (SI) time series were calculated from the multispectral S2 time series. The SIs include NDVI, NDRE, CLRededge, CLGreen and EVI2band. For smoothing of the SI time series, Fourier series (FS) and double logistic (DL) smoothing were tested, as experience with using these smoothing methods on S2 data is less compared to Landsat data. FS smoothing was found to better approximate the SI time series (overall R2=0.76, RMSE=0.65) than DL smoothing (overall R2=0.70, RMSE=0.73).
In a second step, the smoothed SI time series were used to extract features for machine learning (ML) models. These features were fed into a Random Forest regression ML model. Preliminary analyses using the NDVI time series on the year 2017 show a Random Forest model performance R2 of 0.46 with an RMSE of 0.72 t/ha (compared to an average winter wheat yield of 6.3 t/ha). Performance on the test set was slightly better with an R2 of 0.52 and an RMSE of 0.64 t/ha. The years 2018 to 2021, further crops and SIs are currently being analysed and will be presented at the conference.
It is planned to enhance this baseline model with meteorological data, use growing degree days as the time series for cross-year comparability and use a topsoil map of the region. These expansions of the baseline method are expected to increase the model performance. It is also planned to fit an artificial neural network (ANN) to the data set to compare modern ANN based modelling with ‘traditional’, easily interpretable, ML methods such as Random Forests.
With these analyses, we aim to show the applicability of the combination of readily available, state-of-the-art sensor data and methods to estimate and predict the crop yields in a small-scaled agricultural environment such as that of Switzerland. This will allow for a more site-specific management of the fields in a heterogeneous landscape consisting of different soil types and orography in close proximity.

Grassland covers about one quarter of the Earth's surface. In Austria, grassland even represents the most important land-use system with 1.34 million hectares. From an agricultural point of view, its main function is the supply of forage for livestock. The precise knowledge of cut dates is an important prerequisite for biomass estimation and essential to improve grassland management. Retrieving this information on a large scale using earth observation comes with advantages regarding timeliness, costs, and data consistency. Radar observations provide gap-free measurements due to their ability to penetrate clouds. Especially Sentinel-1 C-band SAR observations are of high value as they produce dense time series.

This study aims to assess the potential of Sentinel-1 backscatter time series to detect grassland cuts in Austria. In addition, the performance of sigma naught and terrain flattened gamma naught for grassland detection is compared.
In the course of this project, reference data for grassland fields over several seasons is collected via a mobile phone app provided to partner farmers. In addition to the precise cut dates, farmers also take pictures of the grasslands and provide grass height measurements and information on species distribution within the fields. Sentinel-1 sigma naught and terrain flattened gamma naught time series for VV, VH polarization, and their cross-ratio (CR) are derived for around 250 fields in total. Terrain-flattened gamma naught has shown advantages over sigma naught in mountainous areas, as it mitigates the impact of the terrain on the backscatter. This characteristic is of high importance in Austria with a significant number of steep alpine meadows. In the course of time series generation, extensive pre-processing is carried out to remove the undesired impact of radar shadows and mixed land cover, e.g., single trees or small buildings within field parcels.
Using either the sigma or gamma time series and reference cut dates, we train a Gated Recurrent Unit. The designed model predicts for every input time step the probability of being a cut date. To reduce the number of false-positive cut dates we apply knowledge-based rules to this probability time series, e.g., limiting its range to the cut and growing season or defining a certain number of days in between consecutive cuts. Within the accuracy assessment, the performance of the model based on sigma naught and terrain flattened gamma naught will be evaluated and further investigated. While the designed workflow is only applied on previous seasons in a first step, it will later be tested for near-real-time cut detection.

Our preliminary results on a small scale indicate good performances for fields with distinct patterns in the time series: grassland cuts lead to a quick increase in VH polarization and CR, followed by a gradual decrease. However, some fields miss these discriminating patterns. Using a comprehensive reference database will allow a deeper understanding of the backscatter signal before and after cut events and under which conditions a cut detection fails respectively a false positive cut is detected. At a later stage, we aim to improve the result by adding meteorological data (growing degree days, rainfall, soil moisture) to the model.

In summary, this study provides new findings on the potentials and limitations of SAR based cut detection and a comparison of the performance of sigma naught and terrain flattened gamma for grassland applications in mountainous areas.

The increasing demand for food due to the rising population and the simultaneous shortage of agricultural land challenges agriculture in particular in the light of ongoing global change, such as climate change, threats to water quantity and quality, soil degradation, environmental pollution, destruction of ecosystems, and biodiversity loss. Efficient and sustainable solutions for adapting to the effects of global change are therefore of central importance. The digitization of agriculture offers the opportunity to optimize and automate processes, but also poses challenges for farmers.
International activities such as GEOGLAM defined essential agricultural variables and core information products from remote sensing. However, major shortcomings are the lack of workflows for bringing scientifically based knowledge and methods into practice as well as insufficiently specified interfaces between data sources, farmer and machine. Methodological challenges are, above all, the robustness of the procedures, the handling of multi-sensor data, standards regarding data quality and the trustworthiness of scientific models as well as their temporal and spatial transferability. Open and solution-oriented communication with farmers regarding potentials, accuracies and limitations of remote sensing products is also strongly deficient. Technical questions about data management, user-friendliness, data integrity and data protection as well as high investment costs are further barriers. Farmers often lack awareness of the added value of the data, while scientists often fail to provide easily understandable data interpretations. However, it is also evident that practicing farmers are open to new technical solutions.
The project "AgriSens - DEMMIN 4.0" (project duration 02/2020-01/2023), funded by the German Ministry for Nutrition and Agriculture (BMEL), has the goal to identify practical applications based on remote sensing data originated from satellites, aircrafts, and UAV-supported systems, for crop production. It further aims at developing new methods and making this knowledge easily available to the farmer and the public. Therefore, four use casas are implemented targeting at crop growth monitoring and yield prediction, sustainable use of low yield zones, irrigation monitoring, and detection of glacial stones at fields. The presentation will show and explain the first results of these use cases:
The first use case focuses on a more resource-efficient management in winter wheat by the integration of re-mote sensing-based information on current crop status, crop development and potential crop yield. For this purpose, key variables such as above-ground biomass and leaf area index are derived from Copernicus Sentinel images. They are coupled with a crop growth model to provide spatial explicit, daily information on the status of crops as new base layer for the application of plant protection products, fertilizers or growth regulators. On-farm experiments planned for 2022 shall provide insights on the potential economic and ecological benefits of this new source of information. Additionally, the potential of airborne images is evaluated
The second use case is dedicated to sustainable management of agricultural land. Considering intra-field heterogeneity (e.g., through precision farming) can increase or stabilise yields while reducing the use of operating resources. Furthermore, this can contribute to the reduction of ammonia and nitrogen oxide concentrations in the soil and thus to the improvement of water quality. The aim of this use case is to provide an information layer that allows the identification, location, and typification of low-yield areas to support their optimised management. For this purpose, local knowledge about these areas is repeatedly digitally captured by farmers during field work using the "FieldMApp" application on mobile devices. The captured FieldMApp data are fused and afterwards blended with satellite data. The functionality and design of the "FieldMApp" are defined in a cooperative collaboration between farmers and scientists (citizen science approach) in order to create a solution that meets the requirements of both.
The third use case deals with the detection of stones on agricultural land, which can cause major damage to agricultural machinery and have so far been removed manually by driving off the entire field. The aim is to develop a marketable workflow for the drone-based detection of stones in order to reduce personnel and machine costs. In this way, only selected areas on the fields need to be targeted and operating resources are saved and area compaction is reduced. Stones are detected at different field trials in 2021 using different camera techniques, e.g. optical, thermal, Lidar, in combination with object-based image analysis algorithms.
The topic of the fourth use case is irrigation technology. In times of increasing extreme weather events, which became particularly visible in 2018 and 2019 due to a pronounced drought throughout Germany, many farmers are focusing on the expansion or re-installation of irrigation infrastructures. In order to apply the resource water in a demand-oriented and cost-conscious manner. The use case presents field trials in 2021 with different irrigation strategies and the corresponding detection of growth and quality parameters for potato. This information is combined with soil water modelling and remote sensing data analysis for supporting future site-specific irrigation strategy. The use cases are supporting with its results for greater acceptance and wider use of these valuable data sources for operational processes in crop production. This presentation spotlights on the experimental field DEMMIN, the project AgriSens - DEMMIN 4.0 and first results from it.
AgriSens DEMMIN 4.0 located at the only German test site in the Joint Experiment of Crop Assessment and Monitoring (JECAM) an initiative for product development and validation in GEOGLAM. Thanks to its many years of research activities and its national and international networking, DEMMIN is ideally suited as a test site for the AgriSens - DEMMIN 4.0 project. DEMMIN (Durable Environmental Multidisciplinary Monitoring Information Network) experimental field is located about 180 km north of Berlin in the federal state of Mecklenburg-Western Pomerania in the north-eastern German lowlands. The young moraine landscape with its numerous lakes and bogs is characterised by typical periglacial landscape elements such as extensive, flat sandy areas, hills and depressions. DEMMIN has been operated as large facility for calibration and validation of remote sensing data by the German Aerospace Centre for since 2000. Among others, it is equipped with 43 environmental measuring stations, 63 soil moisture stations, a lysimeter hexagon, an eddy covariance measuring station and a research crane. Since 2011, it has also been part of the TERENO Observatory Northeast of the German Research Centre for Geosciences Potsdam.
The combination of the infrastructure facility and close cooperation between farmers and researchers at the site DEMMIN enables a large potential to combine high quality method development with quality assessment based on in-situ data and farmers information. Furthermore, the research can consider directly the needs of farmers related to remote sensing based information products. This exchange is a further key element of the AgriSens – DEMMIN 4.0. Regular local workshops, like AgriSens DEMMIN field day 2021 with 125 participants are supporting this exchange, but also includes now aspects like the knowledge and role of agricultural advice services.
The oral talk of AgriSens DEMMIN 4.0 will present the described project and especially focus on the first results of the use cases and field trials achieved in 2021. Furthermore, other aspects, like the information of the status of use of remote sensing-based products and data handling options within the project, including access to farmers for the developed services will presented.

The Mekong River Delta, popularly known as the "Rice Bowl" of Vietnam, has a key role in determining rice prices not only in South-east Asia but also in the world market. However, the rice production in the region has already been impacted by climate change, and projections for the next few decades indicate that the climate warming trends and anthropogenic pressures are likely to be accelerated, which has consequences for local and global food security.

The objective of this study is to develop projections on rice production in the Vietnam Mekong Delta (VMD), under the effects of climate change. In this research, we assess the possibility to use remote sensing for regional upscaling of the models. Specifically, we evaluate the use of Sentinel-1 data to provide spatial and temporal information related to agro-practices (e.g., crop establishment dates) and seasonal crop development (i.e., phenology) and vegetation status (e.g., height, biomass, etc.), in order to apply the models on a pixel basis for regional application. The research combines modelling, Earth Observation, in situ data, and climate data to assess and to simulate the effects of climate change on the two components of the rice production, rice yield and rice grown area.
For rice yield, the process-based model ORYZA2000 developed for rice has been calibrated in the VMD for the period of 2015-2020, before its use for future projections for 2030 and 2050 rice productivity taking into account the evolution in crop calendar and cultural practices which is currently observed in VMD. The remote sensing data (Sentinel-1) acquired since 2015 have a high temporal frequency (every 6 days) and high spatial resolution (10 m), allowing to access to the cultural practices and rice phenology at field level. The model parameters has been refined by a comparison of model outputs with yield data from National Statistics Office. Rice grown area and rice cropping density (the number of rice crop seasons per year, typically from 1 to 3 across VMD), has been updated by using remote sensing data. For future projections of rice grown area, a first approximation consists in estimating the reduction of rice grown area caused by salinity intrusion. Finally, the projections in rice grown area and in rice yield has been combined to simulate the future rice yield of each province in the Mekong Delta, under different climate scenarios for 2030 and 2050.

The impacts of climate change are quantified by a comparison between simulated annual rice yields of Mekong Delta provinces for Winter-Spring, Summer-Autumn and Autumn-Winter crops for the period from 2020 to 2050, and the average yield of the last five years. The yield reduction ratios obtained have been used to discuss about rice production in 2030 and 2050 for different scenarios of climate and its adaptation strategy.

The mapping of field sizes is advancing rapidly. Current approaches based on high-resolution satellite imagery and machine learning algorithms provide quasi-operational workflows in consolidated agricultural systems across the globe. Contrastingly, field size mapping in smallholder agricultural systems in Sub-Saharan Africa is far from operational. With field sizes commonly not exceeding half a hectare, the spatial resolution of globally available optical satellite imagery such as Sentinel-2 and Landsat can be insufficient to delineate individual fields. Furthermore, the lack of distinct spatial boundaries between fields creates additional challenges for approaches relying on the identification of field edges. Therefore, assessing field size in smallholder landscapes requires alternative approaches which go beyond the delineation of individual fields and instead focus on distribution parameters of field sizes, such as average or maximum field size. Such approaches can allow for assessing field sizes in regions or for periods where the mismatch between spatial resolution and field size would not allow for the identification of individual fields.
This work explores multiple pathways for assessing local-level field size distribution combining field-based reference data, a variety of satellite image archives with varying spatial resolution, and machine learning regression models. Our reference dataset consists of 23 UAV-based images distributed across three provinces in Northern Mozambique (Niassa, Zambézia, and Nampula). The reference images cover between 6 and 10 hectares and contain between a few and up to dozens of fields. We demarcated individual fields from these images and derived local field size distribution parameters. We then generated a set of satellite image features, such as spectral-temporal and textural metrics, from intra-annual PlanetScope, Sentinel-2, and Landsat time series for the field sites. Finally, we trained multiple machine learning regression models to predict local field size distribution parameters and assessed the goodness of fit of the different models. While uncertainties remain high across all tested approaches, we discuss how particular combinations of sensors, image features, and distribution parameters could potentially be used to approximate field size distributions in complex smallholder systems across large areas.

Climate change affects food security, as changes in temperature and precipitation patterns affect vegetation, specifically crops. Copernicus Sentinel-2 mission offers new opportunities for crop management using remote sensing data. It provides information at a spatial and temporal resolution of 10 m and 5 days, respectively, due to the existence of two polar-orbiting satellites placed in the same sun-synchronous orbit, phased at 180° to each other: Sentinel-2A, launched in 2015, and Sentinel-2B, launched two years later. However, this unprecedented time series of high resolution satellite imagery, requires from approaches to extract meaningful agronomical information and reduce spectral and temporal dimensionality. It can be solved using land surface phenology (LSP), the study of the cyclical and seasonal phenomena of vegetation, as well as its relationship with biotic and abiotic environmental. Regarding crops, it consists in estimating key phenometrics related to agronomical events from time series of vegetation indices (VIs). Specifically, knowing the dynamics of crop phenology is essential for the correct monitoring of the European Common Agrarian Policy (CAP). The new CAP approach aims to monitor 100% of croplands using Sentinel-2 data, developing methodologies for each agroclimatic region in Europe. In this study we used EVI2 (Enhanced Vegetation Index 2) time series of Sentinel-2 data for the period 2018-2020. NDVI (Normalized Difference Vegetation Index), the most used VI, tends to saturate in areas with high biomass density, such as a crop during its growing season, but EVI2 has a better response in this situations. Moreover, EVI2 is more sensitive to variations in some plant species. The lack of information caused by cloud presence or sensor failures in selected parts or the whole image was addressed using the Double Logistic smoothing method. A threshold of 0.15 was implemented to identify the beginning and the end of the growing season. We selected parcels of some winter cereals; common wheat, durum wheat, sorghum, barley and triticale according to the Geographical Information System for the CAP (GISCAP-CAP) declarations in Andalusia, Spain. This is an official database where farmers state the main crop type that they have in their lands, if there is a secondary crop, or the presence or absence of irrigation systems, including the boundaries of the cropland too. The phenometrics extracted were start of the season (SOS), middle of the season (MOS), end of the season (EOS), their respective values of EVI2 for each moment, and length of the season (LOS). The aim of this study is to characterise the LSP of different winter cereals, through the extraction of meaningful phenometrics, and to evaluate whether these latter measures can serve to distinguish them. Results show that the response is quite similar between them, with subtle differences. They had their SOS in the first half of March, reached their MOS by the end of March and beginning of April and had their EOS in the second half of April. The exception was sorghum whose phenometrics occurred around 20 days later than the average of the rest of crops. Common wheat shows the earliest SOS and EOS, followed by barley, durum wheat, triticale and sorghum for SOS and by durum wheat, barley, triticale and sorghum for EOS. Overall values for each phenometric are very similar, although a combination of all of them gives a specific phenometrics signature that may help to distinguish them and they can be used as input for classification models. Moreover, this methodology can be used to elaborate a crop calendar in Andalusia that can be useful for farmers, CAP management, researchers and general public.

Phenology mapping allows monitoring vegetation dynamics that could significantly assist agricultural management practices and serves as an indicator to environmental changes. Particularly for cropland monitoring, adequate spatial and temporal resolutions are required to be able to characterize their different phenological stages .

The Northern Nile Delta represents about half of all agricultural lands of Egypt. In this region, intensifying farming systems are predominant and multi-crop rotations schemes are increasing, requiring a high temporal and spatial resolution monitoring for capturing successive cropland growth cycles. This study presents a workflow for cropland phenology characterization based on time series of green Leaf Area Index (LAI) generated from NASA's Harmonized Landsat 8 (L8) and Sentinel-2 (S2) surface reflectance dataset from 2016 to 2020.

Two different LAI time series were processed for each satellite dataset, which were used separately and fused to identify seasonal dynamics for a selection of crop types (wheat, clover, maize and rice). For fusing L8 and S2 LAI products, we proposed a time series smoothing and a fitting method: (1) the Savitzky-Golay filter, and (2) the Gaussian Processes Regression (GPR) technique. Single-sensor and L8-S2 fused LAI time series were used for the calculation of key phenological metrics (start of season, end of season, length of season) applying two different threshold based approaches to detect cropland growing seasons defining a seasonal or a relative amplitude value.

Overall, the developed phenology extraction scheme enabled identifying two successive cropland cycles within a year over four consecutive years, with a superior performance observed for the seasonal than for the relative threshold method, in terms of consistency and cropland season detection capability, revealing that setting a fixed value as a fraction of the amplitude tends to fail detecting successive growing seasons, as implies a fixed start/end LAI value throughout for the whole studied time window. Differences between the four LAI time series generated were analyzed by comparing the derived phenological metrics per crop type and year. Results suggest that both SG and GPR similarly captured seasonal patterns in LAI corresponding to phenological events. Compared to the single-sensor LAI time series, L8-S2 fused LAI data streams lead to a more precise detection of the start and end of the within-year multiple growing seasons for most crop types, reaching a maximum detection of 75% over the total planted crops. Finally, phenology mapping allowed us to evaluate the spatial distribution of single and multi-cropping parcels and the temporal evolution of the croplands over the study site in the Nile Delta from 2016 to 2019 differentiating crop phenology seasonality for winter and summer crops.

Although only tested at a local scale, this study demonstrates the application of high spatial resolution satellite-derived LAI time series for accurately estimating cropland phenology over within-year multi-cropping systems. Our findings also show the great potential for fusing optical satellite-based estimations to improve cropland phenology mapping.

Meadows belong to the most diverse ecosystems, both in terms of plants and animals. They are among the agricultural areas with the most extraordinary biodiversity. However, their diversity is declining due to various influences, such as abandonment of agriculture, over-fertilization and frequent mowing. Extensive meadows, in particular, are among the most endangered biomes, as they are heavily influenced by various factors, including more intensive use. With the demand for more yield, meadows are fertilized, mowed early, or fall into disrepair because management is no longer practical or possible. In all of these, diversity decreases.
These events or their effects can be tracked from space. With the availability of Sentinel-2 imagery, we can create dense time series that can be used to distinguish between different types of meadows and activities. The problem lies in the availability of reference data, which is often outdated or not available. This is especially true for historical data. In our study, we used data collected by the Institute of the Republic of Slovenia for Nature Conservation. The data were repurposed for classification and might include inconsistencies. They were collected sparsely and concern mainly extensive meadows that are of interest for protection. We have eleven regions which contain 10061 extensive and 1851 intensive meadows. This data sample does not reflect the actual distribution as extensive meadows are scarcer than intensive meadows. The data were collected over many years using different methodologies, which has resulted in an inconsistent reference. The reference was first matched with aerial imagery and Google satellite imagery to remove areas overlapping with other land use or cover types such as forest, urban or agricultural. In this way, we obtained 1133 intensive and 711 extensive meadows. Based on this reference, we acquired Sentinel-2 imagery. The time window considered was March to September, when most changes occur in the field. The data were atmospherically corrected using the Sen2Cor pipeline to generate a cloud-free time series for each meadow. We created an NDVI time series by interpolating the available observations to a common interval for the 1844 polygons. With each polygon corresponding to a meadow. For the unlabelled data we used meadows from a public dataset which also includes different other land use types. With it we created over 500 000 NDVI time series of meadows with unknown labels.
The constructed time series and the reference were used to prepare the dataset for machine learning. The task at hand was that of classifying meadows as extensive or intensive. We usually use supervised learning methods for such tasks, but they work best with large and reliable sets of reference data. For smaller set of reference samples we can use semi-supervised learning, which allows us to use both labelled and unlabelled data, because the latter is often readily available. To this end, we used the CLUS framework, which supports semi-supervised learning and can use the distribution of unlabelled (time series) to provide additional information. The framework offers different machine learning methods. The results presented here are produced by using Random Forests, as they achieved the best results for the given task.
We learned predictive models by both supervised and semi-supervised machine learning using the manually reviewed data as the training set. The results are shown in the left half of Table 1, where the manually reviewed data were used for training, and the Institute data were used for testing. Comparing the results of supervised and semi-supervised learning, we notice that both achieve similar results on the training data but cannot replicate the success on the testing data. The semi-supervised model performs better than the supervised one on the test data, by 2 percentage points.

Better results can be achieved by training on the data provided by the Institute and evaluating the model using the manually reviewed data. In both cases we used the same unlabelled data. As shown in the right half of Table 1, we achieve lower accuracy in training, and the semi-supervised model performs better on both data sets. With an accuracy of 79% on the test set, it outperforms the supervised model by 10%. Thus, we can conclude that the unlabelled data improves the accuracy of the model as it provides more context about the distribution of the data points (time series). The difference in performance between the two datasets could be due to the limited diversity in the manually reviewed dataset, as we have fewer regions and fever samples. This could affect the ability of the model to generalize across different regions of the country.
When only a few reference examples are available, we can improve performance by using semi-supervised learning, which can exploit unlabelled data. We have achieved 69% accuracy with supervised learning and semi-supervised learning with unlabelled data improved it to 79%. The 94% accuracy on training data drops by 15% on test data, which is due to over-fitting the training data.

Cloud cover is a common issue with optical satellite images, severely affecting the quality and the spatiotemporal availability of surface observations. Clouds can critically impact the utility of satellite images by completely covering the ground below, or distorting the measurements collected. While a generally known practice is cloud identification and removal of affected areas, there is also a growing interest in filling the gaps created with the use of powerful data-driven deep learning methods. Images produced by the Sentinel-2 mission come with such cloud occlusions and a common alternative is the use of Synthetic Aperture Radar (SAR) data, as they are nearly independent of the atmospheric conditions and solar illumination. However, SAR data share entirely different characteristics compared to optical data. Even though Sentinel-1 has the ability to provide continuous, day-and-night observations and to overcome various kinds of bad weather conditions or poor air quality (e.g., clouds, rain, fog and smoke), the information captured by this mission is less descriptive and more complex to interpret than that of optical images. In that case, Generative Adversarial Networks (GANs) are used to translate SAR to optical imagery, while many researchers have recently proposed the information fusion of SAR (Sentinel-1) and optical (Sentinel-2) images with different motives.
In this work, we investigate the possibility of optical data simulation for cloud gap filling, by fusing both SAR and optical data. In this context, a hybrid deep learning model is developed, targeting applications of cropland monitoring, where reliable information on crop development is essential to support the transition towards precision agriculture and evidence-based verification of compliance to agricultural policies. Optical imagery can provide invaluable insights in crop growth and development, but is acutely hampered by cloud cover. The proposed architecture considers a combination of a state-of-the-art GAN and a deep residual neural network. In contrast to previous approaches, GANs in this work are used to translate only the image's pixels/patches that contain clouds, instead of the whole image, avoiding in this way the deformation of already clear parts. Then, the residual neural network is trained to successfully fuse the original Sentinel-2 image with the Sentinel-2 patches that are generated by the GAN, towards the reconstruction of a cloud-free image.
Acknowledgements: This work has been supported by the EU's Horizon 2020 research and innovation programme under grant agreement H2020-101004152 CALLISTO.

Spatially explicit information about agricultural fields (e.g., field location and size, crop type) is a prerequisite for the effective monitoring of agricultural lands and provides the basis for an area-wide evaluation of the implementation of the Common Agricultural Policy (CAP). Through remote sensing, such spatial information can be obtained over wide geographical areas at a high temporal resolution. Whereas many recent studies have demonstrated the progress in mapping different crop types over large regions and countries, the robust delineation of agricultural field boundaries at a large scale remains challenging. Therefore, this research evaluates different methods for accurately segmenting agricultural fields from satellite images specifically Sentinel-2 (S2). To segment agricultural fields from satellite images, the multiresolution segmentation (MRS) algorithm (Baatz and Schäpe, 2000) has widely been used in the literature. In recent times, the use of deep neural networks (DNNs), particularly convolutional neural networks (CNNs), for land-cover and land-use mapping has been gaining traction in the RS world. Therefore, this research compares the MRS algorithm with three state-of-the-art CNNs namely U-Net (Ronneberger et al., 2015), ResUNet-A (Diakogiannis et al., 2020), and Mask R-CNN (He et al., 2017) for segmenting agricultural fields from S2 images acquired in May of 2018 over Lower Saxony, Germany. The geographical area of Lower Saxony was divided into 8417 tiles with each tile being 2.56 km by 2.56 km. The red, green, blue, and near-infrared bands of S2 were extracted and subsequently clipped to the tiles. To apply the three CNNs, 70% of the tiled images were used for training, 15% for validation, and 15% for testing. For the MRS algorithm, the Bayesian optimization approach of Tetteh et al. (2020) was applied to the test tiles. The segmentation results obtained at the test tiles by the CNNs and the optimized MRS approach were validated against the reference parcels contained in the Geospatial Aid Application (GSAA) dataset of Lower Saxony, Germany. The GSAA dataset contains agricultural parcels declared by farmers within the European Union (EU) intending to access the subsidies provided within the CAP framework. Based on a set of derived performance measures, the pros and cons of each algorithm are highlighted and the remaining challenges regarding the segmentation of agricultural fields from S2 images are discussed.

 
References
Baatz, M., Schäpe, A., 2000. Multiresolution Segmentation: an optimization approach for high quality multi-scale image segmentation, in: Strobl, J., Blaschke, T., Griesebner, G. (Eds.), Angewandte Geographische Informations-Verarbeitung XII. Wichmann Verlag, Karlsruhe, pp. 12–23.
Diakogiannis, F.I., Waldner, F., Caccetta, P., Wu, C., 2020. ResUNet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing 162, 94–114. https://doi.org/10.1016/j.isprsjprs.2020.01.013
He, K., Gkioxari, G., Dollár, P., Girshick, R., 2017. Mask R-CNN, in: 2017 IEEE International Conference on Computer Vision (ICCV). Presented at the 2017 IEEE International Conference on Computer Vision (ICCV), pp. 2980–2988. https://doi.org/10.1109/ICCV.2017.322
Ronneberger, O., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation, in: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science. Springer International Publishing, Cham, pp. 234–241. https://doi.org/10.1007/978-3-319-24574-4_28
Tetteh, G.O., Gocht, A., Conrad, C., 2020. Optimal parameters for delineating agricultural parcels from satellite images based on supervised Bayesian optimization. Computers and Electronics in Agriculture 178, 105696. https://doi.org/10.1016/j.compag.2020.105696

Evidence suggests, cropland abandonment became a global problem, with multiple implications to the environment, food security and societal well-being. Satellite remote sensing serves as a valuable approach to assess land-cover transitions, including cropland abandonment, due to ability to have repetitive observations, cover large areas, apply mathematical approaches to perform supervised land-cover change classification. Mountainous regions, for instance in China, can be prone to cropland abandonment due to unfavorable locational characteristics compared to plains. At the same time, mountainous areas can be difficult to map with optical satellite data, due to cloud presence and fragmented fields. Using multi-source fusion optical satellite data (HLS data, Landsat and Sentinel-2 images) as well as support vector machines classifier, we developed a new framework of mapping abandoned cropland in subtropical hilly and mountainous areas of China by integrating methods of the penalized cubic smoothing splines and feature optimization. We test such approach in Nanjing county in Fujian province of China as a case study. Further, we assessed the spatial determinants of cropland abandonment at different spatial scales with kernel density and redundancy analysis. Analysis showed fused Landsat and Sentinel-2 data resulted in accurate mapping of cropland abandonment. However, classification accuracies varied depending on image feature combinations with and without ancillary data (e.g., terrain). Cropland abandonment rate was as high as 32% from 2000 to 2018 (1.7%/year). Statistics based on the raster scale show that differ from the decreasing trend in areas of cropland abandonment along with the geographical condition and location (except for terrain) goes bad, along with each of the geographical condition and location goes from good to bad, cropland abandonment rates show the trend from low to high. Among the potential factors, the slope has the greatest effect on the change of cropland abandonment and can explain 53.3% of the variation of cropland abandonment at the township scale, followed closely by the distance to the rural resident area (18.2%). Research results can provide technical reference for mapping abandoned farmland and quantitatively measuring the chief influencing factors in subtropical hilly and mountain areas, which is relevant to better understanding the cropland abandonment process and its socio-economic and ecological landscape effects and assist the formulation of land-use policies.

Maize is one of the most prominent crops for human consumption in the world and a staple source of food supply that supports over 200 million people in developing countries. A third of the total areas under cereal cultivation in Africa is planted with maize. South Africa is the leading maize producer in the continent, and ranks one of the top ten in the world with a steady increase of production over the last century from a total of 328 000 tons in 1904 to a staggering record of 14.92 million tons in 2014 from 3 million ha of largely rain-fed farms. Irrigation only accounts for less than 1% of the total maize grown areas. The country is a net exporter of maize with about 25% of its production sold to countries in southern Africa. In Poland, spring cereals and maize accounts for 36% of the total cultivated land of which 31% is occupied by maize. Its production increased from 69 000 ha in 1996 to 1 200 000 ha now grown for grain and silage use. This study investigated crop growth monitoring and yield forecast in at the Agricultural Research Council-Natural Resources and Engineering (ARC-NRE) in South Africa and the Institute of Geodesy and Cartography (IGIK) in Poland. The unabated global warming is increasingly becoming worrisome for agriculture with substantial crop yield reduction already been reported globally. For instance, for every 1ᴼC increase in global mean temperature, maize production decreases by 7.4%. Recurring drought incidences and climatic variabilities are often the main factors for crop yield reduction. Poland is the second biggest producer of cereals in Europe. Nevertheless, the average yield remains lower than most countries in western Europe. This, however, is compensated by the extensive agricultural lands acquired for cultivation, making the country more competitive in the market, with 21.8 % of its cereal production exported in 2015.

The joint project between the two countries, investigated crop growth conditions and yield forecast using the Terra MODIS, Sentinel 2 and meteorological data in both countries. The MODIS database consisted of nine years of observation (between 2003 and 2019) and covered about 100 crop maize fields in Poland. The same number of maize fields with two white and yellow maize crops cultivated in rain-fed and irrigated practices were used for South Africa, spanning three provinces (Free State, North West and Mpumalanga provinces) for the period of five years from 2015 – 2020. Plant parameters such as crop phenology and crop yield data were recorded from field works, while the air temperature and rainfall data were acquired from the Agricultural Research Council weather stations and incorporated into the crop yield model. The data were analyzed using the accumulated eight days of NDVI (MOD09Q1) and accumulated 8 days’ differences between LST (MOD11A2) and air temperature (TA) from meteorological data. The preliminary results showed that the rapid increase in accumulated NDVI curve occurs at lower accumulated difference between LST and TA (∑LST-TA) and this resulted in high value of yield at the end of the season. During the dry season, however, the accumulated difference between LST and TA increased enormously resulting in lower rate of accumulated NDVI, thus this led to crop phenology occurring at different times. At good crop growing season, crop heading occurred earlier at lower accumulated difference in temperature (∑LST-TA) than in dry season and this has a direct response to crop yield. The FPAR was used to determine the different crop phenologies.

Keywords: FPAR; Yield Forecast; NDVI; Land Surface Temperature; Climatic Variability; Phenology

Introduction

Earth Observation (EO) techniques represent the most significant methodology to map the land cover owing to their cost-effectiveness and temporal repeatability of observations. The public access to datasets such as the Sentinel imagery, with their medium - high resolution, has represented a new paradigm to globally monitor our planet and, in particular, for agricultural purposes. In this sense, the joint use of optical (Sentinel-2, S-2) and Synthetic Aperture Radar (SAR; Sentinel-1, S-1) data for vegetation studies, such as crops classification or above-ground biomass estimation, is currently widely accepted on the scientific literature as they offer complementary information. This use has been proved to be an improvement for the crops classification as it implies the use of both active and passive sensors, enhancing the product. In addition, the near independence of SAR to weather conditions, being able to acquire images even in cloud-cover conditions, completes optical information of the system, representing an important improvement for regions that remains partially cover by clouds the great part of the year.

However, success is closely linked to the limitations of sensors, especially the spatial resolution, which is 10 m in the case of Sentinel-1 and 2. To avoid these weaknesses, new tools have been developed to improve the resolution of the Sentinel-2 imagery up to x4, achievement a final resolution of 2.5 m while preserving the temporal frequency. Crop phenology, monitored by both optical and SAR indexes and bands, have been largely used for crop classification (Phan et al., 2020; Maskell et al., 2021 and references therein) since more information is key to differentiate between types. Super-resolved Sentinel 2 imagery generated every few days allows more accurate crop classifications to be generated, and could facilitate the detection of some types of crops which are problematic with the native spatial resolution of sentinel 2, including challenging herbaceous and woody crops. Furthermore, this methodology improves significantly the parcel limit identification.
This study comprises two different regions of Spain (ROI-1 and ROI-2), corresponding to different climatic areas with own crops and diverse landscapes, for the period 2019-2021. ROI-1 includes an area of 850 km2 in Lugo, at the NW of Spain. This area is characterised by being wet and showing low temperature oscillation throughout the year (Atlantic Climatic) and having herbaceous crops. ROI-2 comprises an area of 700 km2 in Cuenca, corresponding to a typical Mediterranean climate (dry and hot summers and cold winters) and a combination of arable lands and permanent crops, including vineyards and olive groves.

Data and methods

Our studies are based on the development of novel algorithms and tools to derive crop types in wide Spanish areas, capturing different landscapes. Sentinel-1 and Sentinel-2 imagery for each ROI has been pre-processed and corrected (geometric, atmospheric and topographic) for each measurement that has been taken by the sensor from 2019 to 2021. After this initial stage, S-2 imagery has been processed to obtain super-resolved information. This has been achieved by the use of Artificial Intelligence based on antagonistic generative neural networks, applied to S-2A/2B images.

The algorithm IRIX8, developed by COTESA, uses SRGAN (Super Resolution Generative Adversarial Network) for the Super Resolution implementation. This framework is capable of inferring photo-realistic natural images for 4x upscaling factors (Ledig et al., 2017). A model has been trained and fed with the large VHR COTESA’s dataset. This includes multi-temporal VHR satellite imagery (WorldView 2 imagery with a resolution up to 0.5 m) from different European regions and ortho-photos (from Spanish National Geographic Institute). The main principle, within the model, is to progressively train the generator and discriminator. Our algorithm has been designed to search for balance in layers between performance and computational cost. In this sense, the deep residual network is able to recover realistic textures from S2 up sampled images.

The crops classification methodology is based on the well-established Random Forest Classifier. Our innovation lies on the joint use of radar (S1) and the VHR-S2 products obtained from the SR workflow. VHR indexes (i.e. NDVI) were calculated from the super-resolved S2 imagery and combined with the SAR backscattering and the interferometric coherence. This methodology takes advantage of all the potential of these products, gathering all these data on an innovative data cube, which permit to manage the information in a more efficient way.

The methodology selected for the temporal aggregation of data was the median metric. Nevertheless, this aggregation is not arbitrary and needs a preliminary analysis of data where the phenology of the main crops or each ROI was analysed (Wang et al., 2019 and reference therein). Spectral indexes time series from both optic and SAR have been generated by the strictly use of valid pixels (no clouds or shadows), capturing the spectral signature for every class to be determined. This methodology has been applied to a scope of the 4 last years. This process provides us the necessary information to determine the green up, peak and down of a particular crop, which were used to divide the phenological year into different stacks, which represent the target crops.
In order to achieve satisfactory and realistic results, we fed the model with field data coming from the ESYRCE dataset, which is the Spanish survey for crops and yield carried out by the Minister of Agriculture, Fishing and Food. This study is composed of a random selection of 1 km2 cells that cover between 8% and 10% of the cultivated area in the country. In order to minimize the effect of spatial autocorrelation while preserving the gradient of land cover type within each class, we generate a cloud of points from the original polygons, between 25,000 and 50,000 for each classification, randomly reserving the 25% for the validation of the results.

Results

The phenological approach for the temporal aggregation has led to achieve a very detailed crop classification. The integration of SAR (including backscattering and interferometric coherence) into the classifier has led to a direct improvement between 5% and 10% on the accuracy of the results according to the kappa indexes obtained in both ROIs. The incorporation of S1 data turned essential in the case of ROI-1, partially cover by clouds over the year, where the accuracy improved by almost 10%. However, this increasing had a special impact on the classification of woody permanent crops such as vineyards or olive groves with an improvement close to 15%. Nevertheless, our results indicated the margin of error could be minimised for this crops, as the accuracy for some specific zones only reached the 70%.
To solve this issue, the next stage of our research involved the application of our SR algorithm to super-resolve the S2 imagery. In first place, the linear correlation analysis carried out between the radiometric values for the original S2, the VHR and the S2 super-resolved imagery resulted in R² coefficient values close to 1. This allowed the use of the SR products to calculate accurate VHR indexes. Secondly, preliminary results indicate that the use VHR-S2 imagery defines correctly the border of the crop parcels, allowing the discrimination of those crops with areas that are close to the S-2 pixel resolution.

Our first tests indicate that the application of super-resolved S2 images to our classifier results in an improvement on the classifications, especially for permanent crops. This methodology has achieved kappa indexes greater than 0.90 for arable and woody crops. These results have been validated with the ground truth from the ESYRCE survey.

References

Ledig Ch., Theis L., Huszar F., Caballero J., Cunningham A., Acosta A., Aitken A., Tejani A., Totz J., Wang Z., Shi W., 2017, “Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network”. Computer Vision and Pattern Recognition. arXiv:1609.04802v5
Maskell G., Chemura A., Nguyen H., Gornott Ch., Mondal P., 2021, “Integration of Sentinel optical and radar data for mapping smallholder coffee production systems in Vietnam”. Remote Sensing of Environment 266 (2021) 112709.
Phan T.N., Kuch V., Lehnert L.W., 2020, “Land Cover Classification using Google Earth Engine and Random Forest Classifier—The Role of Image Composition”. Remote Sens. 2020, 12, 2411; doi:10.3390/rs12152411.
Wang H., Magagi R., Goïta K., Trudel M., McNairn H., Powers J., 2019, “Remote Sensing of Environment 266 (2021) 112709”. Remote Sensing of Environment 231 (2019) 111234

Within the context of agriculture, climate change poses challenges regarding long term planning of which crops to cultivate.
In this work we research the effect of extreme weather conditions such as drought and frost on the yearly yields for a variety of crops which are most relevant to the South Moravian Region in the Czech Republic, such as grapevine, wheat, and rapeseed.
We employ data retrieved from remote sensors, which can be useful in areas where ground meteorological stations are scarce.
By using remote sensed data and historical weather data we analyse the changes in frequency and duration of those events in order to predict which steps can be taken to reduce the impact of climate change on agricultural productivity.
For the assessment of droughts, we mostly focus on the Standardized Precipitation Evapotranspiration Index (SPEI), one of the well-established metrics for monitoring drought world-wide, and calculated it for several time scales. Positive correlation was found, for example, between rapeseed yearly yields and SPEI values for all time scales on the months December and April (r = 0.4 to 0.6). This suggests the high influence of weather conditions during these months on the growth of this crop.
For some plants, frost is very damaging during growth, while for others it is important. Therefore it is necessary to know when the last day in spring with a temperature of lower than 0° C. By applying this threshold as a criterion for a frost day, we calculate trends for the duration of frost-free periods and the occurrence of the last and first frost days each year.
While climate change increases risks of long drought periods and therefore poses challenges with regards to taking steps in order to mitigate effects of those events, it can also improve conditions for certain crops. Thus, another goal of this research is to find out whether some currently grown crops are likely to become unsustainable in the long term for the studied region, or if others that were not previously native may prove suitable.
We also investigate how phenological and agricultural phases of the different crops interact with these meteorological conditions.
Due to the high resolution of the sensor grid, we can also gain insights into local differences, e.g., instances the duration of a drought period differs between nearby areas.

Cocoa has been the main driver of deforestation in West Africa for decades, leading to substantial loss in rainforest, protected areas and national parks. The main global producer, Côte d’Ivoire and Ghana, experience the highest rates of increase in deforestation in the last years, yet accurate, transparent and country-wide mapping of cocoa farms is still an open issue, both in public and industry. Here, we use extensive cocoa farming data and publicly available Sentinel-2 satellite imagery in a deep learning framework to create country scale high resolution maps of cocoa plantations for both countries. Our cocoa map is the first to be validated by an in situ verification procedure. Our model can predict full-sun and agroforestry cocoa farms across countries with high precision generalizing to regions outside training area. In particular, our findings suggest dramatically larger harvesting areas compared to official figures, leading to diverse implications such as illegal growing and decreased yield per hectare.

To obtain a reliable neural network for large-scale cocoa detection, we have created a dataset that contains more than a 100,000 polygons including full-sun cocoa farms, agroforestry cocoa plantations and a diverse set of background boundaries, such as open land, villages, national parks and other crops. While this data is used as the ground truth reference, we feed Sentinel-2 images as input to our model. The network is trained on images captured within the last three years to diminish atmospheric noise and seasonal characteristics.

In particular, the training is performed in two steps. First, we generate a vegetation height map from a model trained with the same Sentinel-2 images and GEDI's lidar waveforms and subsequently used as an additional input to the actual classification network. The classification network is trained on 9 out of 13 Sentinel-2 bands and the vegetation height map. The output is a binary per-pixel mask indicating if a location contains cocoa plants or background. To increase robustness and reliability, the final cocoa map is averaged over ten differently initialized models and over a series of Sentinel-2 images from the same location.

Our final map is the first country-wide cocoa map that has been trained on a large dataset and the first to provide an additional per-pixel uncertainty quantification on the prediction. It has been validated in situ, highlighting the superior performance over existing maps. In particular, we achieve 96.5% precision and 86.1% recall on cocoa and an overall accuracy of 88.1%. Our results indicate extensive cocoa farming in large regions including illegal deforestation and plantations in protected areas and the need for improved cocoa maps in practial applications and decision making.

The latest European Commission report on the implementation of the Nitrates Directive expresses concerns about the nitrate concentrations in surface water in the EU (European Commission, 2021). This Directive is a main component of the Water Framework Directive (WFD) which requires all European surface waters – lakes, rivers, transitional and coastal water, and groundwater – to reach “good status” by latest 2027.
For the period 2016-2019, across all member states, nearly 15% of groundwater monitoring locations exceeded the nitrates concentration limit set for drinking water. Moreover, 81% of marine waters, 31% of coastal waters, 36% of rivers and 32% of lakes were reported as eutrophic (European Commission, 2021). These high nitrate levels may cause oxygen depletion and eutrophication, which negatively impact both human and ecosystem health. Excessive fertilization on agricultural fields is the main cause of nitrate leaching to surface and ground water resources.
The Nitrates Directive requires Member States to (i) identify waters affected and at risk of being affected by nitrates pollution and declare the areas draining into these waters as Nitrates Vulnerable Zones (NVZ) and (ii) develop action programs with specific measures for reducing and preventing nitrates pollution, and to reinforce these measures.
The latest reform of the Common Agricultural Policy (CAP) provides for tools to address nutrient pollution. CAP aims at giving farmers levers to protect natural resources, enhance biodiversity, and contribute to the fight against climate change. Through a number of rules and measures, CAP supports farmers in the sustainable use of nutrients. For example, farmers can receive financial compensation through CAP if they comply with a set of management requirements and good agricultural and environmental conditions, including measures from the Nitrates Directive.
In Flanders, the northern region of Belgium, the implementation of the Nitrates Directive was realized through so-called “Manure Action Plans” (MAPs) (. The cultivation of catch crops is one example of a mitigating action against nitrate leaching from the MAPs. These crops absorb residual nitrogen and prevent its leaching to the surface water. Farmers in areas where the water quality is too low need to cultivate these crops in winter. The implementation of the mitigating actions is monitored and enforced by the Flemish Land Agency (VLM). Non-compliance leads to financial penalties for the farmers. Farmers have to register the winter crops they will put on each of their parcels before the start of the season through the so-called ‘Single Parcel Registration’ (SPR). In this registration, they also have to report the intended sowing and harvesting dates. This is important, as the effectivity of catch crops is directly proportional to the length of their growing period.
VLM uses the data from the SPR registration to check the compliance to the MAP and calculate the associated financial compensation each farmer is due to receive. The Flemish Environmental Agency (VMM) also uses the data provided by the farmers to model the impact of catch crops on surface water nitrogen loads. Their algorithm calculates how much nitrate leaching to surface water is prevented by the cultivation catch crops. However, random field visits by the department of Agriculture has uncovered that for 30% of the agricultural parcels, the farmer reported a winter crop that was not actually present on the field. This means that farmers are overcompensated for crops they have never sown, and that the algorithm used for calculating nitrate leaching is using incorrect input data. It is not feasible to go out and check each individual parcel in the field.
Satellite images can offer a solution to the unreliability of the current data sources. The Sentinel-1 and Sentinel-2 satellites of the European Copernicus program have been collecting data since 2016. The spatial resolution of 10 to 20 meters and the revisit time of 5 to 20 days makes them a good source for the creation of time series for the identification of the crops and for estimating their sewing and harvest date. Remote sensing is increasingly used in policy support. It is an additional semi-automatic instrument which provides objective information in near real-time and also for less accessible areas. Based on this information, it is easier to decide whether and more importantly where field interventions are needed. This spatial guidance can thus support field campaigns and as such render the field data collection more efficient.
Catch crops specifically have already been detected using Sentinel-1 and Sentinel-2 images but they were rarely the main focus. Their short time on the field and the large diversity in crop types makes their identification challenging. Timeseries of a vegetation index derived from satellite spectral measurements can be exploited to model the phenology of vegetation. An often used vegetation index is the Normalized Difference Vegetation Index of NDVI. Extracting parameters describing the seasonal development (sewing date or harvesting date) is difficult, only a few methods have been developed for this purpose.
The objective of this research is thus (i) to detect the presence of catch crops and (ii) to estimate their sowing and harvest date using freely available Copernicus Sentinel images.
First, the common winter crops in Flanders were subdivided into 10 classes, where crops with similar growing patterns and phenology were combined. A separability analysis was conducted to gain insight into the spectro-temporal features needed to effectively discriminate between these classes. Then, the most discriminative features were used in a Random Forest model that classifies each parcel into one of these 10 classes.
The sewing and harvest date will be estimated using extracted NDVI time series from the Sentinel-2 images. These NDVI time series are a measure of phenology. As Sentinel-2 data often have data gaps due to cloud cover, especially in winter, we rely on a model-based combination of Sentinel-2 and Sentinel-1 to acquire sufficiently dense NDVI time series. We then fit a double logistic function to these NDVI time series to extract information about crop emergence and crop senescence or harvest.
As reference data, we have collected data about 1000 agricultural parcels across Flanders (100 parcels per class) using the CropObserve app developed by the European e-Shape project (https://e-shape.eu/). 70 parcels (7 per class) were monitored weekly during the emergence (September-October 2021) period, while another 930 parcels (93 per class) will be visited once between December 2021 and February 2022.


Reference
European Commission (DG Environment). 2021. Report on the implementation of Council Directive 91/676/EEC concerning the protection of waters against pollution caused by nitrates from agricultural sources based on Member State reports for the period 2016–2019.

Ongoing changes in climatic conditions, increased stress due to human activity and inadequate management contribute to both decreased grassland productivity and habitat quality. In order to manage grassland areas properly, and mitigate or avoid stress, precise information about grass growth conditions is needed. The main objective of GrasSat project is a fully operational system in form of desktop and mobile application, which provides a complementary tool for managing grassland production, mainly for medium and large farms in Poland and Norway. Combining the effectiveness of the application with the support of external advisors is the key to improve grass production management.

The methodology for monitoring grass growth conditions and yield forecast will be based on synergistic use of remotely sensed data, process-based grassland models and reference in-situ data, indispensable for elaborating reliable models characterizing plant development. In the first stage of work ca. 50 farms managing grasslands in valley and non-valley lowland sites were selected. A minimum of three observations have been taken around each of measurement plot. In the course of field campaigns the following in-situ measurements have been carried out in each test site: volumetric soil moisture - TRIME-FM probes and ML3 ThetaProbe Sensor, Leaf Area Index (LAI) - LAI2200 Plant Canopy Analyzer, Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) - AccuPAR LP-80, surface temperature - Infrared Thermometer, greenness index (SPAD) - SPAD-502 meter, height of vegetation - Jenquip EC20 Bluetooth plate meter, fresh and dry matter sward yield collected from 1 x 1 m samples (accomplished in laboratory), spectral reflectance - ASD FieldSpec hyperspectral instrument to assess diversity of grass vegetation.

At first stage of the works is to prepare classification maps of grasslands for the selected test areas, with the general division into high-productive grasslands and low-yielding unimproved grasslands. The classification approach will be based on applying time series of high-resolution Sentinel-2 satellite data for discrimination of grassland types. The classification maps will be necessary for further works related to grass yield modelling and for appropriate establishment of ground measurement networks. In addition to the analysis of the current satellite data acquired during the project, historical satellite data from Landsat 5 and Landsat 8 since 1995, and MODIS since 2000, will be analyzed for Poland. Vegetation indices derived from these images, and the relationships determined in the first part of works, will be used to map the general grassland conditions over 20 years and identify trends and changes between the years, and the best and worst conditions during this time interval. The changes will be correlated with meteorological data. To monitor and classify the drought areas additionally the data from Sentinel 1 satellite will be applied to determine soil moisture variation. Also the surface temperature from Sentinel 3, Landsat 8 and MODIS will be applied to analyse spatial and temporal variation of evapotranspiration for the seasons of 2020-2022. Evapotranspiration and vegetation indices from previous works will be analysed for drought determination. Also meteorological data and calculated Standard Precipitation Index (SPI) will be included into the drought model. Two indices calculated from Terra MODIS data as the Temperature Condition Index (TCI) and Vegetation Condition Index (VCI) will be applied to describe actual towards historical grasslands conditions and the deduction of drought. The last part of works will be concentrated on creating the website dedicated to presentation of the project’s results i.e. information on grass growth conditions, indications of areas affected with drought, winter kill and impacted by different management, as well as the prognosis of grass yields. In parallel, the mobile application will be developed in order to deliver the same products to individual farmers.

Using remote sensing to estimate the expected yield in grassland can help farmers to prepare for importing forage and to detect areas with high water stress. Novel approaches will be generally based on innovative use of satellite data in the grassland management to increase yield and monitoring of grassland status.

The research work was conducted within the project financed by the National Centre for Research and Development , titled "Tools for information to farmers on grasslands yields under stressed conditions to support management practices".

Sustainable management of water resources is of critical importance worldwide and particularly in the South-Mediterranean, which is suffering from water scarcity intensified by the impact of climate change. In this context, early detection of water stress by monitoring the water status of canopies can contribute to an optimal use of irrigation water use.
C-band radar data have shown great potential for monitoring soil and vegetation hydric conditions. Over forests, several studies have observed a diurnal cycle in the backscattering coefficient that can reach up to 1 dB between morning and evening measurements acquired by sun-synchronous satellites. This cycle is assumed to be related to the physiological functioning of trees, in particular to the diurnal cycle of the vegetation water content. Recently, some studies have shown the existence also of a diurnal cycle of coherence over tropical forests where they found that transpiration can be the main driver of the coherence decrease at dawn when winds are generally near zero. The objective of this work is to present a preliminary investigation of this behavior over annual crops such as wheat by assuming that water movement in the plant could lead to a daily cycle of the interferometric coherence and backscattering coefficient.
In situ measurements of the interferometric coherence and the backscattering coefficient are acquired every 15 minutes by a radar set-up composed of six C-band antennas (full polarization), installed on a 20 m high tower and targeting a winter wheat field in Morocco (LMI TREMA - TOSCA/CNES MOCTAR experiment). The field is equipped with an eddy covariance and weather stations that allow half-hourly measurements of evapotranspiration and wind speed. Measurement campaigns are also conducted to measure soil moisture, surface roughness, vegetation above-ground biomass and cover fraction.
The preliminary analysis of in situ radar acquisitions over the agricultural season from January to June 2020 reveals the existence of a diurnal cycle of the interferometric coherence whose amplitude increases with the canopy development. Concerning the diurnal variation, the drop of interferometric coherence at dawn is concomitant with the increase in evapotranspiration which may indicate that this drop could be due to the sapflow. In contrast, the lowest coherence values at the end of the afternoon are rather related to wind peaks. For the backscattering coefficient, a good agreement is observed between the evolution of its daily average and the evolution of evapotranspiration. These results, which need to be consolidated, demonstrate the existence of important dependencies between the C-band response and the physiological functioning of wheat, which opens insights for the monitoring of crop water status using radar data acquired at sub-daily timescale. This rather highlights the interest of a future geostationary radar mission.

Soil moisture has been identified as an essential variable to be monitored through remote sensing. Sentinel-1 has a strong potential for this, yet in vegetated areas the backscatter response is complex, as it is not only influenced by the dielectrical characteristics of the soil, but also by the physical structure and the dielectrical characteristics of plant elements, among others. Therefore, soil moisture retrieval over vegetated areas requires the coupling of vegetation and soil backscatter models in order to separate both contributions.

Wheat is the most important crop worldwide in terms of cultivated area. Previous research showed that, in C-band and VV polarization, wheat plants attenuated the surface scattering component from the underlying soil during a significant part of its growth cycle, coinciding with the stem elongation phase. This behavior needs to be modeled or corrected before soil moisture retrieval is attempted. The objective of this work is to propose a new method for wheat attenuation correction (WATCOR) that is applicable to Sentinel-1 VV time series.

With this aim, a large dataset of ~80,000 wheat parcels observed by Sentinel-1 during four years was built. Three Sentinel-1 orbits were used making a total of 563 Sentinel-1 acquisitions processed. A descriptive analysis revealed the systematic seasonal pattern of wheat, with the characteristic attenuation period between the stem elongation and booting phases of its phenological cycle. Then, the new correction method (WATCOR) was designed, based in sound time series analysis procedures like time series smoothing, change point detection and lower envelope curve delimitation. The main hypothesis of WATCOR was that with no attenuation, backscatter time series during the agricultural season would be composed of a high-frequency component caused by soil moisture dynamics over a smooth low-frequency trend. The method does not require any external vegetation descriptors nor parameter fitting, and can be applied by only indicating rough periods, according to local phenology, when attenuation is likely to occur.

Finally, WATCOR was evaluated in a case study where the soil moisture content of two wheat parcels was measured during three agricultural campaigns (six parcels in total). As a measure of the success of vegetation correction, the case study relied on the correlation between backscatter and ground measured surface soil moisture, comparing the correlation coefficients obtained before and after correction and their eventual improvement. For comparison, the Water Cloud Model (WCM) was also applied, considering three different variants that used different vegetation descriptors (NDVI, NDWI and VH/VV ratio). Results showed a rather low correlation between the original backscatter time series and surface soil moisture (with a median R=0.14). Correlations improved to R=0.24-0.32 for the WCM, with NDVI appearing as the best option. WATCOR obtained the highest correlation, with a median R=0.47. Furthermore, the corrected backscatter time series showed a successfully removal of wheat attenuation. Altogether, WATCOR appeared to be an appealing vegetation correction methods, due to its relative simplicity and independence to external data. Future studies should evaluate WATCOR as a pre-processing step before soil moisture retrieval with other state of the art methods, to confirm its validity and interest.

Agriculture contributes to air pollution and is affected by atmospheric composition, meteorology and climate change. One of the important gases emitted from agricultural activities is ammonia (NH3), which makes up a large portion of the anthropogenic reactive nitrogen in the environment. Agricultural ammonia emissions contribute to the formation of inorganic fine particulate matter (PM2.5), causing multiple negative effects on human health and the overall air quality. Since NH3 abundance drives PM2.5 formation, reduction of its emissions can make an important contribution to air quality control. Several studies proved the efficiency of the Infrared Atmospheric Sounding Interferometer (IASI) instrument aboard the Metop satellites in measuring ammonia from space. The instrument was launched in 2006, providing since a continuous view of the global atmosphere and allowing us to study many pollutants relevant to air quality. With IASI, we were able for the first time to retrieve NH3 concentrations from space.

In this presentation, we explore the interaction of atmospheric ammonia with land and meteorological conditions. We look at the temporal variability of ammonia in different regions of the world, mainly agricultural fields. The relationship land-ammonia is assessed by comparing the variability of surface land surface temperature product (skin temperature) from the European centre for medium-range weather forecasts (ECMWF) latest reanalysis (ERA5) with IASI NH3 total columns. The meteorology-ammonia relation is examined, by looking at air temperature, humidity, precipitation, and wind speed. The regions examined have been identified as point sources and/or hotspots of ammonia of agricultural sources.

Meteorology is not the only factor affecting the abundance of ammonia in the atmosphere. In fact, human-induced events, such as wars and conflicts affect land-use management and the abundance of agricultural ammonia. As an example, we show how the war in Syria affected land use and NH3 concentrations. We examine the changes in NH3 close to a fertilizer industry in central Syria, whose activities were suspended due to conflict-related events. We also explore the effect of war-induced land use/land cover changes on agriculture-emitted ammonia in north-east Syria that has witnessed battles between different groups. The interpretation of the changes in NH3 is supported by different datasets: visible satellite imagery to assess the effect on industrial activity, ERA5 (temperature, wind speed, and precipitation), and land cover and burned area products from the moderate resolution imaging spectroradiometer (MODIS) to examine land use/land cover changes and fire events during the study period. We show that the NH3 columns are directly affected by the war. Periods of intense conflict are reflected in lower values over the industry reaching -17%, -47% and -32% in 2013, 2014 and 2016 respectively, compared to the [2008 – 2012] average, and a decrease reaching -14%, and -15% in the croplands’ area in north-east Syria during 2017 and 2018 (compared to 2011), respectively. Towards the end of the control of Islamic State in Iraq and Syria (ISIS), an increase in atmospheric NH3 was accompanied by an increase in croplands’ area that reached up to +35% in 2019 as compared to pre-war (2011). This study shows the relevance of remote-sensing data of atmospheric composition in studying societal changes at a local and regional scale.

In a context of climate change and population growth, agriculture is starting to face new challenges. More frequent and severe droughts are leading to an increase in crop water demand. In Europe, the new Common Agricultural Policy (CAP) reform encourages farmers to use their water supply in a safe and sustainable way. Safeguarding water is also a key aspect of the European Green Deal. One cross-compliance rule relevant to water includes compliance with authorization procedures for irrigation (GAEC 2). Having that in mind, regular information on the spatio-temporal variation of variables driving water and energy cycles, such as evapotranspiration (ET) is required to support decision makers as the climate changes. In agriculture, ET is a key variable to identify crops water deficit and forecast potential yield. A few in-orbit instruments can provide the necessary energy variables as well as the biophysical state of the land surface to feed ET models. Until now, there was no combination of spectral, temporal, and spatial resolution to estimate ET at parcel-level. The newly developed tool Sentinels for evapotranspiration (Sen-ET), recently launched by the European Space Agency (ESA), opens that possibility.
The Sen-ET open source code makes use for the first time of the unique combination of Sentinel 2 (S2) and Sentinel 3 (S3) to derive ET at agricultural-parcel scale resolution (20-30m). The multispectral observations of S2 and the sharpened S3 SLSTR thermal band at S2 spatial resolution are the main inputs of Sen-ET. Combined with meteorological and land cover data as well as a digital elevation model, ET is estimated through the Priestley-Taylor Two-Source Energy Balance Model (TSEB). While this is a major step forward for ET estimate building on the unique four Sentinel satellites constellation, the performance of the ET retrieval is still to be assessed along agricultural seasons for given crop types and the source of errors characterized. The objective of this study is thus to assess its accuracy over a field in Belgium, equipped with an ICOS Flux tower ecosystem station. For the years 2017 to 2020, four variables of interest provided by the station, naming the radiation Rn, the soil heat flux G, the sensible heat flux H and the latent heat flux LE are used to assess the accuracy of the energy balance modeled based estimate of ET for 3 crops of the field rotation naming winter wheat, potatoes and sugar beet. Besides, since 2017, Belgium has known consecutive and unprecedented drought periods up until 2020. This latter year was recently recognized as an agricultural disaster by the Walloon government for the period from May to September. Therefore, thanks to the overlapping availability of sentinel data, the actual and reference ET are calculated from ICOS data to assess the ability of the SenET estimate to capture crop water stress episodes. The impact of the satellite observation configuration and conditions on the estimation is statistically analyzed which includes the effects related to the viewing and azimuth angles with regards to the sun position (i.e. the bidirectional effect on surface reflectance being higher for the low sun angles occurring with S3). The potential improvement of accuracy of SenET estimate over cropland is thus evaluated by defining the optimal conditions for Sen-ET use and adapting its generic ancillary input data with ones more suited to our area of interest in terms of spatial resolution.

The heterogeneity of soil properties is the one of the main reason for the spatial distribution of crop yields within the plots of arable land. These are mainly differences in agrochemical soil properties and soil moisture availability, which can be a limiting factor for crop yield. The implementation of precision agriculture procedures, such as variable rate application of fertilizers, variable seed rate and machinery guidance by GNSS seems to be appropriate management on these plots. With regard to the time and cost labor for traditional soil mapping, more effective methods are evaluated to capture the field spatial variability with high reliability, informative value and optimized costs for soil sampling. The refined soil maps together with crop yield maps serve for improvement of nutrient balance model and more efficient use of agrochemicals.

As part of the solution of the Czech national project NAZV QK21010247 (2021-2024), the techniques of advanced mapping of soil physical and chemical properties within the fields, their combination with soil moisture monitoring and implementation into crop management technologies by utilization of precision farming tools. For these purposes, the agricultural enterprise Rostěnice a.s. (Czech Republic) with the managed area more than 10,000 ha of arable land was chosen for the research activities. The research part of the project is divided into three areas: (1) Digital soil mapping for the whole-area evaluation of agrochemical properties of soil and its usage for variable rate application of P, K and Ca fertilizers; (2) analysis of crop yield zones and their implementation into nutrient balance models; (3) identification of soil moisture conditions within the fields for correction of site-specific crop management practices, which will lead to stabilization of annual crop yield differences affected by climate change.

For the purposes of digital soil mapping, ground survey and remote sensing methods are combined to design directed soil sampling and to refine the estimation of soil properties. Spatial prediction of soil properties is based on methods of digital soil mapping using data-oriented machine learning approaches (regression trees, neural networks, including convolutional and combined ensemble methods). This step also includes the implementation of soil moisture surface analysis from the Sentinel-1 SAR radar data and the identification of the basic topsoil properties from the Sentinel-2 optical data. Initial research results indicate the robust prediction capability for assessment of soil texture and soil organic matter content from multispectral images of Sentinel-2. At the same time, the underlying area maps from satellite data and DMT are used to optimize the distribution of sampling points across the fields. Information on the spatial distribution of crop yield zones is elementary to refine nutrient balance models. For this, the implementation of yield records obtained by harvesters or yield potential from the time-series of Landsat and Sentinel-2 multispectral data are verified. The results of yield maps analysis showed high differences of relative crop yield within the fields (from 45 to 140 %) which reflects the topography and soil type condition within the selected region of interest. Final application rates of P, K, Mg and Ca fertilizers are determined by the combination of soil maps and distribution of yield zones with a final adjustment respecting the of settings of application machinery.
Mapping of soil moisture as the crop yield limiting factor provides a crucial information to maintain yield stability and increase productivity in low-yield zones. An automated system for monitoring of soil moisture with the validation for dry and wet season is under the development incorporating satellite SAR data of Sentinel-1. Two different approaches are assumed for the determination of soil moisture: a) Change detection method (Wagner, 1999) with the adaptation to crop growth phases using Sentinel-2 optical data, (b) WCM - water cloud model (Attema and Ulaby, 1978) calibrated using ground measurements and Sentinel-2 data using the machine learning principle. The research activities include verification, testing of accuracy, applicability and adjustments for optimal implementation in practice. The model is validated using continuous and ambulant in-situ soil moisture measurement data.
The research project is focused on the application of new procedures for mapping of the spatial variability of soil condition by using Earth Observation data. Their implementation in the form of site-specific crop management can lead to reduction of excessive use of agrochemical inputs and optimal land use while maintaining the soil productivity and ensuring the production of quality agricultural commodities with minimal impact on the environment.

The green area index (GAI) is a key indicator of crop status and is thus widely used for crop growth and stress monitoring, as well as for yield forecasting. Satellite remote sensing is a highly effective means of providing GAI estimates in a timely manner, on a large-scale, and with high temporal frequency. While current techniques relying on optical remote sensing allow for high accuracy GAI retrieval, they are, in many regions of the world, frequently impeded by the presence of clouds obscuring the view of the sensors. This seriously limits the temporal density of GAI estimates and the robustness of operational crop monitoring systems. Contrary to optical sensors, synthetic aperture radars (SAR) do not suffer from this issue.
One of the most widely used approaches for GAI retrieval from SAR data is based on the inversion of the Water Cloud Model (WCM), a semi-empirical model which performs well with even a modest amount of calibration data. However, methods exploiting the WCM often rely on in situ measurements of surface soil moisture to estimate the GAI of a crop. This renders the use of the WCM for operational crop monitoring impractical, even on a moderately large scale.
In this study, we first tested the sensitivity of SAR data in L- and C-band to the GAI and surface soil moisture of maize crops. We then developed an algorithm to retrieve the GAI of a crop based on a machine learning approach using the WCM. The WCM requires SAR backscatter data, the incidence angle of the radar beam, and an estimate of the soil moisture under the canopy to estimate the GAI and vice versa. Calibrating a pair of WCMs in two different polarizations therefore allows in principle to retrieve both variables of interest simultaneously. However, the resolution of the system of WCMs is an ``ill-posed'' problem as several combinations of GAI and soil moisture can correspond to a same SAR backscatter value. To overcome this problem, we used a random forest (RF) regressor trained with backscatter data in two distinct polarizations for each frequency band simulated with a pair of WCMs for a large set of GAI, soil moisture, and incidence angle values. This method allowed for the simultaneous estimation of GAI and surface soil moisture in maize crops from bi-polarized SAR backscatter measurements and their incidence angle alone, without the need for additional field measurements beyond the calibration of the model.
The algorithm was tested on an innovative data set acquired in 2018 in Belgium during the so-called BELSAR-Campaign. The data consist of full-polarization L-band airborne SAR data, bi-polarized C-band SAR data from Sentinel-1, and simultaneous field measurements of biogeophysical variables of the vegetation and underlying soil in eight maize fields. Due to the relatively small size (n=26) of the data set, a leave-one-out cross-validation was performed. Overall, the algorithm performed significantly better with L-band data than with C-band data. Estimates for the GAI in the most accurate configuration, i.e. in L-band and with the vertical cross- and co-polarization pair VH and VV, returned a RMSE of 0,77 m²m⁻². Results were particularly encouraging for later stages of crop development, between the stem elongation phase and the harvest. At the beginning of the growing season, however, the algorithm tended to overestimate the GAI. This issue might be explained by the relatively small, and therefore biased, calibration data set combined with high sensitivity of the WCM. Regarding surface soil moisture retrieval, neither the estimates obtained from L-band data nor those obtained from C-band data seemed usable for accurate soil moisture monitoring under the canopy. This is likely related to the poor sensitivity of backscatter measurements in both frequency bands to soil moisture in well-developed maize crops, as highlighted by the sensitivity analysis.
In conclusion, this study demonstrates the potential of bi-polarized SAR backscatter data in L-band, and to a lesser degree in C-band, and their incidence angle for the operational monitoring of the GAI in maize crops.

The spaceborne imaging spectroscopy mission PRecursore IperSpectrale della Missione Applicativa (PRISMA), recently launched by the Italian Space Agency (March 2019), represents the first example of forthcoming spaceborne hyperspectral missions for Earth Observation (EO). PRISMA, with its high spectral resolution in VNIR-SWIR region and high spatial information from the additional panchromatic camera, opens new opportunities in many scientific domains, including precision farming applications and contributions to sustainable agriculture monitoring. Crop biophysical variables (BVs), such as Canopy Chlorophyll Content (CCC) and Canopy Nitrogen Content (CNC), provides important information on crop nutritional status, allowing to map Plant Nitrogen Uptake (PNU) which is essential for the rational implementation of site-specific crop fertilisation strategies. In this context, the accurate assessment of key BVs from this data stream provided by new spaceborne hyperspectral sensors, requires reliable, efficient and computational cost-effective approaches.
This study evaluates the performance of the hybrid approach (HYB) for the retrieval of CCC and CNC to estimate PNU, from PRISMA data. The tested HYB approach combined the radiative transfer model PROSAIL-PRO and the machine learning algorithm Gaussian Process Regression (GPR). The PROSAIL-PRO model was calibrated to generate a look-up-table of maize reflectance spectra, associated to the corresponding CCC and CNC values. This database was used to train GPR models for the estimation of CCC and CNC. The developed HYB models for CCC and CNC were applied to PRISMA-like data, simulated from images acquired by the airborne hyperspectral sensor HyPlant in the 2018 Italian ESA-FLEX campaign. The validation of the estimated BVs, performed against in-situ data, showed promising results (R² = 0.79, NRMSE = 0.19 and R² = 0.84, NRMSE = 0.15 for CCC and CNC, respectively). Estimated CCC and CNC values were also used to find a linear regression model with PNU, which proved to be slightly better correlated to CNC than CCC (R² = 0.82 and R² = 0.80, respectively). The HYB model for CNC was then applied to two actual PRISMA images, acquired in summer 2020, and the CNC-PNU relationship was finally used to generate PRISMA PNU maps for the study area. These PNU maps showed values within maize expected ranges and temporal trends compatible with plant phenological stages identified according to different sowing dates and crop development. These findings open new perspectives for the generation of spatial explicit geo-products related to crop nutritional status, representing a fundamental prerequisite for the adoption of precision farming management strategies.

Agriculture production boosted since the end of WWII and increased linearly starting from the early 1960s. This increase caused several environmental impacts such as emissions of greenhouse gasses, nitrogen leaching in groundwater and water bodies eutrophication. A great share of these side effects is due to fertilizer exploited to increase yield. Moreover, in the last 20 years, a stagnation of yield was observed in some western countries. It is likely that in these countries crop yield reached its physiological maximum given current climate condition, available plant genotypes and means of production.

As a consequence, given the increasing global food demand and climate change issues, a promising way to increase or maintain cropping systems productivity is to increase the efficiency, basically by producing more with less. Promising tools to do that are information extracted from remote sensing data to monitor field condition in space and time and technological innovations of digital agriculture to take site-specific management strategies.

In this framework, this work shows a validation of crop biophysical parameters estimated from Sentinel 2 and PRISMA data with i) a machine learning (ML) and ii) a hybrid (HYB) approach. Once validated, the estimated maps are exploited to investigate their potentiality for an early detection of corn yield performance.

Side goals of the work are i) comparing the goodness of ML and HYB approaches in estimating LAI ii) checking the temporal coherence of biophysical parameters estimated with HYB using PRISMA and multispectral Sentinel-2 data and iii) exploring the contribution of canopy nitrogen content (CNC) from hyperspectral imagery to investigate yield patterns (among and within fields).

Estimation are obtained for the 2021 summer season using Sentinel-2 and PRISMA imagery and validated with field data collected by SARAGRI project in Bonifiche Ferraresi farm (Jolanda di Savoia - Italy). The estate covers about 3850 ha mainly cultivated with winter and summer crops. Corn was chosen for the present study because it is of great agro-economic importance. A field of 26 ha was identified and 3 elementary sampling units (ESUs) were placed on it considering soil and vegetation variability.
Field campaigns of vegetation biomass, LAI and leaf chlorophyll concentration were conducted from sowing till harvesting (6th May to 12th August) obtaining 21 averaged samples of biomass and 45 of LAI and chlorophyll concentration. All measurements are collected following the recommendations of the (VALERI) protocol (w3.avignon.inra.fr/valeri/). Yield maps and ancillary information (sowing dates, corn variety, etc.) were provided for all the corn fields of the estate.
Sentinel-2 data were automatically downloaded and masked from cloud cover for the whole 2021 summer season using the R package {sen2r}, while PRISMA data were provided in the framework of ASI-CNR PRISCAV project. Two cloud-free hyperspectral data were available on the study area in the period covered by field campaigns (4th and 21st June).

LAI time-series, obtained for all cloud-free Sentinel-2 data, were obtained exploiting both the i) ML model set in De Peppo et al. (2021,) based on Gaussian Process Regression (GPR) algorithm and ground data for different crops in the calibration phase, and ii) HYB developed specifically for corn exploiting PROSAIL-PRO model and GPR algorithm as described in Ranghetti et al. (2021). HYB model calibrated on PRISMA spectral configuration allowed also to estimate CNC maps. Validation of LAI estimations from the two methods was performed using field data acquired along the season.

Finally, biophysical parameters maps were analyzed to assess which seasonal metric is able to highlight differences in crop yield comparing fields with the same variety and management. Metric tested are single LAI map estimated at different key phenological stages (e.g. sixth-leaf, tassle, etc.) and LAI seasonal cumulative (e.g. start of season to peak, start of season to end of season, etc.). CNC maps estimated from hybrid model and PRISMA data in key moment were also analysed in order to assess the added value of nitrogen uptake information to analyse yield variability among fields and within field.

To conclude, the results show approaches to estimate crop parameters that are suitable to be exploited in near-real-time applications thanks to independence from field data. Information derived from estimated time series, once provided to farmers with the timing requirements of farm activities, can be exploited as an early warning useful to set management strategies (also site-specific) with modern tools of digital agriculture.

Agricultural production relies on a few breadbasket regions (i.e. Europe, USA, Russia, China), who are the largest producers and exporters of wheat, one of the top four most consumed cereals worldwide.Wheat production is heavily dependent on accurate climate, soil and management conditions in order to maximise yields. Climate change triggers changing crop growing conditions, especially in rainfed production systems. Yield forecasts based on analyses of past yield and climate time series have shown that wheat yields will probably benefit from climate change in northern latitudes which will experience a longer growing season, whereas in southern countries yields will decrease, due to high temperatures and extended droughts. With climate change, anomalous weather or climate events (AWCE) frequencies increase too. AWCE, in an agrometeorological context, can be described according to three criteria: (i) magnitude, (ii) duration and (iii) timeline. For example, in the case of a late frost event the magnitude refers to the question “how cold?”, the duration answers the interrogation “for how long?” and the timeline assesses “during which growing stage did the event happen?”. AWCE are responsible for major crop failures, endangering global food security and impacting markets. For instance, the 2007 late frost event in Kansas, happening during the flowering stage of wheat, nearly caused a total crop failure state-wide. Similarly, 80% of the United States’ agricultural land was impacted by a destructive drought in 2012, accounting for $14.5 billion in crop insurance payments. Remote sensing data provides insights on intra-year crop dynamics and has been extensively used for crop monitoring and yield forecasting. However, there are only a few studies focused on using remote sensing for assessing the impacts of AWCE on wheat yields and phenology. In this study we aim to answer two research questions: (a) Given the three criteria in the AWCE definition, when does wheat yield get significantly impacted by an AWCE? (b) Does the phenological curve of wheat during a yield reducing AWCE year significantly diverge from a “normal year” curve? To answer those questions, we first identify low yield anomalies in historical yield records over major wheat producing countries. We then assess whether or not an AWCE happened during those years and collect information about magnitude, duration and timeline of the event. Finally, we analyse the phenological response of wheat during AWCE occurrence years and areas, compared to curves in optimal growing conditions.

Estimating and forecasting wheat yield at different relevant scales with sufficient accuracy for decision-making remains a major scientific challenge from crop growth modeling and Earth Observation (EO). SenCYF (Sentinel for Crop Yield Forecasting – ESA Open Call Project) research strategy relies on a unique data set combining multi-sensor high-resolution satellite time series (TS) and a set of more than 10,000 yield in situ data at holding level distributed over whole France, the main EU winter wheat producer.
Machine Learning (ML) algorithms were trained to relate the reference yield and a large set (n=71) of potential explanatory variables and yield’s proxies derived from optical EO (Sentinel-2 and Landsat 8), meteorological (ERA5-Land) time series, and simulations from a crop growth model calibrated at parcel level for the three studied years (2017 to 2019). Optical EO data were pre-processed from level 2 to biophysical variables, and TS were smoothed. A logistic function was fitted based on the EO observations to aggregate over time the explanatory variables. ML models were designed to provide a yield estimation at three times of the season: one month before the harvest, after the harvest, and after the acquisition of the yield survey.
At parcel level, 61% of the yield variability was explained by models trained on the current year’s survey, without any bias. The transferability of models trained on other year surveys, applied to the current year, better retrieved the yield variability (56.6% for estimation after harvest and 54.8% for in-season forecasting) than most of the recent comparable literature. The most relevant features were the parcel location (latitude, longitude), being a synthetic proxy of climate and soil conditions, and the maximum of vegetation related variables derived from satellite remote sensing. The peak of vegetation occurs early enough for the feature extraction of the forecast mode, explaining the close performance between both modes. Parcel-level estimations were aggregated at the NUTS3 level and compared to official sub-national statistics. The performance considerably increased with the level of estimation. Including yield observation of the current year in the calibration dataset, estimation models retrieved between 84.1% and 91.7% (depending on the year) of the NUTS3 yield variability according to the different years. Similarly, high coefficients of determination from 77.8 to 81.5% were obtained for the in-season forecasting mode.

Droughts lead to crop yield losses worldwide. The severity of a crop yield loss depends on the stress duration and on its intensity. Crop growth models describe the plant development throughout the growing season using soil and meteorological data as input. The AgroC crop growth model describes the plant development as well as its carbon and water fluxes with an hourly resolution (Klosterhalfen et al., 2017). To describe the carbon fluxes, it makes use of the mechanistic description of photosynthesis (Farquhar et al., 1980). Key variables in this approach are the photosynthetic electron transport Jp and the maximum carboxylation capacity VCMax. Both variables reduce in case of a drought stress. Simultaneously, the non-photochemical quenching component (NPQ) tends to go up. Both these effects cause a change in the emission of chlorophyll fluorescence. This variable is quantified by means of the fluorescence emission efficiency (εF), which corresponds to the top of canopy sun-induced chlorophyll fluorescence SIF, normalized for photosynthetically active radiation (PAR) and canopy structure, based on the Fluorescence Correction Vegetation Index (FCVI; Yang et al., 2020). The SIF, as well as the PAR and FCVI are measurable at the canopy scale with a field spectrometer.
To quantify the drought stress, the AgroC model makes use of a stress factor (Feddes et al., 1974). Here, the stress factor (α) is described in function of the soil water availability, expressed in pressure heads. The formula contains a threshold value h3, describing the pressure head by which the plant switches from a light- to a water limited regime. This parameter is to be determined for every field individually, as it depends on various soil and plant properties.
By coupling the AgroC crop growth model to the Soil Canopy Observations Photosynthesis Energy (SCOPE) model (Van Der Tol et al., 2009), it is possible to forward model εF for a given plant size and a given stress condition. Inversely, the εF can therefore serve as a real-time proxy for plant stress conditions. By feeding field-based observations of εF over a sugar beet stand to the AgroC crop growth model, the value of h3 could be determined through an inverse estimation. With the better estimate of h3, an improvement in the agreement between the AgroC-based carbon/water fluxes and the fluxes measured by a reference eddy-covariance station is found. Using only a limited number of εF observations, it was possible to calibrate the h3 parameter (De Cannière et al., 2021).
The presentation proposes a framework, in which εF is used as a real-time proxy for the plant stress status in order to calibrate a water stress function, is particularly interesting in the light of the FLuoscence EXplorer (FLEX) mission. This satellite mission will sample sun-induced chlorophyll fluorescence data at the global scale with a spatial resolution of 300 m. The mission comes with a temporal resolution of 27 days, allowing sampling a few observations of εF over a plot during the growing season. This would allow calibrating a stress function for each FLEX pixel. A more precise and site-specific calibration of a water stress function improves the modelling of carbon and water fluxes in land surface models as well as for crop yield forecasts.

References
De Cannière, S., Herbst, M., Vereecken, H., Defourny, P., Jonard, F., 2021. Constraining water limitation of photosynthesis in a crop growth model with sun-induced chlorophyll fluorescence. Remote Sens. Environ. 267. https://doi.org/10.1016/j.rse.2021.112722
Farquhar, G.D., von Caemmerer, S., Berry, J.A., 1980. A biochemical-model of photosynthetic CO2 assimilation in leaves of C3 species. Planta 149, 78–90.
Feddes, R.., Bresler, E., Neuman, S.P., 1974. Field test of a modified numerical model for water uptake by root systems. Water Resour. Res. 10, 1199–1206.
Klosterhalfen, A., Herbst, M., Weihermüller, L., Graf, A., Schmidt, M., Stadler, A., Schneider, K., Subke, J.A., Huisman, J.A., Vereecken, H., 2017. Multi-site calibration and validation of a net ecosystem carbon exchange model for croplands. Ecol. Modell. 363, 137–156. https://doi.org/10.1016/j.ecolmodel.2017.07.028
Van Der Tol, C., Verhoef, W., Timmermans, J., Verhoef, A., Su, Z., 2009. An integrated model of soil-canopy spectral radiances, photosynthesis, fluorescence, temperature and energy balance. Biogeosciences 6, 3109–3129. https://doi.org/10.5194/bg-6-3109-2009
Yang, P., van der Tol, C., Campbell, P.K.E., Middleton, E.M., 2020. Fluorescence Correction Vegetation Index (FCVI): A physically based reflectance index to separate physiological and non-physiological information in far-red sun-induced chlorophyll fluorescence. Remote Sens. Environ. 240. https://doi.org/10.1016/j.rse.2020.111676

The objective of this study is to evaluate the ability of data driven approaches to derive GAI and fAPAR from LANDSAT-8 and SENTINEL-2. They rely on machine learning algorithms trained on actual ground measurements and satellite reflectance data. The use of ground data is expected to mitigate issues related to 1D and 3D radiative transfer model inversion, such as model realism, and the representativeness of the simulations including crop architecture, optical properties of the canopy elements, and background reflectance. However, data driven approaches require a large training dataset with a wide and representative diversity of growth stages, background, and environmental conditions.
We measured GAI and FAPAR over ten crops and four sites located in France and Belgium during the 2017-2018 season, representing about 800 data points. We complemented this dataset by 17 additional ones available within the scientific community distributed over the United States, South America, Eastern Europe and China. This resulted into more than 4000 data points acquired with a consistent methodology mostly based on hemispherical digital photos.
Since some of the ground data were acquired before the launch of SENTINEL2, we developed a machine learning regression algorithm (MLRA) to estimate SENTINEL2 reflectance from that of LANDSAT-8 and reciprocally. For this purpose, we selected 427 agricultural sites among the EUMETSAT SAVS database and extracted simultaneous acquisitions (+/- 1 day) of LANDSAT-8 and SENTINEL-2 during 2019. To account for the differences in spatial resolution between both sensors and the registration uncertainty, we used a 120 m x 120 m window. Results show very good performances when estimating TOC and TOA reflectance of one sensor from the reflectance of the other for all the bands considered. Marginal effects of the acquisition conditions were observed (atmosphere, sun and view angles, spectral differences).
We selected the most optimal MLRA to estimate GAI and FAPAR from LANDSAT-8 and SENTINEL-2 TOA and TOC reflectances. The whole dataset was carefully split into training and validation datasets. We discuss the performance and robustness of the algorithm depending on the input data (TOA or TOC reflectances) and size of the training dataset for each crop, and how it compares with a 1D generic radiative transfer inversion algorithm.

Remote sensing is a key technology in agriculture, since it provides invaluable spatial-temporal data for supporting the decision-making in issues related to fertilization, irrigation or pest management. The products most often used in the agricultural context are those obtained from optical multispectral sensors, i.e. reflectance values in the visible, NIR and SWIR regions of the spectrum. These can be interpreted in terms of biophysical parameters of interest (e.g., pigment content, cell structure or wetness). Yet, multispectral information might be too large and even redundant, so it can be efficiently summarized through vegetation indices, like the Normalized Difference Vegetation Index (NDVI), which has become a standard product for agricultural applications.
Many agricultural applications require timely observations that might be interrupted by the presence of clouds or by failures in the system. This introduces uncertainty and reduces the reliability of these data sources for modeling and monitoring the land surface. Gap filling techniques make use of valid observations to estimate the missing data, assuming that the seasonal path of vegetation growth can be adjusted using a mathematical model. The objective of this study is to evaluate two gap filling techniques in NDVI time series obtained by Sentinel-2. With this aim, two gap filling techniques are evaluated, one that used as input Sentinel-1 data and another one based on the NDVI time series itself. The two methods were applied on a synthetic analysis that was carried out by creating gaps of different durations at different points in the time series.
The study area was an agricultural district of ~73,000 ha, located in the province of Navarre (Spain), where more than 73% of the study area was dedicated to agriculture. The main crops in the area were rain-fed herbaceous crops (mainly barley and wheat), followed by irrigated herbaceous crops (mainly corn, wheat, barley and sunflower) and irrigated woody crops (mostly vineyards). The Land Parcel Identification System (LPIS or SIGPAC in Spanish) vector file was available for this study, as well as an anonymized version of farmers’ Common Agricultural Policy (CAP) declarations for year 2019. In total ~28.000 agricultural parcels and 95 different crop types were present in the area. However, the study focused on a selection of the five most representative crops (wheat, barley, maize, sunflower and vineyards), that were randomly sampled to obtain a 150 parcel sample per crop, that was subsequently divided at random into training and test datasets.
The study period ranged from September 2018 to December 2019. All available Sentinel-1 (S1) scenes and Sentinel-2 (S2) scenes with a cloud cover lower than 60% were downloaded and processed. Sentinel-1 processing consisted on thermal noise removal, orbit correction, calibration, speckle filtering, terrain-flattening and orthorectification, resulting in backscatter coefficient values (gamma0) in VH and VV polarizations. Sentinel-2 processing consisted on an atmospheric correction using Sen2cor and a mosaic of the four granules comprising the study area. Finally, the NDVI and four different Sentinel-1 features (VH, VV, VH/VV and VHxVV) were calculated and their time series per field extracted.
The first gap-filling technique (S1GF), reconstructed missing data in the NDVI time series by using univariate linear and polynomial regression models trained on Sentinel-1 features. The second technique (S2GF), was the methodology proposed by Chen et al. (2004), consisting of the recursive application of the Savitzky-Golay (SG) filter to the NDVI time series. Both techniques were evaluated by calculating the RMSE between the reconstructed NDVI time series and the real NDVI time series.
The results showed that, in general, the S2GF technique outperformed S1GF, especially for short gaps (1 to 4 S2 acquisitions). In extreme cases with very long gaps (15 acquisitions or longer) S2GF was unreliable and S1GF obtained generally lower RMSE values. For intermediate cases (5 to 15 acquisitions), results varied depending on the particular crop studied, and on the date and duration of the gap in the time series. It appeared that barley was the crop most favorable to S1GF, in particular when gaps occurred in early spring or during crop senescence. On the contrary, vineyards time series, with a less marked seasonal pattern, produced consistently best results with S2GF, regardless of the duration and moment of the gap. Future studies might extend this analysis to other crops and land covers, and eventually explore other modeling techniques to take advantage of Sentinel-2 and Sentinel-1 complementarity.

BIBLIOGRAPHY
Chen, J., Jönsson, P., Tamura, M., Gu, Z., Matsushita, B., & Eklundh, L. (2004). A simple method for reconstructing a high-quality NDVI time-series data set based on the Savitzky-Golay filter. Remote Sensing of Environment, 91(3–4), 332–344. https://doi.org/10.1016/j.rse.2004.03.014

ABSTRACT
Sentinel-2 satellites offer a global coverage of Earth surface with a high spatial resolution (between 10 and 60 meters for a pixel) and an image frequency of 5 days on average. Sentinel-3 satellites offer images at a lower spatial resolution (300 meter pixels) but at a higher frequency (almost daily images). Combining Sentinel-2 and Sentinel-3 images allows us to produce a high resolution albedo product with sufficient temporal resolution to track monthly changes on agricultural fields over the year. First results calculated since the beginning of the 2020 summer are presented with comparisons with ground data gathered over a French prairie.
INTRODUCTION
The advent of satellite fleet like the satellites Sentinel-2 (A \& B) and Sentinel-3 (A \& B) fosters the environmental research to enter into a new era owing to both an appropriate revisit - typically 5 days - of the whole globe at an enhanced time and spatial high resolution (HR). High quality pre-processing of Sentinel-2 radiometry supports the dissemination of scenes finely calibrated and corrected from atmospheric effects. Though less popular than Sentinel-2, Sentinel-3 offers a frequent revisit - almost daily - but a moderate spatial resolution. The routine distribution of Sentinel products supports many applications, notably in agriculture, food security, weather forecast, climate change impact studies, water use, forest and natural resources management. A merged reflectance product in a spectral domain covering visible, near infrared and mid-infrared offers a new challenge of collecting cutting-edge information at the benefit of crop monitoring. The outcomes of the dissemination of quality-checked HR product will certainly benefit programs like GEOGLAM (Group of Earth Observation for Global Agricultural Monitoring) for which main concerns are the onset and decay of crops, and early warning. The presentation will highlight the operational methodology to be implemented in order to perform a measurement of the HR surface albedo and also ensure a trimmed monitoring of the worldwide crops.


Ruminant livestock contributes to climate change (CC) through its enteric methane emissions, which are partly offset by the use of grasslands as a carbon (C) storage factor. The ALBEDO project proposes, in addition to the well-known mitigation levers of CC, which are the reduction of greenhouse gas emissions (GHG) and carbon storage, to study a third innovative abatement: albedo. This power to reflect solar radiation by grasslands helps mitigate climate change. The primary objective of this study is to better characterize the spatio-temporal variability of grassland albedo in France. For this purpose, measurements are carried out on experimental farms and by high-resolution optical satellite, for different grassland management and pedoclimatic situations. The aim is to identify and quantify the abatement for increasing albedo in order to mitigate CC, from the plot to the territory scale, using remote sensing analyses performed by the Sentinel 2 and Sentinel-3 satellites. This project aims to strengthen the arguments in favor of sustainable grassland farming based on optimal use of grasslands.

SURFACE ALBEDO PRODUCT
The surface albedo is an Essential Climate Variable (ECV) that needs to be generated on a regular basis in order to ensure continuous estimates of the net radiation and besides the water and carbon balance. Among the key issues, are a timely production, the availability of historical archives, and the more consistency of the long-term archive. First of all, the removal of atmospheric effects must be properly handled. For Sentinel-2, cloud removal and aerosol correction rely on the MAJA method proved to be efficient for processing multi-temporal and multi-spectral data sets. For Sentinel-3, freely distributed by VITO for the present study, less effort was performed regarding the atmospheric correction, in particular the removal of aerosol effects.


Herein, the surface albedo was first calculated using Sentinel-2 data. The choices were a 60-days synthesis period and a 10-days composite period in order to ensure that sufficient observations were available to inverse a BRDF model and thereby estimate the surface albedo. Capturing surface albedo variations along the season is meaningful for most agricultural practices. Since the aim is also to make the method fast and operational, computation time is accounted for and codes were optimized to run with a single core of the CNES’ HAL supercomputer. The method – to turn operational - makes use of the well-established approach based on a semi-empirical BRDF kernel-driven. Given the scarcity of cloudless Sentinel-2 observations (on average 50% of images are cloudy in France, but that can go up to 70% in certain regions), adding Sentinel-3 was necessary to collect enough observations and get a timeliness and reliable albedo product.

In addition, Sentinel-3 has a wider range of viewing angles compared to Sentinel-2 (less than ±15° for Sentinel-2 viewing zenith angle, compared to ±55° for Sentinel-3), which allows to better constrain the inversion of a BRDF model. FIGURE BRDF_B4 and BRDF_B8A shows the BRDF simulated with the PROSAIL model for 2 Sentinel-2. It can be seen that the only availability of data for viewing zenith angles less than 15° for Sentinel-2 cannot provide a mean estimate of the reflectance and thereby albedo. This outlines clearly the added-value of Sentinel-3 and allowed for a reduction of synthesis and composit period to 30 and 5 days respectively. To prepare the BRDF model inversion, Sentinel-3 tiles at 300 meters resolution were resampled to 10 meter on the Sentinel-2 grid using a nearest-neighbor interpolation scheme.

Broadband albedo products were derived using narrow to broadband conversion coefficients based on numerical experiments using the PROSAIL radiation transfer model. BRDF coefficients can also serve to perform a normalization of the data. In order to best answer users' requirements, the surface albedo products are delivered with a quality flag and an uncertainty assessment. Also the true age of the product is indicated as being the median value of the clear sky scenes used during the composite period. The methodology is displayed in FIGURE Flowchart.

The first step of the albedo product validation use ground measurements gathered on the 6 prairies in France. These prairies are equipped with weather stations founded by the French institute for livestock (Institut de l'Elevage). On each station, standard meteorological variables (temperature, wind speed and direction, precipitation) are collected from tower fluxes also equiped of pyranometers CNR1 measuring downward and upward shortwave radiation.

The result of albedo calculation with Sentinel-2 and Sentinel-3 data over a prairie located at Pradel, France (31TFK Sentinel-2 tile) can be seen in FIGURE Albedo_Pradel. It is compared with ground measurements. The measurements shown in figure one are for solar noon. The calculated albedo is the DHR component (Directional Hemispherical reflectance) which is an integration of the bi-directional reflectance over the viewing hemisphere. Note that a BHR (Bi-Hemispherical Reflectance) product is also an output of the processing chain.

Here are a few points we can observe on this graph :
- There is a high dispersion of the albedo measurements, especially during winter. This is probably due to low and varying illumination conditions and changes of the canopy wetness.

- The albedo calculated from satellite data is lower than the ground measurements. This may be due to a low signal value coming from Sentinel-2 data.

- We can clearly observe a link between the NDVI estimates and weather conditions : there tends to be a rise of NDVI value after precipitation episodes, and a decrease in NDVI value during dry periods (like the 2020 summer). NDVI changes can also be linked to specific agricultural practices like mowing (in June 2020 and September 2021, before dry periods).

The actual albedo product will be improved in different ways:
- The filtering is today too severe for the input data, which arbitrarily consists to discard some input data and leads to having more gaps in albedo products time series

- MAJA products for Sentinel-2 include an aerosol optical tickness (AOT) from which it is possible to derive a diffuse fraction, which is not actually the case for Sentinel-3. Giving a weight between DHR and BHR to get a blue-sky albedo will lead to improvement in wintertime when the sun is low above the horizon. Besides, it should act as a noise reduction.

- Gap-filling and consistency will enhance the reliability of the surface albedo product from implementing a recurrent method previously considered in the frame of the Copernicus Global Land Service.

SUMMARY AND FUTURE PROSPECTS

This method could be applied worldwide and treat any kind of target. Herein, the focus was carried on the monitoring of agricultural lands for grazing. Combining Sentinel-2 and Sentinel-3 satellites improve the temporal shape of the surface albedo time series to capture fine agricultural practices or meteorological events on a given target.

Specific attention will be carried on the validation of the surface albedo product in considering the 6 stations measuring the surface albedo set up as part of the study over prairies dedicated to livestock. In fine, the goal is to convert quantitative estimates of the surface albedo into radiative forcing terms and then in carbon fluxes in order to demonstrate the role of prairies for a sustainable agriculture and as an abatement of global warming.

Sentinel-1 provides dense time-series regardless of cloud cover and day light illumination, enabling
nearly daily monitoring of crop development, if considering multiple orbit paths. However, the use of
dense time-series is challenged by signal discrimination associated with different factors –
environment and topography related, sensor specific factors etc. Often the signal isinfluenced by more
than one factor and signal discrimination is difficult and sometimes even impossible. In order to extract
and utilise useful for crop monitoring information, it is important to define and quantify the
contribution of these factors to the overall backscatter signal. This can be achieved by isolating one
factor from another. Authors of earlier studies have attempted a similar analysis (Gauthier et al. 1998;
Peinado 2001; Wegmüller et al. 2011), yet there are no studies that isolate these factors while
considering time‐series.
This work intends to give an overview of how the backscatter changes associated with incidence angles
(θ) and different sensor look directions (ascending/descending) along with different plant row
orientations, affect the time-series interpretation of Sentinel-1 backscatter signal of agricultural areas.
For this purpose, time-series of a radiometrically corrected gamma nought backscatter were
considered for both ascending and descending directions, as well as for two incidence angles (θ≈33° &
42°, considering multiple orbit paths and days without precipitation). Different weather conditions
(especially precipitation forms) and topography variations (as an indicator of soil properties) were also
considered, as they highly affect the backscatter signal. Soil properties, along with weather conditions,
are the most important factors determining plant growth over space and time. Spatial information on
soil variability in the field is often not available and very expensive to collect. Therefore, spatial
information on topography (DEM, 5×5 m grid cell) was used to relate to soil heterogeneity and soil
properties in the field. Based on this topographical information, backscatter averaging was performed
for homogeneous spatial units (classification by slope and aspect) to capture and isolate topographyand soil-related effects in the backscatter time series. Through spatial averaging in the defined units,
the impact of the speckle, which is a challenge when tracking a signal pixel by pixel, could be minimized.
For the analysis, three years (2017-2019) of co – and cross-polarized (VH, VV) scenes were acquired
over agricultural fields with six crop types (Winter Wheat, Summer & Winter Barley, Rapeseed and
Corn) located in Thuringia. They were pre-processed and primary statistics were calculated to extract
and explore the crop related signal.
The resultsshow that the topography affect the backscatter signal of the different crop type, i.e. affect
their growth. This is also partly due to different sun-illumination and water run-off depending on slope
and aspect (azimuth). For instance, dry slopes with sand texture or waterlogged lowlands have poor
crop development because of lack or excess of soil water availability, thus, low biomass production,
and, thus, low backscatter in comparison to favourable spots. Differences of up to 2dB are observable
for the study area considering different local incidence angles (θ) which could become greater with ∆
θ > 9°. Analysis in regards to the backscatter differences due to the plant row orientation has shown
that there is a trend that is different for the different crop types: 1) the backscatter is higher when the
microwave signal targets the rows perpendicularly than in parallel (Sugar Beet, Corn); 2) vice versa:
the backscatter is lower when the microwave signal targets the rows perpendicularly than in parallel
(Winter Wheat, Spring Barley); 3) the backscatter is higher when the microwave signal targets the rows
perpendicularly or in parallel and lower by targeting at 45° (Rapeseed, Spring Barley). Differences are
significant, from 2 to 4 dB; however, the results are inconclusive as to whether the difference is due to
different image acquisitions (ascending/descending) or to the row orientation. Effects due to different
acquisition times and geometries are present and most significant (∆≈6 dB for this study area), but
they are more pronounced for Rapeseed and less pronounced for cereals. It was also observed that
the effect is more pronounced in dry years, indicating the influence of different weather conditions in
combination with changes in a plant structure at late phenological stages.

References:
Gauthier Y, Bernier M, Fortin JP. 1998. Aspect and incidence angle sensitivity in ERS‐1 SAR data.
International Journal of Remote Sensing. 19(10):2001–2006. doi:10.1080/014311698215117.
Peinado OL. 2001. Comparison of change detection methods for the extraction of land cover
parameters. München: Utz. 209 S (Geowissenschaften). ISBN: 3‐89675‐862‐4.
Wegmüller U, Santoro M, Mattia F, Balenzano A, Satalino G, Marzahn P, Fischer G, Ludwig R, Floury N.
2011. Progress in the understanding of narrow directional microwave scattering of agricultural fields.
Remote Sensing of Environment. 115(10):2423–2433. doi:10.1016/j.rse.2011.04.026

Assessment of plant status and crop biophysical parameters is important information to correct crop management practices, especially for the application of nitrogen fertilizers. Remote and proximal sensing of the crop has been widely used in the last decades for agricultural applications, both for assessing vegetation conditions and for subsequent yield prediction. In this work, we take advantage of vegetation indices for effective full-area monitoring of spatial variability of winter wheat biophysical parameters, nitrogen status and prediction of crop yield. The aim of the study was to verify the use of proximal sensing and satellite imaging to estimate agronomically relevant biophysical parameters and nitrogen status of winter wheat crop stand and to predict differences in crop yield.

Input data were obtained from farm field trials with winter wheat realized in the period 2019 and 2021 at three localities in the Czech Republic (Zdounky, Rasovice, Risuty) with a total area of more than 150 ha. To estimate the crop parameters, a plant sampling was carried out during the vegetation in stem elongation (BBCH 32–39), heading phase (BBCH 45-55) and before harvest (BBCH 91). Plant samples were analyzed in the laboratory for estimation of nitrogen content [%], the total amount of aboveground dry biomass [t/ha] and the harvest parameters (number of spikes and TGW) and grain yield. To complete the determination of the nutritional status of plants, a nitrogen uptake (Nupt, [kg/ha]) and Nitrogen Nutrition Index (NNI) was calculated from nitrogen content and dry biomass. Spectral properties were obtained from the satellite imagery of Sentinel-2 (Level L2A) and multispectral UAV imaging (Micasense Altum, Parrot Sequoia, DJI P4M) as the set of vegetation indices (GNDVI, NDRE, NDVI, NRERI, RENDVI). Simultaneously, proximal crop sensing was realized by Fritzmeier ISARIA or AgLeader OptRx. All experimental fields were harvested with the acquiring of the crop yield maps. Spatial data were processed and analyzed using GIS tools and then statistically evaluated relationships between variables by using correlation and regression analysis.
The results showed a moderate to a strong level of correlation between selected vegetation indices and Nupt (r = 0.875 at Rasovice). Proximal sensing systems performed similar correlation results as the remote sensing methods with a tight relationship among the vegetation indices (r = 0.942 Sentinel-2 NDRE to OptRx NDRE). Despite the free availability of Sentinel-2 data for users, their main disadvantage in comparison to the proximal crop sensors is the occurrence of clouds and other atmospherical objects in the provided image.
The spatio-temporal variability of crop growth and differences of yield formation at the experimental sites were observed by the statistical evaluation of the results of crop yield mapping. The highest correlation coefficient in this study was achieved for the NDRE vegetation index (r = 0.740 and 0.741), which concluded in the high potential of early stage in-vegetation sensing to predict the main spatial differences in crop yield, for example for yield zoning. The investigation of the main driven factor for crop yield formation differs at the observed location. Results from the experimental site in Zdounky identified the only moderate influence of plant nitrogen status (Nupt) on the crop yield values, while in Rasovice it proved almost the highest level of correlation (r = 0.823). The sampling of crop yield parameters highlighted the number of ears as the main factor for yield formation at both sites. The finding of a high level of correlation between in-vegetation crop sensing and grain yield showed the possibility to identify yield spatial variability by proximal and remote sensing systems in the early stage of crop growth. This can be implemented for the development of decision support tools for yield zoning in site-specific crop management.

Estimating crop yield early in the growing season represents a valuable asset for policy makers, who want to address national food security concerns, and for farmers, who want to improve crop management decisions for their fields. Agricultural surveys are often performed to estimate crop yield using in-field measurements and sampling. Despite their reliability, these approaches are well known to be time-consuming and expensive. In addition, in case of damage to crops, e.g., due to droughts or plant diseases, results from surveys are only available after the damage occurs, preventing corrective measures to be performed during the season to avoid yield losses.

In the last decade, technological advancement in sensor payload and miniaturization has paved the way for the development of small satellite remote sensing platforms, which can deliver timely and cost-effective information about agricultural systems as never before. Amongst recently-establish commercial satellite enterprises, Hydrosat is building a constellation of 16+ satellites to provide thermal infrared (TIR) and visible/near-infrared (VNIR) imagery of the earth surface at daily frequency and sub-field-level ground sampling distance. This enhanced resolution, combined with synergistic sensor use will open new avenues for commercial sectors and researchers to retrieve plant-related information at the sub-field-level, and to improve agricultural monitoring and management to address food and water security goals.

Empirical regression methods that relate satellite-based vegetation indices (VI) and historical yield data have usually been used to estimate the end-of-season yield. These methods use statistical models and cannot predict time-dependent processes of crop growth. Besides, the relationship between VI and yield might not be adequate under extreme weather conditions. To overcome this limitation, mathematical crop simulation models provide users and researchers with the ability to estimate crop development and yield during the growing season, and to improve management interventions (e.g., irrigation and fertilizer scheduling, disease treatment, and replanting of damaged areas) aimed at improving production. Yet, despite their wide utilization, the use of these models is often limited by uncertainties in input parameters such as soil condition, sowing date, planting density and initial field conditions. Apart from application in controlled experimental field trials, many of these parameters are often unknown, and their uncertainty impacts the reliability of model predictions. High-resolution remote sensing platforms can play a key role in identifying field and crop status from estimated biophysical variables, providing data that can be assimilated into crop models to improve their overall performances.

In this study, satellite-based leaf area index (LAI) data has been assimilated into the Agricultural Production Systems sIMulator (APSIM) to re-initialize model input parameters and estimate crop yield and soil water components in both irrigated and non-irrigated fields in the state of Nebraska, US. The impact of the number of satellite images required to re-initialize input parameters on yield estimation accuracy was evaluated. Overall, results show that the estimated crop yield and water balance compared reasonably well against in situ collected measurements throughout and at the end of the season and that the number and timing of satellite observations impact the effectiveness of the model parameterization. In this case, the assimilation of satellite imagery into the model allowed us to predict end-of-season yield reliably up to three months before harvest, with a strong correlation to independently collected measurements (R2 = 0.87 and rRMSE = 7 %). The detection of intra-season biophysical crop development is particularly valuable since it may be employed to infer inefficiency problems at different stages of the season, and hence drive specific and localized management decisions for improving crop yield.

The recently launched and upcoming hyperspectral sensors, featuring contiguous visible-to-shortwave infrared spectral data, open opportunities for the accurate retrieval of crop traits at leaf and canopy levels by new-generation models. Crop trait mapping is essential in agriculture to monitor crop status such as plant nitrogen deficiency and water stress useful in precision farming.
In the framework of the ESA Copernicus Hyperspectral Imaging Mission for the Environment (CHIME) requirement consolidation studies (RCS), which are devoted to the demonstration of hyperspectral data benefits, a hybrid retrieval workflow for retrieving plant traits in agricultural crops has been developed by exploiting images provided by the PRecursore IperSpettrale della Missione Applicativa (PRISMA). The investigated traits include chlorophyll, nitrogen, water content at both leaf and canopy level.
A unique PRISMA dataset consisting in eight images was acquired on a large agricultural area of about 3850 ha located in North-East Italy (Jolanda di Savoia, 44.855061 N, 11.952233 E), between April and September 2020. This dataset was integrated with five additional PRISMA images acquired between April and June 2021. We conducted field activities to acquire spectral measurements contemporary to PRISMA overpasses for L2D product validation as well as to measure vegetation traits for the model tuning and validation. We collected ground measurements including different crop types from both winter (i.e. durum and soft wheat and barley) and summer crops (i.e. corn, sugar beet, alfalfa, soybean and rice). We measured the leaf area index by a LAI-2200 plant analyser. Furthermore, we collected leaf samples of known area in the field for the laboratory destructive measurements of leaf chlorophyll (µg cm-2), nitrogen (g cm-2) and water content (g cm-2). Leaf pigments were extracted by samples through methanol and the chlorophyll a and b concentrations were determined through spectrophotometry, reading the absorbance at 665.2 and 652.4 nm, respectively. The total N content was determined by dry combustion with an N elemental analyser. Finally, the leaf water content was measured by weighting the leaf samples when fresh and after drying them in an oven for 48 hours at 80°C.
Regarding the retrieval methodology, a hybrid approach based on PROSAIL-PRO radiative transfer model coupled with a Gaussian Process Regression algorithm was deeply investigated. We used PROSAIL-PRO to simulate a set of one thousand vegetation spectra by varying the model input parameters. Machine learning regression algorithms were used to learn the relationships between the simulated spectra and the investigated crop traits. To reduce band collinearity and spectral information redundancy the model was trained using synthetic bands obtained by a principal component analysis. Furthermore, the simulated spectra dataset was also reduced by active learning algorithms, which were used to select only the most relevant spectra from the initial spectral dataset. This allowed maximising the retrieval performance and its speed. To evaluate the retrieval performance, the model, trained on the simulated dataset, was finally applied to the real 2020 PRISMA spectra, extracted from image pixels located where ground measurements were collected. Results were very promising for all the investigated variables, with best performance against the ground data achieved for leaf nitrogen content (R2=0.865, nRMSE=7.46%); slightly worse results were achieved for leaf chlorophyll content (R2=0.671, nRMSE=11.65%) and leaf water content (R2=0.627, nRMSE=17.09%). A significantly higher accuracy was observed at the canopy level for nitrogen content (R2=0.921, nRMSE=5.51%) and chlorophyll content (R2=0.823, nRMSE=10.17%). We used the retrieval models to map the spatial and temporal variability of the vegetation traits from the PRISMA multi-temporal images. To test the robustness and exportability of the model, we finally applied the models to the PRISMA images acquired in 2021. The consistency of the maps demonstrates the usefulness of space spectroscopy for crop monitoring to support sustainable agricultural practices.

Large scale Crop type mapping at Field level is an essential data for in-season crop production monitoring, enabling to better quantify the total areas per crop at regional level but also to better monitor crop condition and increase accuracy of the yield forecast. Since the launch of the Copernicus Sentinel missions, the abundance of remote sensing (RS) data with spatial resolution lower than parcel size (for most of the European country) has changed the perspective of crop monitoring system. Operational crop type mapping has already been designed to provide in-season crop type map at regional or national level. However, most of these approaches rely on the availability of in situ data (i.e., parcel crop type identification) of the on-going season. Because of the complexity of in situ campaign, no institution has built the capacity to gather large scale in situ data during the season to perform crop type mapping. Nonetheless, in some European countries, quasi-exhaustive parcel identification (Land Parcel Identification System, LPIS, data set) and crop type declaration (Geospatial Aid applications, GSAA, data set) are made available years after the cropping season ends. The timing and the availability of such data set represents the main limitation for crop production forecasting as information needs to be available during the season. It is thus required to develop approaches that are not based on in situ data from the on-going season and to design a model that are trained on previous season and can be transferred to the on-going one. In this study, we propose to evaluate a deep-learning algorithm trained with past-seasons GSAA data and Sentinel-2 RS data to forecast in season the crop type at parcel level.
As a use case, we considered two studied countries, The Netherlands (NL) and France (FR), which release publicly their GSAA data sets (from 2009 to 2020 for NL, from 2015 to 2019 for FR) and account for a total of 630k and 6M parcels, respectively after filtering of the smallest parcels. The evaluated algorithm is a multimodal and hierarchical Long-Short-Term-Memory (LSTM), validated with last GSAA data set (2020 for NL and 2019 for FR but this might be updated if an additional season is processed) and trained with all other GSAA data set. The source of information for training are threefold:
The post-2016 RS-based smooth time series of red and NIR surface reflectances, LAI and FAPAR were derived from the in-house processing of Sentinel-2 and Landsat-8 Level-2, involving biophysical variable processor, aggregation at parcel level, time series outlier detection and Whittaker smoother. We aggregate the RS at year level using simple concatenation of bi-weekly features, considering basic statistics indicators (e.g., mean, median, min,...).
The crop rotation histories were retrieved by intersecting GSAA data sets since 2009 for NL and since 2015 for FR.
The environmental crop distribution was derived for each parcel, summing the per-crop areas in a 10 km area during the last year of the training data set, providing an a priori information to the model as it is well assumed that crop distribution is quite stable at 10 km for the two studied countries.
We benchmarked four configurations of learning models varying the inputs:
- Crop rotation data only, corresponding to initial output which can be run prior the on-going season without any knowledge of the on-going season.
- RS data only using simple concatenation or a bidirectional LSTM to aggregate the RS features at the year-level.
- RS and crop rotation data, treating the two sources as separate modalities
- RS, crop rotation and crop distribution data, as an additional modality.
Performances increase monotonically from configuration (i) to configuration (iv). Considering the most advanced configuration (iv), promising results were obtained with overall accuracy (OA) in the range of 84% for FR and 90% for NL when considering all the crop categories as defined in the GSAA. When regrouping the class to focus on major crop categories, the OA increased to up to 95%. Winter soft wheat is well classified with F1-score higher than 90%. Comparison with NUTS-2 for NL and NUTS-3 for FR were evaluated and performances were also promising. We investigated the increase of the performances throughout the season by running monthly models from March to September. While the performances keep increasing till last run, OA outreaches good accuracy as early as in July. These performances equal or exceed traditional crop type mapping approaches achieved with in situ data of the evaluated season.
We analysed the use of the probability from LSTM to monitor crop specific LAI time series at NUTS-2 and NUTS-3 level for NL and FR, respectively. By subsampling the data set maintaining only high probability parcels, we obtained aggregated time series with limited contamination of miss-classified parcels. We analysed throughout the season the best compromise between the number monitored parcels (varying the thresholds on the probabilities) and the accuracy of the NUTS-aggregated in season crop specific LAI time series.

Today, farmers increasingly rely on information to drive their decisions not only due to advances in smart and precision farming, but also due to issues related to climate changes, the greening of the agricultural sector, and transparency for customers. For this reason, the agricultural application market is booming because developers are creating the apps and services needed for farmers to compete and prosper in a global economy. The Copernicus Sentinels can help in providing a wide range of measures, data and information about farmers land and their environment that can be linked to field sensors and augment the information used within decision support systems especially for automation of tasks. However, applications developers are not the experts in Earth Observation (EO) technologies and information extraction techniques and therefore do not integrate this valuable resource into their solutions.

Through the e-Shape project (https://e-shape.eu/ ), Riscognition (https://riscognition.com/ ) and eVineyard (https://www.evineyardapp.com/) joined forces in order to bridge this gap between application developers and EO services. Applying co-development strategies, the main requirements were identified where Copernicus EO information could expand usage of decision support system for grape growers, providing EO data value to a broader range of farms. Both SMEs are working together to bring the identified services to market quickly.
eVineyard develops a set of digital tools to assist winegrowers that work in concert to provide complete information management platform for growers. This includes providing traceability of conditions, activities, assets, financial and environmental aspects. Riscognition was founded to bring EO to application developers through industry standard Application Development Interfaces (API) and take advantage of the Software-as-a-Service (SaaS) paradigm. This allows faster application development and access to a wide range of EO services reducing the barrier to integration including raster analysis, monitoring, and many others (https://riscognition.io/ ).
Our presentation will not only showcase the Copernicus Sentinel services developed through the e-Shape project but also demonstrate to application developers how easily eVineyard was able to integrate the Riscognition EO APIs into their application in order to improve their services. Furthermore, we will show how any developer can use the Riscognition EO APIs in their applications as well to augment their already developed services, or integrate with other farm solutions such as robots and AI based services as well.

Many countries in the world have developed financial structures to support their farmers and ensure successful agricultural industries. In the European Union (EU), the Common Agricultural Policy (CAP) was launched in 1962 to, among others, support farmers and improve agricultural productivity, monitor sustainable farming and resource management, safeguard farmers' living, conserve both rural areas and economies. By leaving the EU, the United Kingdom has started reforming their agricultural payment systems, previously supported by CAP. The current changes provide an opportunity for the use of accurate spatio-temporal Earth Observation (EO) solutions in agricultural monitoring. EO can support effective public agricultural payment schemes by providing actionable and trustworthy evidence to monitor compliance both at a large scale and at the field-level, increase the area of farmland inspected and reduce intervention costs.
Planet’s Fusion product offers immense value in this context. Fusion is a cloud-free and harmonized daily product generated by combining highly radiometrically calibrated pixels from Sentinel and Landsat sensors with the higher spatio-temporal resolution of PlanetScope data. With daily, gap-free images, Fusion allows to observe farmland at both Very High Resolution (VHR) and Very High Frequency (VHF). This can be extremely useful for helping find evidence to support authorities’ planned interventions pertaining farming policy compliance.
This session will present the results of a proof-of-concept (POC) study carried out to evaluate the use of Fusion for this purpose. Two specific Fusion-based solutions to support the agricultural payment system were developed within the POC: (1) Land Classification and Change Detection and (2) Farming Event Detection. The findings of this POC show the potential of using EO data, particularly Planet Fusion, for monitoring agricultural activity over large areas with temporal and spatial resolution high enough to observe field-level changes. Fusion was used to perform Land Cover/Land Use (LC/LU) classifications at high temporal intervals to observe, among other applications, if landholders or managers (primarily farmers) are using defined land parcels as expected. Furthermore, the temporal resolution of Fusion enabled the derivation of daily LC/LU maps that can be used to observe change, for instance to see when a certain crop was harvested, how long a parcel is kept bare, etc. Event detection using multispectral Fusion data also showed that discrete events, such as planting date, start, peak and end of season, can be derived in order to monitor crop behaviour and regulation compliance.

A well-known advantage of radar satellite data is its ability to penetrate through clouds. This research seeks to make the best possible use of radar satellite data in combination with multispectral satellite data in the evaluation of the Earth's surface. The aim of the research was to determine whether the radar vegetation index RVI4S1 can be considered as a possible substitute of the vegetation index NDVI or EVI for the purposes of mapping agricultural land. Vegetation indices NDVI and EVI are probably the most frequently used indices for identification and determination of vegetation cover state, not only in agriculture. However, during the growing season, increased cloud cover is a common phenomenon, which makes it impossible to accurately map the vegetation cover over time. By using radar satellite data, it is possible to obtain a periodic time series of data which is not affected by clouds or weather conditions. By calculating the radar vegetation index RVI4S1, additional data can be obtained and used in the cases when data from multispectral satellite sensors are not available. The study compared vegetation indices calculated from both multispectral (Sentinel-2) and radar (Sentinel-1) satellites. Testing took place in the CENIA, Czech Environmental Information Agency test area in the north of Czechia, where 7 data loggers (HOBO U23 Pro v2 Temperature / Relative Humidity Data Logger) are located, which measure relative humidity and surface temperature daily. In the first phase, values of vegetation indices over time were compared with each other using statistical and mathematical methods. In the second phase satellite data were compared with ground measurements data. It turned out, that there is a similar trend apparent in the development graphs of the indices, however, the results of the research are not definitive. The issue will be further investigated, and attention will be paid to testing other methods for the application of radar satellite data in agriculture, as such data could be a very valuable input.

The advent of ESA’s Sentinel missions has inspired the development of Earth Observation (EO) technologies for the management of agricultural policies. Key examples of this in the EU include the creation of remote sensing workflows for the management of the Common Agricultural Policy (CAP) at a national level. Notably, Checks by Monitoring (CbM) and the development of an Area Monitoring System (AMS).

Yet, there is still much room for improvement both on the policy, technology and data level. The temporal and spatial resolution of Sentinel-1 and 2 has its limitations. Particularly, “the problem of the small parcels” is posing a significant administrative challenge for EU Paying Agencies. Similarly, using only Sentinel data poses challenges in the identification of ineligible areas within agricultural parcels.

Planet Fusion is an Analysis Ready Data (ARD) product which is designed to be used in combination with Sentinel-2. Fusion brings together the strengths of the Sentinel-2 and Planet’s PlanetScope constellations. The Sentinel missions are a gold standard of radiometric sensor quality. PlanetScope on the other hand has a higher temporal and spatial resolution. Planet Fusion combines these characteristics using the CESTEM methodology (Houborg and McCabe, 2018). The outcome is a daily stream of cloud- and gap-free imagery. Fusion is designed for quantitative analysis and is very well suited to measure the markers of the Checks by Monitoring (CbM) process. Fusion is also harmonised to Sentinel-2 meaning that you can use the same workflow on both data sets and arrive at comparative results.

In this session, we will present early findings of Planet Fusion’s use for monitoring agricultural policies. The Slovenian Paying Agency (ARSKTRP) with the help of their service provider Sinergise, have used Fusion operationally for Checks by Monitoring (CbM) in 2021 resulting in administrative costs savings and greater insights into area based subsidy claims. Likewise, in the Spanish region of Castilla y León, the Instituto Tecnológico Agrario de Castilla y León (ItaCyl) has tested Fusion for identification of ineligible areas in agricultural parcels. ItaCyl uses a pixel-based workflow and Fusion has delivered a 1-2% improvement in the identification of ineligible areas with respect to Sentinel-2. This is significant for the Paying Agency considering the large amount of subsidies at stake.

These initial results highlight the enormous value of enhancing Sentinel data with increased temporal and spatial resolution for the remote sensing management of agricultural policies.

Agricultural intensification has been associated with simplified crop production systems. Crop rotation – growing a sequence of crops with benefits for resource use and soils – is an age-old practice to maintain or restore the quality of productive soils. To receive subsidies under the Common Agricultural Policy (CAP), farmers in the European Union (EU) need to comply with requirements specified under the Good Agricultural and Environmental Conditions (GAEC). In the latest reform, a new GAEC 7 was introduced on crop rotation in arable land. The thinking of introducing such requirement was also apparent in the previous CAP when in 2013 a crop diversification measure was introduced in relation to the so-called greening payment. Crop diversification in this definition refers to an annual spatial requirement at holding level (e.g. farms with more than 10 ha of arable land must grow at least two crops). Diversification can be beneficial for a range of purposes including reduced reliance on agro-chemicals with positive effects for biodiversity. Crop rotation – a multi-annual practice – is much more geared to improving soil quality. Using policy tools such as eco-schemes and agri-environment and climate measures in the new CAP give Member States the opportunity to promote crop rotation schemes.

Here we illustrate how geospatial data derived from Earth Observation and from farmers’ declarations can be used in the context of agri-environmental measures accounting for crop diversity and rotation. Recently, d’Andrimont, Verhegghen et al. (2021) presented a 10-m resolution European crop map made with Sentinel-1 and LUCAS Copernicus in-situ data. Taking advantage of this, we calculated the crop diversity based on Shannon’s entropy at various spatial grid and aggregation levels. While the accuracy of the EU crop map is around 80 to 85%, we illustrate how this geospatial data could provide EU-wide information on crop diversity relevant in various policy contexts, e.g. to eventually target regional interventions.

In the CAP, farmer’s applications for subsidies are done at holding level within the Integrated Administration and Control System (IACS) through the Geospatial Aid Application (GSAA). In some Member States, anonymized parcel level crop declarations from the GSAA have been made publicly available. Using 12 years of parcel level farmer’s declarations for a total of more than 750,000 parcels in the Netherlands from 2009 to 2020, we compute both spatial crop diversity and temporal diversity in crop rotation at parcel level. First, given the context of the greening measure in place since 2013 and in the light of the no-backsliding principle currently in place, we consider how trends in crop diversity at grid and national level have evolved in the Netherlands. Second, we quantify the evolution of the number of crops in any given parcel in different periods: before, during, and after the current CAP.

These examples illustrate the potential of a top-down overview based on Earth Observation data and a bottom approach based on parcel level declarations. We reflect how such geospatial datasets including data from the Farm Structure Survey are relevant in the context of the Performance Monitoring and Evaluation Framework for the new CAP. We also discuss the possibility of combining information from these sources. For this, several considerations need to be made, including a coherent hierarchical matching of crop classes across legends, aggregating across scales from pixels, to parcels, to holdings, as well as the role of temporary grassland in the rotation. This analysis is rather simplistic as it only relies on the main annual crops. Future data reporting will have to include additional information to account for catch and cover crops, intercropping, etc. Importantly, crop rotation schemes cannot be considered independent from soil and environmental conditions nor from farmers implementing more ambitious practices in the field.

The highlands of Armenia are a prime example for mountainous landscapes featuring smallholder-based, dispersed and multifunctional land use. These areas often have high nature value and provide various ecosystem services. Yet, detailed and accurate mapping efforts with satellite imagery are facing difficulties due to the spatial and spectral complexity of intercepted landscapes. This study aims at exploring the potentials and limitations of mapping smallholder farming with an increased thematic detail in a 4726 km2 study area in Armenia using Google Earth Engine (GEE). For this purpose, a coherent classification catalogue was developed featuring 11 land use and land cover (LULC) classes in total and six agricultural target classes. The proposed classification system follows a classic pixel-based approach. The input data comprises intra-annual and multi-temporal optical imagery as well as other multisource data, all available within GEE. The data was processed to construct composites with filtered image spectra and fixed temporal bins. Different input datasets were created to assess the added value for map predictions through the multisource data. The input datasets were classified with random forest and acquired training data. For classification the study area was stratified due to GEE quota limitations. The map predictions were assessed through an accuracy assessment supported by a comprehensive validation sample. Based on the results, a “good choice” dataset was constructed that achieved overall accuracies of 76% and 82% for the two sub-regions. User’s accuracies ranging from 57% to 88% indicate mapping constraints of the highly collinear agricultural target classes. Compared to a baseline dataset with optical data only, an added value of the multisource data could be observed. The improvements are however limited, suggesting other accuracy limitations such as spatial and temporal resolution. In respect to this, the mapping approach can be further improved without sacrificing its simplicity and practicality. The approach shows potential regarding the transferability but also faces limitations regarding the workload for training data. The produced map delivers accurate and detailed information of the extent and the distribution of the LULC classes. Such data is of great relevance for many current land-use challenges in Armenia.

Corn silage is widely used as food for cattle. Corn silage is sown at the end of April and the choice of the harvest date, often between September and October, is an important issue for the farmers. They must harvest on a date that maximizes the yield and nutritional qualities of the corn. The key element that determines these two variables is the dry matter content of the plants. The longer the corn is left standing, the more the total dry matter content increases. But, at the same time, a too high dry matter percentage leads to a loss of yield since part of the ration can’t be digested by the cattle. Conversely, if the maize is harvested prematurely, the high residual moisture leads to problems of preservation during the silage process. On the one hand, the quality of the grain deteriorates, leading to a decrease in yield, on the other hand, the silo can be contaminated by toxins, which constitutes a risk for the herd. Assessing the dry matter content could allow harvesting at the optimum date and to prioritize the harvesting order of the fields.
The images provided by Sentinel-1, are not affected by clouds, it covers Belgium every three days and is equipped with a C-band radar sensitive, amongst other, to the canopy water content and particularly in small canopies like corn. Sentinel-1 images are required because of the short time interval and reliability of regular and dense observations. Using this instrument, prospects for regular monitoring of dry matter at the parcel level could become possible.
To evaluate its potential for dry matter content detection, a set of field data were compared to the backscattering coefficient of Sentinel-1. Field data has been collected between August and October 2020 during the corn growing season. The dry matter content and the biomass have been collected for 18 fields in Belgium by crop cutting. Measurements have been carried out 5 times over the end of the growing cycle et synchronized with Sentinel-1 overpass.
Five variables have been evaluated : canopy water content (CWC), Biomass, dry matter content in percentage and in tons per hectare and leaf area index. The VH polarization shows the best coefficient of determination for all variables compared to VV polarization. The CWC is the best correlated variable to the signal with R2 = 0.71 and in the best conditions for the dry matter content, R2 = 0.69.

Monitoring the dynamics of biomass in the conditions of organic farming is important because it is related to forecasting yields. Monitoring the dynamics of biomass in the field is an extremely labor-intensive process that is limited in time and depends on weather conditions. This requires the use of remote sensing methods, which provide a reliable assessment of the process of accumulation of terrestrial biomass in crops. At present, the use of remote sensing methods in organic farming is not enough. The prospects for the development of organic farming in the context of the Green Deal are promising, which necessitates more and more remote sensing research in the field of organic farming and the creation of new products and services in order to commercialize them. The aim of the present study is to use remote sensing methods to track the dynamics of biomass accumulation in einkorn culture in the conditions of organic farming.
The experiment was conducted on a certified biological field located in the municipality of Parvomay, Plovdiv district, south-central Bulgaria. The field was planted with einkorn in October 2020. The boundaries of the field were determined with the help of an organic farmer in Google Earth Pro. The field of study is divided into three separate parts depending on the condition and development of the crop in early March 2021. The condition and development of the field was determined using the EOS Crop Monitoring platform, which uploaded the KMZ file with the field boundaries and generated the vegetation index NDVI based on which the field was divided into three separate parts with high NDVI values, those with medium NDVI values and low NDVI values. In the respective parts of the field, 3 GPS points were generated in the EOS Crop Monitoring platform. After that, with the help of the mobile application of the EOS Crop Monitoring platform, the corresponding GPS points were found in the field, which represent the three different NDVI values. A square polygon sided 10 m × 10 m was drawn in the corners of which permanent markers of the field were placed, marked with 12 GPS points from where the field samples are taken. On April 2, 2021, at the end of the phenological phase of tillering of crops, the first sampling from the field of fresh biomass was made. The census of einkorn plants in 1 sq. m., as well as the weeds in the respective sq. meter, the density and species composition in each of the 12 permanent plots were determined. All the biomass was collected from each of the plots, and that of the einkorn was separated from the biomass of the weeds. Immediately after collecting the biomass from the field, its weight was determined by the weight method, in grams (using a precision laboratory balance). The fresh biomass was brought to an air-dry state according to an established procedure. This type of sampling and measurements of the biomass from the biological field was performed in the following phenological phases of heading and ripening of the crop. In May and June, the Wingtra unmanned aerial vehicle (UAV) was used with a Micasense multispectral camera and a Sony RGB camera, respectively to map the field. The data from the UAV acquisition were processed with the Pix4D software. The vegetation indices EVI, MSAVI, NDVI, Chlorophyll Index Green, and Chlorophyll Index RedEdge were generated with the help of the same software. The same vegetation indices were generated for Sentinel 2 using SNAP.
It was found that there is a direct relationship between weed density, the number of einkorn plants, their biomass and the vegetation index NDVI. A strong correlation has also been established between the accumulation of terrestrial biomass and the vegetation indices NDVI, MSAVI. With the help of Chlorophyll Index Green, Chlorophyll IndexRedEdge it is possible to distinguish weeding in the tillering phase, when the biomass of weeds in the crop predominates due to the difference in the phenology of weeds and einkorn. The same statistical analyzes were performed with data from Sentinel 2, which confirms the results obtained from UAV imaging. In conclusion, it should be emphasized that the study of the dynamics of biomass in organic farming requires a priori knowledge of the phenology of weeds in order to avoid exaggeration of biomass values in the early stages of crop development. Detailed mapping of weed distribution and density is also needed in organic farming, which requires the use of a new type of hyperspectral sensors.

Earth-i is an innovative, technology-driven company specialising in the application of artificial intelligence techniques to a wide variety of geospatial data types, including very high resolution imaging and video from space, to extract unique insights of commercial value to customers. We use automated data analytics and artificial intelligence to exploit multi-sensor, multi-resolution, multi-operator satellite and complementary data.

In the period 2018 to 2021, Earth-i led the “Advanced Coffee Crop Optimisation for Rural Development” (ACCORD) project. Funded by the UK Space Agency’s International Partnerships Programme (IPP), the project provided proactive, geo-targeted crop management information to improve the livelihoods and incomes of smallholder coffee farmers in Rwanda and Kenya. Information was derived from satellite Earth observation imagery combined with highly localised and accurate weather data and pest and disease alert models. Crop management advice was delivered directly to smallholder coffee farmers via easy-to-understand text (SMS) messages.

ACCORD used a variety of satellite data from different sources to ensure the optimum resolution of data balanced with cost-effectiveness. The information provided by ACCORD helped to make significant improvements to coffee crop quality and yield, as timely guidance and advice enabled farmers to take both preventative measures and more accurate reactive measures in their agricultural management actions, whether deciding when to apply fertiliser or pesticides, or the best time to pick the cherries or prune the bushes. The project helped farmers to produce larger and better-quality coffee crops, in turn helping them secure higher prices and achieve higher incomes.

As part of our ongoing research and development in this area, Earth-i has continued to investigate the use of high resolution EO data combined with machine learning and automated processing pipelines to provide accurate and targeted information about coffee crop performance, not only for the purpose of giving farmers better tailored advice, but also to develop a commercial service for coffee purchasers, wholesalers and international traders, who are very interested in understanding crop performance in real time, and in crop yield forecasts which give them an indication of supply vs demand and the likely impact on coffee prices both nationally and internationally. Once proven for coffee, we plan to extend the techniques to address other staple crop types; doing so will open up opportunities for us to support food security programmes and the related sustainable development goals.

Our research has included the application of machine learning techniques and deep learning models to automate the detection of coffee crops over large areas (regional, state-wide and country-wide) including identification and delineation of individual fields, using the Minas Gerais state in Brazil as a case study. We have investigated use of several Sentinel-2 remote sensing indices including Leaf Area Index (LAI), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI) and Chlorophyll Vegetation Index (CVI) to detect the stage of coffee crop in its phenological cycle for each field and understand the health of the crop; and we have investigated a number of different yield prediction models to correlate between observed crop health and eventual crop yield or output.

Many challenges have been encountered in the course of this work. In countries like Brazil we are dealing with very large areas of coffee cultivation, but often split up into small fields and interspersed with other crops. Cloud cover is a persistent problem in tropical latitudes in the growing seasons, and collecting ground truth data is expensive and time consuming, whilst also being hindered by the Covid-19 pandemic. Different varieties of coffee plants have different characteristics, and the crop is often cultivated in a 2-year growth cycle, with adjacent fields not necessarily at the same point in their cycles, meaning that coffee fields even within a single satellite scene can all look quite different from each other.

Manual labelling of satellite images to compile training data sets that cover all such situations would be extremely labour intensive. For this reason, we are now exploring the development of a tool for assisted labelling, that will use both spectral properties and machine learning techniques for initial “auto-labelling” of training data. Spectral signatures will be used to rapidly mask out obvious non-crop areas from the images, and a machine learning model will then be used to roughly identify what of the remaining data is likely or not likely to be coffee, and tag it accordingly. An analyst can then quickly go through rejecting any incorrect classifications, significantly reducing the amount of manual labour required for preparing the training data sets.

In this session Earth-i will summarise the achievements of the ACCORD project and present the results of our subsequent and ongoing R&D in this domain, with a summary of where further work is needed and our proposed next steps. The results to date have been extremely encouraging and Earth-i is confident of bringing the concept to market as a commercial product in the near future, and extending it to address other crop types and other geographies thereafter.

Three of the world’s largest exporters countries of cocoa; Ivory Coast, Ghana and Cameroon; are located in West Africa. Ivory Coast and Ghana forests have rapidly disappeared in the last 20 years (Vancutsem et al 2021), cocoa production being one of the main driver of deforestation in forests and protected areas in these countries (Abu et al 2021). In contrast, cocoa plantations in Cameroon do not yet cause deforestation to the same extent. In addition, cocoa plantations are often illegal, linked to major land tenure issues and tensions at the community level. National and international actions are taken regarding the living income of producers, the issue of child labour and halting deforestation to reach a more sustainable cocoa production. Recently, the European Commission released a regulation proposal to impose deforestation-free supply chain regarding six main commodities imported in the EU, including cocoa (EC, 2021). Ivory Coast, Ghana and Cameroon are major suppliers of cocoa into the EU market.

In order to report on the environmental sustainability of the cocoa production, Earth Observation data can play an important role. However, there are still various challenges in setting up transparent monitoring systems that report on forest changes and on the impact of agricultural commodities such as cocoa on the forest cover. In addition to the technical challenges for monitoring changes in forest cover and expansion of cocoa plantations, there are other challenges for the implementation of transparent monitoring systems in these countries such as: complex institutional systems, diversity of interests from the private and public sectors and limited technical capabilities.

In this study, we reviewed the different products that are openly available to report on forest and land use changes in those three countries. While annual tree cover change products are available (Global Forest Watch, Vancutsem et al 2021,Verhegghen et al 2016), land use maps are still rare and often nationally owned. We assessed the existing private and public initiatives to map cocoa plantations and other land use classes in those countries and compared their accuracies and openness.

The combination of EO based products and geospatial data from the production farms is also an essential part of a monitoring system. However, in West Africa, those geo-localized information at the farm level are usually not publically available. Using a small set of cocoa farms polygons in Ivory Coast and Ghana, we analyse the information available from the land use maps and investigate the tree cover loss over the period 2000-2021 at the parcel level obtained from the Tropical Moist Forest product (Vancutsem et al 2021). The impact of different forest definitions is assessed.

We conclude with the perspectives for a transparent system to collect in-situ data, using the Copernicus4GEOGLAM service that will be used to survey different tree and annual crops in the North West of Ivory Coast following a pre-established spatial sampling. With such reference dataset, the accuracy assessment and area estimation aspects often missing in previous mapping effort will also be better addressed.

“Monitoring” is here understood as an in-season analysis of satellite-based time-series observations of cultivated crops, agricultural practices, land cover and land-use changes in agricultural areas. Of specific interest are the records collected during the growing season of crops and grassland. The satellite reflectance values of sun illumination and the microwave backscatter is extracted for locations of interest by applying complex archive retrieval. Processing into “Monitoring Products” is then performed, providing the time-dependent characterization of what type of plants are grown on an agricultural parcel, which yield may be expected, or which practices are applied by farmers at given moments in time (e.g. detection of mowing or harvesting events).
EO-WIDGET is being developed and recently put into pre-operations as a high-throughput satellite data processing chain for bulk generation of Monitoring Products which are demanded by Paying Agencies (PAs) in EU Member States for eligibility checking of farmer claims under the Common Agricultural Policy (CAP) regulatory framework. The Area Monitoring System (AMS) denotes the EO components interfacing the PAs’ administration and control systems utilized in this context. The requirements for Monitoring Products are territory-specific, yet thematically similar, which allows adopting a generic processing and data service platform solution equipped with local adaptation features relating to the data access and PA-facing service components, such as an individualized interactive Graphical User Interface (WebGUI)I.
PAs are looking for industrial offers which deliver AMS services fulfilling their needs.The EO-WIDGET initiative addresses this demand. It provides an industry-managed environment running on IT-cloud infrastructures into which the necessary (very large) volumes of satellite data and auxiliary data are pulled into processing workflows for Level-4 production. EO-WIDGET, at its outset, has been deploying open source Sentinel toolbox software and algorithms provided by the ESA Sen4CAP programme. It is evolving the portfolio of software following the PA's various product demands. It replaces and goes beyond the toolbox and Sen4CAP legacy functions and transitions it from a tool to a service. Furthermore, an open-source software set of graphical widgets are being made available, supporting visual analytics of the Monitoring Products within Web applications, like in the provided fully-featured, multi-language, customizable EO-WIDGET WebGUI App (a tool-set for expert judgement users).
EO-WIDGET has been proving its “as a Service” concept and country-wide scalability for the full 2020 and 2021 growing seasons. It has shown to be capable of up to bi-weekly Monitoring Product creations on the basis of the latest farmer declaration data, including a crop type map, the identification of harvest events and monitoring of grassland in the form of a mowing count. The identification of bare soil presence through the season is also offered as a service. In-depth analysis with In-situ data is currently ongoing and will be extended significantly in the 2022 season. Operational service contracts are foreseen to start with the 2023 season.
EO-WIDGET is an “industry-first” in the AMS context that offers a provider-managed processing chain and hosted workspace individually instantiated for each customer PA. It is very much alleviating the IT administration and software implementation complexity for customers and operates the AMS on behalf of a PA under a Service Level Agreement. EO-WIDGET is an industry collaboration open for additional players to join the value-chain with offers for operations platforms, specialized core algorithms, or customized delivery interfaces, ensuring a versatile response to present and future agricultural monitoring needs.

Nowadays, monitoring the crop status, growth and productivity has become crucial to the food emergency responses and planning for a sustainable development strategy.
Considering the constantly increasing population density, pressure on ecosystems, and several diseases impact on food production due to moving and export restrictions, crop productivity needs to be monitored and predicted to ensure food security and good managements. It is therefore of paramount important to develop and apply tools that enable the understanding of crop functioning and the assessment of crop yield over the large area. Furthermore, stakeholders and policymakers such as Governments, NGOs, Companies and Farmers can take several strategic advantages from knowing the crop growth trends, the spatial distribution of yields and yield forecast. In this context, the integration of time series of actual crop development observations from remote sensing and models’ simulations of crop growth processes, can be the only solution to map crop yield over large areas when access to field data is limited.
The aim of this study is to provide a robust assimilation framework based on the Ensemble Kalman filter (EnKF) algorithm that allows for combining Sentinel-2 Leaf Area Index (LAI) data with the Simple Algorithm For Yield estimation (SAFY). The EnKF sequentially intervenes on model simulations by updating the state variables based on LAI observations considering both observations and simulations errors to provide a more accurate estimation of crop yield. SAFY model is based on Monteith’s concept, and it was developed oriented to the remote sensing application. The model requires weather data for the simulation of the crop growing season; in this study, the ERA5-land dataset of the Copernicus Climate Change Service was used. This weather reanalysis dataset provides regularly gridded hourly data with 0.1° spatial resolution.
A sensitivity analysis was carried out to properly select the models’ parameters exerting the most impact on the model simulations. Further, the raster-based method assimilation method was applied to two years data on Winter Wheat and the first field-scale evaluations of the method provided promising results. The assimilation framework was tested against data collected in the fields where high yield variability was induced with different fertilisation rates (using three replicates of four different Nitrogen fertilisation rates comprising 1 ha trial each). The mapping of regional scale yields, together with further optimizations of the computational time of the data assimilation framework, have to be considered and are still being improved.

The Space Climate Observatory (SCO), created in 2019, addresses the need to step up international coordination to enable accurate assessment and monitoring of the consequences of climate change based on space and in situ observations as well as numerical models. It aims at becoming an important tool for decision-making on preparedness, adaptation and resilience to the impacts of climate change at the local and regional levels. The participants of the SCO International can be seen as a formal group of space agencies and international organizations gathering regularly to share experiences, toolkits and methods, to provide projects accelerators, and to discuss and agree projects and actions addressing the identified common goals. These goals are directly linked to international commitments taken by nations to tackle climate change issues and their impacts.

VietSCO, the Vietnam Space Climate Observatory project, began in July 2020 and will run through April 2022. It is a collaboration founded between the French and Vietnamese Space agencies (CNES and VNSC), the key Vietnamese stakeholders (Ministry of Agriculture - MARD, Ministry of Natural Resources - MONRE, National Disaster Management Agency - VNDMA), UNDP, the French research lab Cesbio, and private sector actors GlobEO and Athena Global. VietSCO has two central components:
• Monitoring rice production areas affected by climate change in the Mekong Delta (VIMESCO Rice);
• Monitoring impact of typhoons on agricultural areas in Vietnam in support of more resilient agricultural planning (Viet-ARRO).

The VIMESCO-Rice component enables regular monitoring of rice fields via Earth Observation (EO) satellites. Regular radar remote sensing data interpreted using dedicated methods and algorithms will generate mapping information that supplements in situ data and statistics (area planted, state of development, cropping calendar, productivity indicators, number of annual harvests, salinity intrusion, year over year changes). The main envisaged users for VIMESCO-Rice are MARD CIS and MONRE, although provincial authorities are also interested.

The Viet-ARRO (Agricultural Resilient Recovery Observatory) component of the project is setting up a full-scale demonstrator of a “resilient recovery observatory” largely produced through satellite observations, which can be triggered by the government (local or national) and international development aid organizations as soon as a major disaster occurs. The main user for Viet-ARRO is VNDMA.

The use of Earth Observatory/EO data and derived products allows users to understand changes in rice production areas and impacts of extreme weather events over time and contributes to resilient reconstruction and recovery. After extreme hydrometeorological events affecting agricultural production (mainly rice cultivation), the Viet-ARRO Demonstrator will provide key stakeholders at the local and national level with regularly updated geospatial information, derived from satellite EO.

The Viet-ARRO will provide VNDMA with free and open access to data and information useful for assessing impact, planning, and monitoring recovery, and will serve as a forum for exchange and collaboration on recovery related issues to foster resilience at the community level.

As the first phase of VietSCO draws to a close, there strong uptake and interest from national actors in Vietnam, and a sense that local actors also have much to gain from a satellite-based approach to climate change adaptation. A second phase is currently in planning, with a view to extending the reach and scope of VietSCO to forests and coastal monitoring.

The sustainable development goals (SDGs) developed by the UN include the number 2 zero hunger, in which one of the objectives is to improve the productivity of small-scale farmers. In Sub-Sahara Africa (SSA) there are 37 million ha of maize produced annually, where 95% of farms belong to smallholders with less than 2 ha. These farmers were heavily affected by the Fall Armyworm (FAW) at the start of 2017 and are still losing tons of maize annually. FAW (Spodoptera frugiperda) is a voracious and mobile pest that has been noted to be a robustly polyphagous lepidopteran, though exhibiting a strong preference for maize in early vegetative stages when it will devour tender green leaves. Nevertheless, at the reproductive phase of maize, FAW will cease to eat the green vegetation and move to the cob where it can completely ruin crop yields. Thereby, this pest affects the leaf area and aboveground biomass during the maize vegetative growth stage; the implementation of remote sensing could be a good opportunity for monitoring FAW. The study was in three different countries: Zimbabwe, Tanzania, and Kenya, and visited 50 maize fields from small farmers.
We conducted observations Normalized Difference Vegetation Index (NDVI), Greener Area (GA), and Leaf Area Index (LAI) at four different scales from continental to field levels: 1. Using Sentinel 2 a+b and Planet Scope (satellites and micro-satellites), we calculated NDVI. 2. With a drone phantom 4 with a RGB camera at 50 m of altitude, we calculated GA. 3. With the multispectral camera at 5 m of altitude was calculated the NDVI for each field and with an RGB camera at 5 m of altitude, we calculated GA for each field. 4. Lastly, with the fisheye lens adapter on a mobile phone, we calculate the LAI for each field from below the maize plant. We did correlations between each index taken all the same day and calculated the coefficients of determination. The correlation between the NDVI from the Planet Scope (3 m/pixel) against the multispectral camera at 5m (0.06 m/pixel) was R2 = 0.713. Following the correlation between GA taken from the RGB camera at 5m (0.009m/pixel) against NDVI (Planet Scope) and NDVI (at 5m) were R2 = 0.936 and 0.70, respectively. Finally, the correlation between the LAI taken with a fisheye from below (0.001 m/pixel) against NDVI (Planet Scope), NDVI (at 5m) and GA (at 5m) were R2 = 0.737, 0.617, and 0.684, respectively. We did not compare the NDVI from Sentinel 2 and the GA from the drone because these were not taken the same day.
Regarding the data from two Satellites: The Sentinel 2 (Copernicus, ESA) and the Planet Scope (commercial micro-satellite), we did the time series anomaly change detection and first derivative growth pattern analyses on maize using NDVI. Then, we observed if the anomalies occurring during the vegetative growth stage were from the presence of FAW, which means it will result in a reduction of the LAI or total green biomass (NDVI) of the crop. Furthermore, it was hypothesized that the NDVI time series first derivative would normalize the data by removing the background noise caused by intercropping, weeds, and other abiotic factors. We could not analyze Tanzania, due to cloudy conditions prevailing during the entire maize season. Some graphical results confirm that these anomalies can be contributed to FAW, as the day that the second derivative vegetation curve presented the anomalies were when we were in the field taking the measurements with the other sensors.
One of the possible suggestions for future continuations of this work could be integrating at different scales of remote sensing; for example, companies or governments could conduct regional monitoring of the small farmers’ fields with Sentinel 2 a+b and commercial micro-satellite constellation with a more detailed resolution and more frequent return interval at every five/one day/s (depending on the type of satellite); and with the data from the satellites calculate the NDVI and derivative curves. Then, if the analysis shows that the NDVI is down in the middle of the season, this would be a call to action to check the field using an app or other visual assessment to confirm if the cause is FAW and verify that the dip in the NDVI growth curve was not caused by intercropping, weeds, other abiotic and biotic factors.

Agricultural carbon budget monitoring is a key task in a world undertaking global climate change. This is in this context that our team is part of the New IACS Vision in Action (NIVA) project that is funded by the European Commission [1]. The project aims to modernise Integrated Administration and Control System (IACS) by making efficient use of the wide range of data relevant for agriculture while reducing administrative burden for farmers, paying agencies and other stakeholders.
In the User Case 1b (UC1b) of this project, our team aims to develop several agro-environmental indicators prototypes that could be relevant for the future Common Agricultural Policy (CAP). The development of these indicators focuses on scientific methodologies, remote sensing and IACS data, algorithms, the software designs and the production of maps. Three environmental themes are covered in this project: first, an indicator that assesses the annual CO2 flux of agricultural parcels. Second, an indicator that estimates the risk of nitrate leaching of agricultural parcels. Third, an indicator that gauges the biodiversity potential of agricultural landscapes.
Each of those three indicators may have several levels of calculation complexity, expressed in different “Tiers”, that have increasing levels of complexity and precision. In this work, we focus on the Tier 1 of the annual CO2 flux indicator (CT1) [4].
The CT1 indicator is based on the following approximation [2,3]: for a whole agricultural year, the annual CO2 flux of an agricultural parcel is proportional to the number of days it is covered by an active photosynthetic vegetation. With an assumed level of precision, this proportionality relationship is independent of the crop species and variety (from a list of about 15 selected different kind of crops) that grows on the given parcel. The key point of the methodology is that the vegetation cover can be estimated thanks to remote sensing imaging.
In this context, the Sentinel-2 mission that provides systematic global acquisitions of high-resolution multi-spectral data with a high revisit frequency, was used to calculate NDVI for each date of acquisition and each pixel after filtering clouds and cloud shadows by applying the MAJA processing chain. Then the NDVI time series were interpolated on a daily basis in order to calculate the duration of active vegetation coverage at 10m resolution for croplands over several case studies (in France, Netherlands, Denmark and Spain) by considering a NDVI threshold of 0.3 below which it was considered that there was no active vegetation. Then by applying the linear relationship described above the net annual CO2 flux was be estimated over those areas of study. For each test area, statistical analysis per crop specie showed results that were coherent with the literature.
Recently, thanks to the technical support of the French Space Agency (CNES), the indicator has been successfully produced on the entire French metropolitan territory over about 9 millions of parcels (cf. figure) by using the IOTA2 processing chain [5,6]. Preliminary results shows interesting and coherent patterns in net annual CO2 fluxes such as :
- contrasts in areas dominated by winter or summer crops (the later being dominant in the very South west). As winter crops have a longer vegetation cycle than summer crops, the fix annually more CO2.
- regional differences due to the difference in implementation of cover crops in the crop rotations (e.g ; mandatory in Bretagne but less present in south east France), Implementation of cover crops during fallow increase the duration of soil coverage and allow fixing more CO2.
- effects of pedoclimatic conditions that affect the crop phenology.
Of course the relative simplicity of this approach raises several questions, such as the impact of a fixed threshold value for detecting soil coverage or the accuracy in clouds detection on the calculation of the duration of soil coverage. We are currently working on those issues in order to produce a good measure of uncertainties.. Yet, the simplicity of this approach allows its deployment at large scale in order to produce CT1 indicators at national scale for the CAP.



[1] https://www.niva4cap.eu/ (H2020 project)
[2] Ceschia et. al. Agriculture Ecosystems & Environment 139(3):363-383 (2010)
[3] Smith et.al, AGEE3621 (2010)
[4] https://gitlab.com/nivaeu/uc1b_indicators_tool
[5] Jordi Inglada, Arthur Vincent, Marcela Arias & Benjamin Tardy, iota2-a25386, Zenodo (2016)
[6] https://doi.org/10.5281/zenodo.58150

Crop inventory is a foundational GEOGLAM Essential Agricultural Variable for food security programs and growth in open access and operational moderate resolution. Satellite observations provide mechanisms to monitor extent and within season production over large regions operationally. This can help optimise resources and implement risk management tools. This research application illustrates a training pipeline for a Deep Learning model, capable of generating crop type predictions across production hot spots in Canada. A hierarchical model design is able to use multi-source training data to classify the crop types into 6 categories, namely Canola / Rapeseed, Cereals, Corn, Pulses, Soybeans and Other prior to harvest. Multi-source Sentinel 1 and Sentinel 2 satellite imagery, throughout the growing season (here dating from April 1st to August 19th), were fused to create time-series data which resulted in a unique time-signature for millions of fields across agricultural regions of Canada for the standard growing seasons 2017 to 2021. Specifically, to create the dataset that was used to train the Crop Type Classification model, after acquiring the Sentinel-1 and Sentinel-2 images for the areas of interest, a Field Boundary Detection algorithm was employed to delineate the field boundaries in these areas. Then, for each field in these areas, zonal statistics of every Sentinel-1 and Sentinel-2 bands were extracted and combined into a feature vector. By concatenating the feature vectors of multiple dates during the growing season, feature vectors were created representing a temporal signature for each field. These time-signatures were then provided as input to a 1-Dimensional Deep Convolutional Neural Network which was trained to predict the crop type of a field given the corresponding signature. We use a combination of publically available data and 40769 windshield observations to train and evaluate the models. Regional fine-tuning of the model was employed, to improve its performance in specific provinces within Canada. Currently, the Crop Type Classification model for Canada is able to predict the crop type of a given field with 94% and 91% overall accuracy for the years 2020 and 2021, respectively, as measured on a held out test set of our in-house ground truth dataset. Overall accuracy refers to the accuracy of the model averaged over the different crop types. F1 scores for each individual crop type are presented in Table 1. In the future, we plan to retrain the model using more public domain data, as well as increasing the size of the ground truth dataset, to allow predictions to be made in different geographical regions and improve the overall model architecture.

Comparing uncertainties provided by machine learning methods for the retrieval of vegetation variables: applications for Sentinel-2

José Luis García Soria, Ana Belén Pascual, Adrián Pérez-Suay, José Alberto Estévez, Juan Pablo Rivera Caicedo, Katja Berger, Jochem Verrelst

Optical Earth observation missions acquire spectral information over the land surface. This spatiotemporal data stream can then be processed by machine learning algorithms (MLA) that convert the reflectance data into biochemical and structural vegetation variables. However, the majority of commonly used MLAs do not provide a standardized uncertainty interval along with the retrieved retrievals. The absence of a confidence interval in the retrieval makes these statistical models less reliable, as they lack fidelity of the predictions in space and time. In this scenario, the MLA Gaussian Processes Regression (GPR) is found as an outstanding algorithm given that it provides an estimation of the uncertainty of each prediction (Verrelst et al., 2013). However, it is crucial that also other MLAs are equipped with uncertainty intervals in case GPR does not perform as well as required and alternative algorithms may become necessary.
The main objective of this work was to compare provided uncertainties of GPR models against those of Random Forest Regression (RFR) given the retrieval of several biochemical and structural variables, such as fractional vegetation cover (FVC), canopy chlorophyll content (CCC), canopy water content (CWC), and leaf area index (LAI). To achieve this, we employed the Automated Radiative Transfer Models Operator (ARTMO) scientific software framework. ARTMO allows the user to operate several leaf and canopy radiative transfer models (RTMs). The generated training data sets can then be further processed through several retrieval toolboxes, e.g. through the machine learning regression algorithm (MLRA) toolbox, which allows the user to train regression MLAs and to establish retrieval models for the chosen variables. In this work, we implemented a quantile estimation of the uncertainty of the RFR (Meinshausen, 2006) in the MLRA toolbox. Then we compared the performance of the provided uncertainties for GPR and RFR output maps of the vegetation products. To do so, by making use of coupled Prospect-4 and SAIL (PROSAIL) simulations, a training data set with band settings of Sentinel-2 was generated to train and validate the GPR and RFR models of the targeted variables. A sub-objective of this work involved the implementation and comparison of uncertainty estimation as provided by Partial Least Square Regression (PLSR). To achieve this a bootstrapping methodology was developed. (Zhihui Wang et al., 2019).
First results indicated that GPR outperformed RFR for all the developed and validated variable-specific models, showing normalized root mean squared error (NRMSE) going from 0.7 to 1.6 percentile points (p.p.) smaller and a coefficient of determination (R2) going from 1.6 to 6.6 p.p. greater in the distinct studied variables. Though the training times of the GPR are four times greater than for RFR, processing time for retrieving the variables is shorter for GPR models (by a factor of 6). Moreover, we found that in general produced RFR uncertainties responded larger than those produced by GPR. This initial study led to a general impression that GPR is preferred; both for variable retrieval, as well for associated uncertainties. Results are to be confirmed with provided maps.
This research line will be continued by analyzing GPR and RFR uncertainties in multiple scenarios (e.g. for hyperspectral imagery). In addition, alternative MLA uncertainties will be tested, such as those provided by PLSR. Although more research has to be done in this field, this work opens opportunities to develop a new generation of retrieval models that encompass uncertainty estimates.

References:
Meinshausen, N. “Quantile Regression Forests.” Journal of Machine Learning Research, Vol. 7, 2006, pp. 983–999.
Verrelst, J.; Rivera Caicedo, J.P.; Moreno, J.; Camps-Valls, G.; “Gaussian processes uncertainty estimates in experimental Sentinel-2 LAI and leaf chlorophyll content retrieval.” Journal of Photogrammetry and Remote Sensing, Volume 86, December 2013, Pages 157-167.
Zhihui Wang; Philip A. Townsend; Anna K. Schweiger; John J. Couture; Aditya Singh; Sarah E. Hobbie; Jeannine Cavender-Bares; “Mapping foliar functional traits and their uncertainties across three years in a grassland experiment”, Remote Sensing of Environment, Volume 221, 2019, Pages 405-416.

New missions are more demanding in terms of high spatio-temporal resolution requirements, in special those related to agriculture. For agriculture and food security applications, a high revisit time is needed to provide data in a timely manner, which, in turn, requires an instrument with a wide swath. At the same time, very high spatial resolution is also necessary to provide observations at the scale of crop fields.

In order to provide accurate data to the scientific community, the emphasis is on state-of-the-art instrument concepts. In particular, for the same swath size, a whiskbroom instrument typically has a greater resolution when compared with a push broom instrument. This derives directly from the more complex design of a whiskbroom instrument. While a push broom instrument is characterised by the collection of a full line of pixels per measurement in the across-track (ACT) direction, a whiskbroom instrument uses a rotating mirror to perform multiple acquisitions along the ACT direction. In this sweeping movement, the same location on Earth can be acquired by different pixels on the detector at different times. This can be accomplished by a judicious selection of instrument parameters. The accumulated data of these pixels can be combined to obtain a single measurement of that location with an increased signal-to-noise ratio (SNR). It is of upmost importance that the centre of all these pixels is collocated at the same location so that, when all combined, the Time Delay Integration (TDI) de-synchronization is minimal.

This analysis considers a whiskbroom instrument focusing on the parameters that need to be optimised so that the co-registration of the centre of pixels is maximised. The instrument design parameters are the satellite attitude model - Yaw Steering Mode (YSM) or Local Normal Pointing (LNP) -, the relative orientation between the satellite platform and the instrument, the sampling rate of the acquisitions, the angular velocity of the field-of-view (FOV), and the detector positioning angle around the optical axis (also called detector clocking angle in the following paragraphs).

Two different study approaches are considered: (i) using as input the line(s)-of-sight (LOS) modelled by a fixed pixel FOV; (ii) having as input directly the LOS of different pixels in the detector. In order to perform these calculations, the Earth Observation CFI software is used. This software is a collection of multiplatform precompiled C libraries for timing, coordinate conversions, orbit propagation, satellite pointing calculations, and target visibility calculations, specifically parametrized and configured for EO satellites.

The obtained results showed the major impact of the detector clocking on the pixel overlap. In the analysed test case, the introduction of detector clocking improves the spatial co-registration of the pixel centres from 19 m to less than a meter for two measurements separated by twelve frame periods (less than 10 cm between consecutive pixels). Regarding the pointing model, we could conclude that the use of YSM does not affect significantly the overlap of the pixels, with a distance between the centre of two overlapping pixels of around 1.6 m for LNP, while this distance decreases to 0.9 m (only by 0.7 m) when YSM is used. On the other hand, the sampling rate of the acquisitions and the angular velocity of the mirror have a significant impact on the overlap and need to be fine-tuned concurrently. The results of varying the relative orientation between the satellite platform and the instrument depend on the detector clocking angle. Towards the edge of the swath, a lower angle is more optimal, while towards nadir its increase is beneficial. Currently under study is how to further maximise the overlap between pixels, in order to enhance the correlation accuracy.

The unprecedented availability of optical satellite data opens the possibility to develop new algorithms and applications to address important challenges such as food security, irrigation management and monitoring vegetation dynamics. In particular, the data provided by Sentinel-2 (S2) constellation is very convenient for agricultural applications, providing relatively short revisit time, spatial resolution up to 10m and adequate spectral resolution with covering the visible, the near-infrared and the shortwave infrared domains. To take advantage of this vast amount of S2 data, cloud computing platforms, such as Google Earth Engine (GEE), open new possibilities to develop crop trait retrieval models applicable to any corner of the world. GEE includes the S2 MSI collection and it also provides a powerful computational capability for data processing from local to planetary scale.

All the retrieval methods have in common that they rely on a model that converts a measured spectral quantity (usually reflectance) into vegetation traits, which describe status and vitality of vegetation. Hybrid retrieval models are of interest to run on these platforms as they combine the advantages of physically-based radiative transfer models (RTM) with the flexibility of machine learning regression algorithms (MLRAs). Among the MLRAs the Gaussian Processes regression (GPR) models excel with efficiency in the training process and fast and excellent retrieval performance.

Despite their diversity, the large majority of developed retrieval methods exploit bottom-of- atmosphere (BOA) reflectance, i.e., after an atmospheric correction algorithm has been applied to acquired top-of-atmosphere (TOA) radiance or reflectance. To avoid the uncertainty that the atmospheric correction step may introduce an alternative approach is to upscale training data simulations from canopy to atmosphere levels and derive the vegetation variables directly from TOA data.

In the present study, we implemented hybrid models directly in GEE for processing S2 Level-1C (L1C) top-of-atmosphere (TOA) reflectance data into crop traits. To achieve this, a training dataset was generated using the leaf-canopy RTM PROSAIL in combination with the atmospheric model 6SV. GPR retrieval models were then established for 5 essential crop traits namely: fractional vegetation cover, leaf area index, and upscaled leaf variables canopy chlorophyll content, canopy water content and canopy dry matter content.

An important prerequisite for implementation into GEE is that the models are sufficiently light in order to facilitate efficient and fast processing. Successful reduction of the training dataset by 78% was achieved using the active learning technique Euclidean distance-based diversity (EBD). With the EBD-GPR model, highly accurate validation results of LAI and upscaled leaf variables were obtained against in situ field data from the validation study site Munich-North-Isar (MNI), with normalized root mean square errors (NRMSE) from 6% to 13%. Model performance showed moderate to good performance on an independent validation dataset of similar crop types.

Obtained maps over the MNI site were further compared against S2 Level-2 Prototype Processor (SL2P) vegetation estimates generated from the ESA Sentinels' Application Platform (SNAP), proving high consistency of both retrievals. Finally, thanks to the seamless GEE processing capability, the TOA-based mapping was applied over the entirety of Germany at 20m spatial resolution including information about prediction uncertainty. The obtained maps provide confidence in the developed EBD-GPR retrieval models for integration into the GEE framework and national scale mapping from S2 L1C imagery. In summary, the proposed retrieval workflow demonstrates the possibility of routine processing S2 TOA data into crop trait maps at any place on Earth as required for operational agricultural applications.

Estévez, J., Vicent, J., Rivera-Caicedo, J. P., Morcillo-Pallarés, P., Vuolo, F., Sabater, N.,Camps-Valls, G., Moreno, J., and Verrelst, J. (2020). Gaussian processes retrieval of LAI from Sentinel-2 top-of-atmosphere radiance data. ISPRS Journal of Photogrammetry and RemoteSensing, 167(March), 289–304.

Estévez, J., Berger, K., Vicent, J., Rivera-Caicedo, J. P., Wocher, M., Verrelst, J. (2021).Top-of-atmosphere retrieval of multiple crop traits using variational heteroscedastic gaussian processes within a hybrid workflow. Remote Sensing, 13(8), 1–26.

Pipia, L.; Amin, E.; Belda, S.; Salinero-Delgado, M.; Verrelst, J. Green LAI Mapping and Cloud Gap-Filling Using Gaussian Process Regression in Google Earth Engine. Remote Sens. 2021, 13, 403.

The launch of sun-synchronized satellites such as ESA’s Sentinel-1 mission, the Radarsat Constellation Mission (RCM), and future missions such as NiSAR and ROSE-L improve the potential of near real-time agricultural monitoring. Furthermore, new SAR systems in Low Earth Orbit (LEO) such as those from Iceye and CapellaSpace could provide new opportunities for sub-daily monitoring of soil and vegetation. However, the presence of surface canopy water (SCW), dew or interception, during single acquisitions can affect the relationship between radar observables and crop biophysical variables and also can affect the attenuation of the microwave signal through the vegetation layer.

The aim of this study was to quantify the influence of surface canopy water (SCW) on radar observables and on the relationship between L-band radar backscatter and biophysical variables of interest in agricultural monitoring. In addition, the effect of SCW on vegetation optical depth (VOD) estimation and on the linear relationship between VOD and vegetation water content (VWC) was analyzed.

In order to conduct this analysis, an intensive fieldwork campaign was performed in Florida, USA, during a full growing season of corn. Fully polarimetric L-band data were collected 32 times per day using a truck-mounted scatterometer. To capture vegetation water dynamics and dry biomass, pre-dawn destructive sampling was conducted three times a week, and plant geometry was measured once a week for a full growing season. Three leaf wetness sensors, installed on different heights, were used for continuous monitoring of SCW. Soil moisture, meteorological data, and SCW were measured every 15 minutes for the entire growing season.

Results show that the presence of SCW can result in an increase in backscatter of up to 2-3 dB, and also affect the relationship between radar observables and crop biophysical variables. In corn, the spearman rank correlation between backscatter and biophysical variables is, on average, about 0.2 higher for dry vegetation compared to wet vegetation. The results presented here underscore the importance of considering the effect of SCW on the retrieval of biophysical variables in agricultural monitoring. In particular, they highlight the possible influence of overpass time on the interpretation of radar data for vegetation monitoring.
The estimated VOD from vegetation with SCW were generally higher than those estimated from vegetation without SCW. Therefore the surface canopy water considerably affected the regression coefficient values (b-factor) of the linear relationship between VOD and VWC from dry and wet vegetation. This finding proposed that considering similar b-factor for the dry and the wet vegetation will introduce error in soil moisture retrieval and highlight the importance of considering canopy wetness condition when using tau omega model to estimate VOD from VWC.

Remote monitoring and verification of climate positive practices requires knowing the location and boundaries of farming fields. On a small scale, these can be provided by farmers or can be accessed through various data sources such as common land units (CLUs). However, field boundary delineations frequently change, even over the course of a single growing season, with farmers opting to purchase or expand into neighbouring land, or subdividing their fields into sub-fields. Natural delineations between fields are also variable between geographic regions. For instance, fields in the UK are often geometrically complex with more permanent and natural boundaries such as hedgerows, whilst fields in the USA are geometrically simpler but have more frequently evolving boundaries in response to weather conditions and planting requirements. Existing databases of field boundaries are therefore never up-to-date, they do not account for sub-fields, and are severely limited in their availability. To address this problem and acquire up-to-date field boundaries for downstream monitoring and verification of climate positive practices, we use a state-of-the-art FracTAL ResUNet to detect and segment fields in satellite imagery [1,2]. To ensure global utility of the model, we use Sentinel-2 (S2) data which has global coverage, a temporal revisit window of five days, and a geospatial resolution of 10 metres. Field boundaries were hand labelled and converted into a three channel ground truth input for the model, containing the field extent, field boundary and boundary distance map at the field level. For the input image we only utilise the R-G-B-NIR bands. Using this approach, a Central Europe model was trained from data across France, Netherlands, Denmark, Germany and the UK, achieving an F1 score of 0.933 and a MCC score of 0.842. A model was also trained in Canada, achieving an F1 score of 0.947 and an MCC of 0.840. In extension to these models, we compensate for two additional challenges. First, different stages of crop growth can impact the differentiation between agricultural and non-agricultural land. Second, cloud cover in S2 data can occlude large volumes of image data and can create false boundaries. We compensate for these challenges by predicting over multiple timepoints across the growing season and average the predicted extent and boundary layers before binarising with an optimised threshold value. The resulting binary layer is vectorised and can be used for downstream monitoring and verification applications.

[1] Diakogiannis, F., Waldner, F., Caccetta, P. and Wu, C., 2020. ResUNet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing, 162, pp.94-114.
[2] Waldner, F. and Diakogiannis, F., 2020. Deep learning on edge: Extracting field boundaries from satellite images with a convolutional neural network. Remote Sensing of Environment, 245, p.111741.

Phenological parameters, curated from satellite data, provide a powerful tool for documenting large-scale, seasonal cycles and impacts of climate change on ecosystems. This is of special interest in mountainous alpine areas, such as the European Alps, which are particularly vulnerable to climate change and expect accelerated changes in the future. As shown by e.g. Xie et al. (2021) for 1991 until 2014, start-of-season (SOS) trends occur earlier in the Alps with distinctive regional and elevational variation. Our understanding of inter-annual variations and long-term trends of phenological parameters such as SOS, can be fostered by analyzing multiple domains of earth-based parameters.

This contribution presents the potential of AVHRR-retrieved Essential Climate Variables (ECVs; e.g. GCOS-200, 2016) time series for cross-domain analysis based on a semantic Earth-Observation (EO) data cube. Satellite-based ECV datasets have strengthened the observational component towards better understanding and prediction of their critical role in the climate system for a specific domain: for example, vegetation dynamics of the biosphere and snow cover extent for the cryosphere. One of the key sensors for continental to global ECV datasets spanning ~40 years, is the heritage Advanced Very High Resolution Radiometer (AVHRR). The ongoing SemantiX research project (https://www.semantixcube.net) has made selected and consistently pre-processed AVHRR ECV time series (1981-2020) of vegetation dynamics, snow cover extent (SCE), and lake surface water temperature (LSWT), processed at the University of Bern accessible in a semantic EO data cube, which cover the entire Alps. The concept of a semantic EO data cube, newly coined by researchers at the University of Salzburg, refers to a spatio-temporal EO data cube, where for each observation at least one nominal (i.e. categorical) interpretation is available and can be queried in the same instance (Augustin et al., 2019).

The foundational implementation step of this data cube offers numerous possibilities, such as cross-domain analysis of these ECVs based on the long EO time series on-the-fly using semantic models/queries in a Web-based architecture. By looking at specific climatological regions of the Alps, we analyse SOS, end-of-season (EOS), and the active vegetation period (EOS – SOS) based on the NDVI in relation to SCE for specific alpine catchments. Phenological parameters are analysed for the main alpine vegetation zones e.g., broad-leaved forest, needle tree forest, grassland or sparse vegetation based on the ESA CCI Land Cover Map 2000 (v2.0.7). In terms of SCE, we focus on the duration calculated from fractional snow cover extent, which impacts the phenological parameters explaining spatiotemporal variations from 1981 until 2020. In addition, we show trend anomalies of NDVI and SCE for different seasons as well as the time periods of SOS and EOS over the entire time range to analyse long-term trends in these highly sensitive ecosystems. To summarize, focusing on phenological parameters, we aim to assess intertwined ecosystem changes using AVHRR time series of multiple ECV domains stored and made accessible with the new semantic EO data cube technologies. This leverages climate-relevant spatial-temporal analysis to a next level to fill knowledge gaps and serve the needs of the climate community.


Augustin, H., Sudmanns, M., Tiede, D., Lang, S., & Baraldi, A. (2019). Semantic Earth Observation Data Cubes. Data, 4(3), 102. https://doi.org/10.3390/data4030102

GCOS-200, 2016: The Global Observing System for Climate: Implementation Needs. GCOS-200 (GCOS-214). World Meteorological Organization (Lead by Alan Belward, JRC, Italy), p. 316.

Khlopenkov, K., Trishchenko, A. P. and Luo, Yi (2010). Achieving Subpixel Georeferencing Accuracy in the Canadian AVHRR Processing System. In: IEEE Transactions on Geoscience and Remote Sensing 48.4, 2150–2161, https://doi.org/10.1109/TGRS.2009.2034974

Xie, J., Hüsler, F., de Jong, R., Chimani, B., Asam, S., Sun, Y., et al. (2021). Spring temperature and snow cover climatology drive the advanced springtime phenology (1991–2014) in the European Alps. Journal of Geophysical Research: Biogeosciences, 126, e2020JG006150. https://doi.org/10.1029/2020JG006150

This study investigates the phenological cycle of Paphos forest in Cyprus using SAR data from 1995 till 2020. Observing phenological changes are important for understanding forest ability to regenerate themselves since as we have seen from the literature timing of blossoming changes are influenced by climate change, while coniferous forest depends on seeds to reproduce themselves. Monitoring Mediterranean forests are also extremely important since they are expected to face first the most adverse consequences of climate crises.

As we observe average phenology from each three-satellite mission (ERS-2, ENVISAT and Sentinel-1), there are two main blooming peaks each year, one in January and one in July, while NDVI only provides information from the first peak only. We suspect that the 1st peak relates to increased foliage, while the 2nd one to fruition (cones) since it occurs two months after blossoming timing and SAR data penetrate the forest canopy and collect information from both structure and moisture (e.g., water content of the cones).

The increased frequency of Sentinel-1 data allows us to observe the phenological cycle in more details, so from Nov 2014 till-Oct 2020, we measure the start and end of each peak. The first phenological cycle recorded, November 2014 – October 2015, contained only one peak, while there were small peaks between Jan 2018 and July 2018 resulting into a month delay of the 2nd main peak that appeared in August 2018. To better understand those alternations of the phenological cycle, meteorological data were analysed. It was shown that November 2014 was colder than the average, suspecting that the act of pine processionary (Thaumetopoea pityocampa) was reduced the following year resulting into a single peak in July 2015, i.e., not a reduction of foliage around March. Additionally, May 2018 was on average around 3 degrees Celsius warmer than previous years, followed by a slightly warmer than usual April. The increased temperature may have been the cause of the blooming delay in August 2018.

Furthermore, the 6 years of investigation were divided into 4 years of testing period and 2 years of predicting period and the predictions were evaluated using rRSME. It was shown that on average there is an increased in the average backscattered coefficient but a reduction of the amplitude at the peak blooming, usually occurring in July. Nevertheless, the peak of 2020 was higher than the predicted one and this could be explained since the average precipitation of 2019 was 2.67- 0.99 higher than the average.

Overall, observing phenological changes of forests are important for understanding how climate change affects them, while long time alternations may relate to difficulty of plants to regenerate themselves and/or ecosystem biodiversity alternations. Due to the increased temporal resolution of Sentinel-1, we are now able to observe phenological changes derived by the structural information acquired from that mission. In this study, we investigated blooming timing changes and showed that, similarly to literature, alternations are most probably related to climatic conditions, while we should also be able to relate the data with pest attacks and fruition time.

The Project "ASTARTE"(EXCELLENCE/0918/0341) is co-financed by the European Regional Development Fund and the Republic of Cyprus through the Research Innovation Foundation.

Time series of optical remote sensing data are widely used to determine land surface phenology (LSP) and hence the timing of key stages, such as start (SOS), peak (POS) and end (EOS) of season. LSP is an indicator for the dynamics of vegetation growth and its response to environmental changes such as a warming climate. Therefore, the importance of accurate LSP retrieval is enlightened by the need of stakeholders and policy makers to take well-informed and reliable decisions.
The satellite observations underlying the time series used to derive LSP metrics are physical measurements. A measurement, however, is only complete if it is accompanied by a statement about its uncertainty. Here, uncertainty is understood as a statistical quantity that denotes the degree of doubt about a measured value. Applying the concept of uncertainty to optical satellite data, it follows that the amount of radiance measured by an optical sensor cannot be known exactly. In particular, the measured radiance can be seen as a single realization of a random distribution. The shape of this probability distribution indicates which other radiance values are possible apart from the measured one.
The scientific literature lacks a discussion about the influence of radiometric uncertainty in spectral satellite measurements on the timing of phenological stages. While there is awareness that the choice of phenology extraction algorithm significantly impacts the results, the uncertainty of the underlying measurements is mainly not addressed. However, all phenological models either rely directly on the spectral data or use them to derive plant physiological traits, whose temporal change in turn indicates phenology, uncertainty impacts most published LSP studies. Finally, it should be also determined whether the influence of the radiometric uncertainty is of a similar magnitude to other, already quantified sources of uncertainty. Consequently, the radiometric uncertainty of optical satellite data should be taken into account for LSP retrieval.
For demonstration purposes, we conducted a little case study. We acquired a time series of cloud-free, atmospherically corrected Sentinel-2 (S2) images using Google Earth Engine. Due to its high spatial, temporal and radiometric resolution, S2 is used for a wide range of phenological studies, including in particular agricultural and forestry applications. The time series of S2 images covered an entire growing season (Aug. 2017 to Aug. 2018) over a winter rapeseed field parcel in Northern Bulgaria. In total, 64 S2 images were available. For each image acquisition date we calculated the Normalized Difference Vegetation Index (NDVI), as it has been used by the majority of LSP studies and can be considered as a reliable reference product. To eliminate the potential impact of undetected clouds and shadows, we discarded the lower and upper 5% percentile of NDVI values and averaged the remaining pixels.
We assumed that the uncertainty u_NDVI of the NDVI values originating from the uncertainty in the original spectral bands was 2%. Though this is a rough estimate, it is still within the range reported by previous radiometric uncertainty studies for S2. Assuming that the uncertainty follows a Gaussian distribution with standard deviation u_NDVI and mean equal to the original NDVI value, we sampled 10 000 possible realisations for each acquisition date. Consequently, we obtained 10 000 NDVI time series realizations.
LSP metrics were then extracted using the in-house developed DATimeS (Decomposition and Analysis of Time Series Software). DATimeS is a stand-alone image processing GUI toolbox that enables to perform different advanced time series tasks for: (1) generating spatially continuous maps from discontinuous data using conventional fitting and smoothing functions as well as advanced machine learning regression algorithms, and (2) quantifying of phenological metrics (e.g., SOS, EOS, and POS) throughout multiple seasons. First, we linearly interpolated the time series to daily values and subsequently smoothed it using a Savitzky-Golay filter. Second, we extracted the SOS, EOS, and POS from the smoothed time series. SOS and EOS were derived using a seasonal amplitude threshold of 30%, while POS was defined at the date where the NDVI curve reached its global maximum.
We analyzed the spread in the timing of the phenological stages among the 10 000 time series realizations and calculated the standard deviation to quantify the resulting uncertainty. In addition we extracted the min-max spread to show how far the scenarios and, thus the uncertainty range diverged. For the sake of reproducibility, we have published the exploited data and generated scripts as well as the results obtained from DATimeS on Github (https://doi.org/10.5281/zenodo.5654859). The DATimeS software framework can be freely downloaded at: https://artmotoolbox.com/plugins-standalone/91-plugins-standalone/34-datimes.html.
The results reveal an influence of the radiometric uncertainty in the NDVI data on the timing of the LSP metrics: For the SOS, the standard deviation of the scenarios was one day; however, the maximum spread reached about 4 days. For POS, the standard deviation was highest (4 days) and the min-max spread reached 15 days. In the case of EOS, the standard deviation was less than one day and the min-max spread was only 2 days. These results are also shown in the attached figure. The figure visualizes the mean NDVI time series before and after smoothing alongside with the three LSP metrics and their uncertainty (red bars). We further applied the methodology developed in the exercise to winter cereal fields in Switzerland and obtained similar results. The uncertainty in the determination of SOS and EOS was between zero and several days. In the case of POS, the uncertainty was mostly larger, ranging from weeks to entire months when looking at min-max spread. In detail, the uncertainty of the derived LSP metrics depended mainly on the shape of the NDVI time series: steep increases or decreases resulted in a smaller uncertainty compared to a moderate change in canopy greenness.
The results of this exercise affirmed that radiometric uncertainty in S2 data impacts the retrieval of LSP metrics. Therefore, we suggest that radiometric uncertainty should be explored when deriving phenological phases. Our approach, however, is based only on a rough estimate of the radiometric uncertainty. In an extended setting, the radiometric uncertainty should be quantified using the S2 Uncertainty Toolbox (S2-RUT), which is an extension of the widely-used ESA SNAP software. Moreover, an uncertainty propagation chain is required to replicate the data processing workflow: Starting from the Level 1C Top-of-Atmosphere reflectance product, the uncertainty should be propagated through an atmospheric correction routine (e.g., Sen2Cor), scene classification to detect clouds and shadows, and finally a time series reconstruction algorithm.
This article is based upon work from COST Action CA17134 Optical synergies for spatiotemporal SENsing of Scalable ECOphysiological traits (SENSECO), supported by COST. In addition, we acknowledge the funding of the Swiss Science Foundation for the projects “PhenomEn” (grant number IZCOZ0_198091).

Satellite image time-series methods have contributed notably in monitoring plant phenology since we can observe how climate change has affected these ecosystems and how this will change in the following years. Researches report that heat stress and water scarcity which are expected in the Mediterranean will increase the frequency of forest mortality events thus, will affect forest diversity (Behrens, Georgiev, and Carraro 2010). Inside from all these predictions, the already affected Cyprus's natural resources are anticipated to further exacerbate climate change pressure due to the island's semi-arid climatic conditions (Zachariadis 2012). Recent climate simulations project a decrease in rainfall during autumn in all forested areas of Cyprus. That will lead to a gradual reduction of forest growth since autumn follows a continuous dry summer period where forests are under stress (Giannakopoulos et al. 2012). This study examines potential abrupt shifts in forest phenology influenced by LST, utilising Synthetic Aperture Radar (SAR) backscatter coefficient and Land Surface Temperature (LST) from Sentinel-1 and Landsat-8 datasets, respectively, within the period 2014-2021. Moreover, it will assess how Landsat LST and Sentinel-1 C-band SAR time series are correlated and investigate how the Sentinel-1 backscatter observables (VH, VV, VHVV-ratio) are correlated with LST. Moreover, in this study, the seasonal trend loess (STL), breaks for additive season and trend (BFAST) and Signal Extraction in ARIMA Time-Series (SEATS) will be explored and evaluated their performance in monitoring and detection of possible abrupt changes in forest seasonality. The study area is the Paphos Forest, managed by the Department of Forest. That forest could be described as a representative Mediterranean forest; thus, it is vital to monitor it because Mediterranean forests are expected to experience first climate change in Europe. More specifically, the study focuses on the Northeast, East and Southeast areas for the SAR images from descending orbit (Figure 1). Concerning the expected outcomes, as the temperature increases, we anticipate a decrease in the radar backscatter, thus finding a correlation between them to observe more efficiently the plant phenology and alternations in them.

Vegetation phenology is an important driver of water, carbon and energy surface fluxes and need to be correctly parameterized in land surface models to accurately represent the seasonal variability of the land-atmosphere interactions. Therefore, the calibration of the ORCHIDEE model (Krinner et al., 2005) against observations of vegetation phenology is a prerequisite of any attempt to study and integrate a new vegetation type. Here, we focus on mangrove forests which are growing along tropical coastlines in the intertidal region between sea and land. These ecosystems, although covering less than 1% of the total tropical forests of the world (Giri et al., 2011), play an important role in the coastal carbon cycle.
Mangroves are among the most productive vegetation types with rates equal to those of tropical humid evergreen forests and coral reefs (Alongi, 2018). They also offer a variety of ecologically and economically important resources for coastal human population as well as coastal protection. Despite these fundamental roles, mangroves are subject of numerous disturbances such as construction of shrimp farms (Proisy et al., 2016), deforestation, overfishing, pollution, erosion and urban development (Nathan et al., 2017). Thus, it has been estimated that over the last two decades, about 35% of the world’s mangroves has disappeared, with potential large impacts on the carbon processes and this decrease is still ongoing, especially in South-East Asia at a rate of 1 to 2% per year.
Numerous studies have been carried out in different parts of the world to measure the various sources/sinks of carbon in mangrove forests, but the spatial variability induced by climate (precipitation, temperature, wind, radiation), ocean features (water salinity, tide amplitude and currents), geomorphology (sediment accretion and erosion) and soil properties (organic content, temperature, …) increases the complexity of global scale estimations. In this context, it appears important to represent explicitly mangrove forests in Earth System Models (ESMs) in order to predict future changes, and this work started in the Institut Pierre Simon Laplace (IPSL) ESM (Boucher et al., 2020), through new pending developments in the ORCHIDEE component.
The monitoring of such complex coastal landscapes from space, requires high resolution imagery. To assess trees phenology, we have worked with Sentinel-2 data Multispectral Instrument (MSI) optical sensor and the Copernicus dataset at Level 1C produced by ESA and available on the Google Earth Engine Platform. The dataset covers the time period from mid-2015 up to nowadays, but, in order to benefit from the increased revisit thanks to the two satellites 2A and 2B, only the data after 2017 were processed. To estimate the phenological metrics on the five-year time series, the HANTS (Roerink et al., 2000) algorithm was applied on the time series of various vegetation indices (VI). The intercomparison of VI allowed us to select the ARVI indice (Kaufman and Tanré, 1992) among the others because of its capacity to better account for the atmospheric effects and reduce the noise linked to the time/spatial variability. Five regions of mangroves selected in Caribbean islands (Haïti, Guadeloupe), in Madagascar, French Guiana and Vietnam have been studied. For these sites, ARVI time series were extracted over the mangroves’ areas and over tropical inland forests nearby. With the lowest values observed at the end of the dry season, the ARVI seasonal cycles seem likely to be more explained by precipitation than by other atmospheric variables (air temperature, solar radiation, etc.) and the effect is even more significant in mangroves compared to inland forests. In mangroves, as water salinity is strongly linked to the freshwater amounts provided by rainfall and runoff, the dry season may impact both evapotranspiration stress and photosynthesis. This result suggests various pathways for the parameterization of the phenology in ORCHIDEE for mangroves vegetation. Besides, phenological dates have been extracted and compared within the sites. The results of the comparison will be presented.



References

Alongi, D. M. (2018). Mangrove forests. In Blue Carbon (pp. 23-36). Springer, Cham.

Boucher, O., Servonnat, J., Albright, A. L., Aumont, O., Balkanski, Y., Bastrikov, V., ... & Braconnot, P. (2020). Presentation and evaluation of the IPSL‐CM6A‐LR climate model. Journal of Advances in Modeling Earth Systems, e2019MS002010.

Giri, C., Ochieng, E., Tieszen, L. L., Zhu, Z., Singh, A., Loveland, T., ... & Duke, N. (2011). Status and distribution of mangrove forests of the world using earth observation satellite data. Global Ecology and Biogeography, 20(1), 154-159.

Kaufman, Y. J., & Tanre, D. (1992). Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE transactions on Geoscience and Remote Sensing, 30(2), 261-270.

Krinner, G., Viovy, N., de Noblet-Ducoudré, N., Ogée, J., Polcher, J., Friedlingstein, P., et al. (2005). A dynamic global vegetation model for studies of the coupled atmosphere-biosphere system. Global Biogeochemical Cycles, 19(1).

Thomas, N., R. Lucas, et al. (2017). "Distribution and drivers of global mangrove forest change, 1996–2010." PLoS ONE 12(6): e0179302 https://doi.org/10.1371/journal.pone.0179302.

Proisy, C., G. Viennois, et al. (2018). "Monitoring mangrove forests after aquaculture abandonment using time series of very high spatial resolution satellite images: A case study from the Perancak estuary, Bali, Indonesia." Marine Pollution Bulletin 131: 61-71 https://doi.org/10.1016/j.marpolbul.2017.05.056.

Roerink GJ, Menenti M, Verhoef W., Reconstructing cloudfree NDVI composites using Fourier analysis of time series. International Journal of Remote Sensing. Int. J. remote Sens. 21 (2000), 9: 1911-1917. 21. 10.1080/014311600209814.

Monitoring of crop growth, variability and dynamics over agricultural areas is needed to optimize management practices and thus to ensure global food security. Nonetheless, estimation of crop phenology from optical satellite data remains a challenging task due to the influence of clouds and atmospheric artifacts. Therefore, measures need to be taken to overcome these challenges and gain better knowledge of crop dynamics.
Since 2017, the ESA Copernicus Sentinel-2A & B (S2) multispectral imagery (MSI) have been providing high resolution optical imagery all over the globe with an observation frequency of 5 days. With 13 spectral channels and 10-60m spatial resolution, time series of these data offer untapped potential for monitoring cultivated areas. In this respect, the processing of S2 imagery in cloud-based platforms, such as Google Earth Engine (GEE), made tremendous advances in the last few years, allowing large-scale precise mapping of agricultural fields. These developments led to a shift in paradigm moving away from desktop-based processing through providing computing facilities around the clock and across the globe via the internet. With this, high-level satellite-based products can be generated on demand. Thus, the arrival of GEE enabled us to propose an end-to-end processing chain for vegetation phenology characterization using S2 imagery at large scale.
To achieve this, the following pipeline was implemented: (1) building hybrid Gaussian process regression (GPR) models optimized with active learning (AL) for retrieval of crop traits, such as leaf area index (LAI), fractional vegetation cover (FVC), canopy chlorophyll content (laiCab), canopy dry matter content (laiCm) and canopy water content ( laiCw), (2) implementing these models into GEE, (3) generating spatially continuous maps and gap-filled time series of these crop traits, and finally (4) calculating land surface phenology (LSP) metrics, such as start of season (SOS) or end of season (EOS), by using the conventional double logistic approach.
In respect to step (1): variable-specific training datasets were generated in the ARTMO software environment using PROSAIL model simulations, with training samples reduced in number but optimized in quality, i.e. representativeness, using the Euclidean-distance based (EBD) AL technique. In this way, light retrieval models could be generated via GPR, a ML algorithm which builds up a retrieval model by learning the non-linear relationships between the spectral signals and crop traits of interest. Subsequently, (2) the retrieval models were integrated into the GEE environment to perform mean value prediction on-the-fly. In this way, time series of crop traits based on S2 images were produced quasi-instantly over the area of interest. As demonstration of the workflow capability to easily reconstruct time series of S2 entire tiles all over the world, phenology maps from multiple crop traits were generated over an agricultural area in Castile and Leon, Spain. For this region also crop calendar data were available to assess the validity of the LSP metrics derived from crop traits. In addition, phenology derived from the Normalized Difference Vegetation Index (NDVI) was used as reference. NDVI not only proved to be a robust indicator for calculation of LSP metrics, but also served to demonstrate the good phenology quality of the quantitative traits products. Overall, good to high performance was achieved in particular for the estimation of canopy-level traits, such as LAI and laiCab, with normalized root mean square errors (NRMSE) of 9% and 10%, respectively.
Thanks to the GEE framework, the proposed workflow can be carried out anywhere in the world and for any time window, thus representing a shift in satellite data processing paradigm. Ongoing studies over other representative cropland areas will be also taken into consideration to assess the robustness of the proposed processing chain. With our study we provide a path towards reliable LSP metrics calculation at global scale, achieving meaningful insights on crop seasonal patterns in a changing environment that demands for an adaptive agricultural production.

Droughts have been projected to increase in duration, frequency, and intensity under future climate scenarios according to the sixth IPCC report. Many regions globally are likely to experience unprecedented droughts, leading potentially to adverse consequences for livelihoods that depend on water-limited vegetation productivity. Drought-induced vegetation mortality will reduce the land carbon sink and biodiversity, and lead to a positive land-to-atmosphere feedback which consequently may exacerbate global warming and climate extremes. Despite the exponential growth of scientific studies during the last decades, the mechanisms of drought-induced vegetation mortality are poorly understood. More research is needed to better 1) distinguish species-specific responses to drought, 2) understand how plant physiology, plant strategies, and soil and root microbes contribute to coping with water deficit, 3) characterize regeneration pathways and drivers, and 4) uncover compound effects of drought and relatively abrupt disturbances, such as insect outbreaks, wildfires, and anthropogenic activities. A prerequisite for reaching these goals are long-term systematic assessments of ecosystems at large scales and at high spatial resolutions, in particular those areas under threat of increasing drought pressure.

The ongoing DRYTIP project (i.e., drought-induced tipping points in ecosystem functioning) aims to address these research gaps through coupling advanced Earth observation technologies, field ecology, and dynamic vegetation models. Specific objectives of the project are: 1) to advance Earth Observation (EO) methods for robust assessment of change in ecosystem functioning; and 2) to gain new insights on vegetation stability (i.e. resistance to drought and recovery) over areas of known historical drought-induced vegetation mortality.

The first work package is to establish an EO-based framework that allows to effectively capture anomalies in vegetation phenology at high spatial resolutions, allowing to analyze effects of drought impacts on individual trees. To achieve this, we are conducting pilot studies in regions where local-scale woody vegetation mortality has been recorded due to droughts (e.g., California and Senegal). First, to generate a reference dataset, dead trees are visually interpreted from multi-temporal 30cm-resolution DigitalGlobal images and labelled on an existing segmentation map of individual trees (Brandt et al. 2020). A curated list of time series of EO proxies is derived from multi-source high-resolution satellite observations, i.e., PlanetScope, Sentinel-2, Sentinel-1, and Landsat. The model fit-based method used by Cheng et al. (2020) is then applied to extract tree-level phenological metrics followed by the assessment of phenological anomalies attributed to drought-induced mortalities. For comparison we apply a similar framework to MODIS time series. The results are expected to justify the applicability and advantages of high-resolution remote sensing time series in studying climate extremes and ecosystem stability in arid or semi-arid areas.

References

Cheng, Y., Vrieling, A., Fava, F., Meroni, M., Marshall, M., & Gachoki, S. (2020). Phenology of short vegetation cycles in a Kenyan rangeland from PlanetScope and Sentinel-2. Remote Sensing of Environment, 248, 112004.

Brandt, M., Tucker, C.J., Kariryaa, A. et al. An unexpectedly large count of trees in the West African Sahara and Sahel. Nature 587, 78–82 (2020). https://doi.org/10.1038/s41586-020-2824-5

Ongoing global climate change is having a major impact on the state of the environment. In agriculture, environmental changes affect the timing and length of the growing season, biomass production and, consequently, the overall harvest. Climate change can thus play an important role in sustainable food security in the region. Regional insight into the phenological cycles of vegetation, including agricultural crops, is provided by Land Surface Phenology (LSP) derived from high-resolution satellite data. Sentinel-2 and Landsat imagery are suitable data sources for regional LSP and have complementary potential. Landsat time series are suitable for analysing more than 35 years of phenological changes aggregated to the regional level, while the high temporal resolution of Sentinel-2 data allows phenological monitoring at the level of individual agricultural plots.
The aim of the present project is to demonstrate our approach to derive phenological indicators from Sentinel-2 and Landsat data by harmonized calculation of leaf area index (LAI) time series for dominant crop species in the Czech Republic. These are winter wheat (Triticum aestivum), spring barley (Hordeum vulgare), winter rape (Brassica napus subsp. napus), alfalfa (Medicago sativa), sugar beet (Beta vulgaris) and maize (Zea mays subsp. Mays). The calculation of LAI is based on the inversion of the ProSAIL radiative transfer model (RTM) along with regression using an artificial neural network (ANN). LAI values based on both Landsat and Sentinel-2 satellites were validated by in situ measurements and compared with each other. The achieved RMSE values for all crops were 1.28 and 1.43, while r was higher than 0.83 and 0.8 for LAI based on Sentinel-2 and Landsat, respectively. A high correlation was then achieved between the LAI derivatives (Sentinel-2, Landsat): r = 0.96, R2 = 0.9. Long-term phenological changes were then evaluated using a 35-year time series of LAI based on both types of data (Sentinel-2 and Landsat), aggregated to areas with similar climatic conditions.
The presented approach and analyses can serve as valuable input for assessing the impact of climate change on sustainable agriculture and future regional food security.

Forests are one of the largest above ground co2 storage landcovers and therefore, as essential climate variables, an important asset to quantify and monitor. The current availability of more than 7 years of data from the Copernicus program is shifting analysis methods from single time steps to multitemporal and time-series studies with an unprecedented spatial and temporal resolution. In particular, the SAR sensor onboard of Sentinel-1 (S1) A and B enables the estimation of phenologically active phases within days and weeks [1], measure of seasonality of different forest types [2, 3, 5] or fallen trees [4] at regular interval through a weather and daylight independency.

A high repetition rate is especially important in the detection of change points (beginning/end of growing season, or abrupt and permanent changes in land cover through, for example, logging). The possible resolution of changes in the observed area depends on the temporal sampling rate. S1 offers the possibility to increase the temporal sampling rate by using information from the twin satellites, reducing the repeat rate to 6 days over Europe. Additionally, overlapping orbits can be employed to increase data availability while including different viewing directions, resulting in one image roughly every 1.5 days. As recent examples of S1 time series studies, Soudani et al. [1] and Frison et al. [5] combined ascending and descending orbits for increased temporal sampling, relying on the fact that the local incidence angles of sensor and ellipsoid (θ) of both orbits are similar throughout the study area.

However, the combination of sensors and orbits as described above introduces systemic shifts in the data, if done without bias correction. We would like to highlight three mechanisms that result in such shifts.

a) S1 A and B show small differences in the measured intensities; in rainforest, for example, S1 A backscatter values were about 0.2 dB higher than S1 B values [6]. Systematic differences have been measured over various land covers and ranges of incidence angles [6]; they are also visible in our data in low backscatter areas such as lakes and airstrips (~1.5 dB) and even in meadows with low volume scattering (~0.5 dB).

b) The employment of more than one orbit, overlapping the same area, introduces further systematic differences within the intensity of time series of the SAR measurement. Specifically, measured intensities are higher at large θ. The observed differences between sensors and within same-orbit directions are marginal compared to the differences observed between different orbit directions, where more significant issues are encountered. Shifts of 5 dB and larger can be observed at a single pixel location with shallow and steep θ within ascending and descending orbits, respectively.

c) With a focus on the pixel scale, positional errors in measured intensities may result from elevation errors in digital elevation models (DEMs) used for topographically correcting the data. Forests are naturally subject to change over years. However, global SRTM and ASTER DEMs are the most widely used for topographic correction. More recently, in 2015 the TanDEM-X global DEM was released. Although this DEM is more representative of current vegetation canopies, all three global DEMs have a coarser resolution than S1, reducing the effect of radiometric terrain correction (RTC) processing and precision of geocoding. Through deviations in local height from the actual tree crown due to the time lag between S1 data and DEM, positional errors are to be expected when combining different orbits in an analysis. Depending on the scale of the studied features, these errors may affect the interpretation of the time series. Looking at a single orbit, the tree crown will be shifted towards the sensor. When using multiple orbits, these shifts are always in the direction of the respective range direction, rendering small-scale analyses impossible. Digital terrain models (DTMs) are also typically preferred as elevation models. Only Rüetschi et al. [4] use a digital surface model (DSM) for windthrow detection where no DTM was available, but the effect on the results was not discussed. However, the reported accuracies within the DSM-processed area mostly exceeded those obtained with the DTM.

For an improved detection of changes in vegetation, for the fine scale monitoring of phenology, we propose the use of DSMs in geocoding and terrain-correcting S1 data. Furthermore, we see an improvement the normalization procedure for heterogeneous S1 time series based on the orbit’s local incidence angle θ through a more accurate representation of the forest’s inclination with the DSM.

The proposed approach is evaluated in forested areas of the Hainich National Park (HNP) in Germany, a hilly but not mountainous area. We employ a LiDAR-derived DSM, acquired between fall 2016 and spring 2017, which is freely available (Geoportal-Th.de). Geocoding is performed using GAMMA through pyroSAR, a python framework for large-scale SAR satellite data processing [7]. Arguably, the C-Band signal penetrates the tree crown to a certain degree, but the height difference between the theoretical scattering center in the tree crowns and the recent LiDAR DSM is small. The time series of S1 IW GRDH data was processed for the 2017-2020 period. We considered 960 acquisitions from four orbits, each in descending and ascending azimuth direction. Backscatter in VV and VH was extracted in 𝛾0rtc at a 10 m spatial resolution after Small et al. [8], to mitigate influences in the intensity due to the underlying surface inclination. As mentioned above, differences in intensity over the local incidence angles were still present in vegetated areas, where volume scattering is the dominant scattering mechanism (see b).
To mitigate the observed patterns of a systematic shift within the forest we employ an orbit-wise linear regression. Based on the assumption of a homogeneous land cover, we select all pixels within a forest land cover class in the survey area. For the linear regression, temporal mean of both the backscatter and θ per orbit and sensor were used. The resulting coefficients then are used to normalize each orbits data. Opposing to Small et al. [8] our normalization is not based on model theory but on empiric data, also our local incidence angles differ from the incidence to flat terrain used in Dostálová et al. [3].

Within the survey area, there were no easily identifiable targets to quantify the increase of positional consistency. Therefore, we considered trees of two different health classes, measured with optical VHR, that should display different backscatter time-series for validation of the approach. Comparing the signal over time for both conventional processing and our method, the conventional processing shows little difference between the time series of the two health classes, suggesting mixed signals from the different orbits at a specific pixel location; with our method, in contrast, both classes are well separable. The normalization by local incidence angle over homogeneous land cover further increases their separability. The results suggest that our proposed orbit normalization of the local incidence angle together with an accurate DSM could be used for very dense, heterogeneous time series analyses of several orbit directions and incidence angles over forested areas with comparable backscatter.

Following these corrections, we expect that the impact of climate variables such as precipitation and temperature on the SAR time series can be analyzed at a substantially improved, very high temporal sampling rate, with which we aim to observe seasonal patterns and trends in the HNP.

[1] Soudani, K., Delpierre, N., Berveiller, D., Hmimina, G., Vincent, G., Morfin, A., & Dufrêne, É. (2021). Potential of C-band Synthetic Aperture Radar Sentinel-1 time-series for the monitoring of phenological cycles in a deciduous forest. International Journal of Applied Earth Observation and Geoinformation, 104, 102505.
[2] Dubois, C., Müller, M., Pathe, C., Jagdhuber, T., Cremer, F., Thiel, C., & Schmullius, C. (2020, July). Characterization of Land Cover Seasonality in Sentinel-1 Time Series Data. ISPRS.
[3] Dostálová, A., Wagner, W., Milenković, M., & Hollaus, M. (2018). Annual seasonality in Sentinel-1 signal for forest mapping and forest type classification. International Journal of Remote Sensing, 39(21), 7738-7760.
[4] Rüetschi, M., Small, D., & Waser, L. T. (2019). Rapid detection of windthrows using Sentinel-1 C-band SAR data. Remote Sensing, 11(2), 115.
[5] Frison, P. L., Fruneau, B., Kmiha, S., Soudani, K., Dufrene, E., Le Toan, T., Koleck T., Villard L., Mougin E. & Rudant, J. P. (2018). Potential of Sentinel-1 data for monitoring temperate mixed forest phenology. Remote Sensing, 10(12), 2049.
[6] Schmidt, K., Schwerdt, M., Miranda, N., & Reimann, J. (2020). Radiometric comparison within the Sentinel-1 SAR constellation over a wide backscatter range. Remote Sensing, 12(5), 854.
[7] Truckenbrodt, J., Cremer, F., Baris, I., & Eberle, J. (2019). Pyrosar: A framework for large-scale sar satellite data proessing. Proceedings of the Big Data from Space, Munich, Germany, 19-20.
[8] Small, D., Jehle, M., Meier, E., & Nüesch, D. (2004, May). Radiometric terrain correction incorporating local antenna gain. In Proc. of EUSAR (pp. 25-27).

Remotely-sensed crop phenology is a key variable for the assessment of ecosystems functioning, food security, or carbon and water balance. The robust retrieval of crop-specific phenology metrics depends on the density and regularity of valid observations, and on the matching between spatial resolution and field size. Sentinel-2 offers excellent spatio-temporal resolution, but only from 2017 onwards since the launch of Sentinel-2B and only in case of non-persistent cloud cover periods. Medium spatial resolution (MSR) daily observations increase the probability for cloud-free data and have been available for more than two decades, but do not offer an adequate spatial resolution. This study aimed to assess the use of long time series MSR data to derive crop -specific phenology metrics, by disaggregating the MSR pixels using high resolution data. For this purpose, we used daily PROBA-V 300 m, and Sentinel-2 10 m NDVI time series data over a region (12 033 km2) around Paris, for four main crops (i.e. winter cereals, spring barley, oilseed rape and maize) in 2015–2020. To address the mixel problem in PROBA-V data, we implemented a linear spatial disaggregation approach, which estimates crop-specific NDVI based on the crop fraction within a mixel provided by the Land Parcel Identification System (LPIS) map, considering each pixel neighborhood. Crop-specific phenology metrics extracted from both sensors were compared with ground data (1/ BBCH-scale of phenological stages from the French phenology observation network - TEMPO database and 2/ our field observations). Our results were trifold: i) Sentinel-2 provided crop-specific phenology metrics for all crops with an accuracy better than 10 days since 2017, ii) disaggregated PROBA-V increased the number of crop-specific NDVI valid observations at certain stages of the crop’s growth period, when 5-day Sentinel-2 revisit was insufficient with regards with the cloud conditions and iii) disaggregated PROBA-V also provided crop-specific phenology metrics with 10 day accuracy over a larger number of years (2015-2020) but only for the dominant winter crops. Therefore, disaggregated MSR data can be used for characterizing crop-specific phenology, for dominant crops in the study area. In addition, MSR data can improve the crop monitoring, especially during the transition periods when the NDVI quick changes are likely to be unobservable with Sentinel-2 if the cloud cover is persistent. Indeed, the spatial disaggregation could be easily computed at the scale of France, facilitating crop-specific phenological detection in synergy with Sentinel-2. This paves the way for monitoring crop-specific phenology over fragmented landscapes, at large scale, from sensors such as SPOT-VEGETATION, even for years before Sentinel-2 launch, as long as a land use map is available. This would allow the provision of a MSR-derived crop-specific phenological database, taking into account the change of the spatial resolution of the latter before 2013, and also taking into account the fact that the LPIS map gives the main crop type at the scale of the farm and not at the scale of the field before 2015, requiring further development. The spatial disaggregation could also be used over the Sentinel-3 data to ensure the time series continuity after the PROBA-V lifetime. In addition, in terms of perspectives, we are investigating the potential of Sentinel-1 radar data (since 2014) to complement the optical time series, as 1/ radar data is insensitive to clouds and 2/ radar data temporal changes may strongly relate to other phenophases than those observable with optical data.

In their phenological development, crops go through different phases, all of which are important for the final yield result. For example, flowering is a very sensitive stage for yield formation. Accordingly, phenology in agricultural applications is more than just defining the start and end of a growing season. Simple parameterisations of leaf area curves (start, max, end) are not appropriate for these applications. Instead, the phenology needs to be determined in agricultural terms, which is for example the BBCH code, a value ranging from 0 to 90. These BBCH major stages contain leaf development, tillering, stem elongation, inflorescence emergence, flowering, fruit development, ripening and harvest.
Phenological models exist which are mainly driven by temperature sum and describe the phenological progress. These are not only different for different crop types, but also very much depending on the initialization which is the seeding date for agricultural crops. Since the seeding date varies even in smaller regions within a couple of weeks, the phenological development of individual fields vary accordingly not only with climatic conditions but also depending on the farmers’ management decisions. Thus, a phenological model driven only by meteo data cannot adequately simulate crop phenology. Instead, data assimilation using satellite observations is required to constrain the phenological model for each studied field.
Vista established a regional yield prediction system, called Ypsilon, which offers yield predictions in t/ha for the major crops in Europe (wheat, barley, rapeseed and maize) with weekly updates starting in May each year. The forecast uses the mechanistic crop growth model PROMET, weather data and forecasts as well as Sentinel-2 data streams that are assimilated on a daily basis (see Fig.1).
In 2021, this service, which is accessible on the portal https://ypsilon.services, was extended by a crop status layer describing the BBCH major stages as listed above. The workflow includes a crop type classification, the selection of representative samples distributed over Europe (several hundred thousands of fields per crop type) and the simulation of the phenological development of each field, which is “calibrated” with Sentinel-2 observations using data assimilation techniques (ensemble particle filter). These individual fields are then aggregated on regions. Backtests for initial model parameterization like date of seeding and general phenological progress with time were further required, making use of the full Sentinel-2 archive. The presently smallest possible regional units for the concluding crop status in Ypsilon are illustrated in Fig.2. Here the dates for the beginning of flowering for maize was selected for illustration purposes.
The percentage of crops at a certain BBCH are summarized on regional level for each day of the growing season (Fig.3). Another information provided is the comparison of the current crop stage to the last three years at European level (Fig.4).

Land surface processes have been shown to substantially impact weather forecasting at short and medium ranges owing to their influence on the partitioning of energy, mass and momentum fluxes between the land surface and the atmosphere. The impact of these processes also extend over different spatial and time scales and can affect long-term climate projections . As a fundamental component of land surface models, the land use in general and more specifically the vegetation layer play a crucial role in the land–atmosphere exchanges. It was shown that the vegetation state affects the weather conditions via its control mechanism on evapotranspiration, development of planetary boundary conditions and clouds. In addition, it affects the available surface energy through the radiative transfer within the canopy by modifying the surface albedo.
The impacts of considering inter-annual and seasonal variabilities based on satellite derived products of structural changes through the Leaf Area Index (LAI), and land use /land cover (LULC) on surface and near surface parameters derived from offline and coupled runs of the ECMWF land surface system ECLand is investigated for future use in climate reanalyses.
LULC data from ESA-CCI/C3S is adapted to the model classification and resolution to provide a long term time series compatible with the ECLand surface modelling system. The Copernicus Global Land Services (CGLS) and THEIA LAI products are also opportunely processed to provide harmonized and consistent long-term time series that support global model simulations spanning the Satellite Era.
Simulations for three configurations using inter-annual varying land cover and LAI maps are performed. In the first configuration, the land cover is based on the operational GLCCv1.2 LULC map and MODIS-based climatological LAI maps, in the second the ESA-CCI/C3S LULC and CCGLS based climatological LAI map are used, in a third configuration a multi-annual ESA-CCI/C3S LULC and harmonized LAI maps are investigated looking at near surface weather and hydrological impact.

West Africa is an important hotspot of global change facing huge environmental and societal challenges. These include climate and land use change, migration, and conflicts, all of which are a major threat to food security. Food security – and a number of other envisaged achievements of the sustainable development goals (SDG) – depends largely on wise natural resource management. For several reasons, the West African environment and its changes are still poorly understood, although a large number of scientific studies have been conducted over the past years thanks to the establishment of the West African Science Service Center on Climate Change and Adapted Land Use (WASCAL) and other initiatives. However, most of the studies are focused on only few study sites, only very few aim at large-scale assessments of climate change impacts or land use. Little is known about vegetation structure, which plays a crucial role in the estimation of the greenhouse gas budget and the carbon sequestration potential of complex ecosystems such as agroforestry systems. Agroforestry systems are a mixture of different land uses, characterized by a certain tree cover and crops in between. Ideally, the trees do not only shade the fields but also provide fruits that can be used as food or to feed animals. Among the consistent datasets that provide more detailed information about vegetation properties throughout the region is the Copernicus Global Land Cover product (Buchhorn et al. 2020). It is available for multiple years (2015-2019) and provides rich information with regard to vegetation, particularly forests. The spatial resolution is 100 m. However, West African ecosystems are diverse and complex. This complexity is also true for agroforestry systems, which are important agricultural production zones and at the same time fulfill numerous ecosystem services. Unfortunately, none of the well-established nor the recent land cover and land use products such as the beforementioned Copernicus product are able to adequately resolve agroforestry systems. Even the WorldCover 2020 product (Zanaga et al. 2021) with 10 m spatial resolution is not suitable to differentiate between cropland areas, forest cover, shrubland and agroforestry systems. While our hypothesis is that the spatial resolution of the Copernicus Sentinel satellites is limiting the classification of single trees, we expect differences in the phenology within agroforestry systems that can be mapped by means of remote sensing. Phenology, the characteristic, often seasonal life cycle of plants, is an important plant species trait and hence one of the essential biodiversity variables. Many methods exist to retrieve phenology from optical remote sensing data. While the resulting information aids in differentiating plant species or plant functional types, satellite-derived products are usually different from what can be observed in the field. West Africa experiences a strong climate gradient from the hot and dry Sahara Desert region to the moist Guinean forest ecozone. In terms of optical remote sensing, capabilities to retrieve dense time series is limited by frequent cloud cover, particularly in the southern part of the region. Therefore, we propose to use Sentinel-1 Synthetic Aperture Radar (SAR) data to retrieve phenology at pixel level (10 m spatial resolution). In recent years, Sentinel-1 SAR data is increasingly used to characterize phenology of field crops. Little is known about phenology of West African vegetation, particularly non-crops. Consequently, we sampled all classes of the Copernicus Land Cover product covering the ECOWAS region in West Africa and explored Sentinel-1 time series. Our pre-processing includes radiometric terrain correction, speckle filtering and time series smoothing using a Savitzky-Golay filter. For West Africa, only data in ascending orbit is available, resulting in a reduced temporal resolution compared to other regions. From the two polarizations, VV and VH, we computed several well-established indices (e.g. VH/VV ratio, radar vegetation index). We sampled the whole region, resulting in 250 samples per land cover class. For each sampling point we extracted the time series of the backscatter as well as the indices and tested their similarity. As the Copernicus product also provides fractions of each class, we were able to explore the relationship between fractional tree cover (and others) and the SAR backscatter and indices, respectively. Our results show that some of the classes are no longer separable at high spatial resolution (e.g. open evergreen forest vs. closed evergreen forest). After adequate join of similar classes, we were able to use the backscatter information as well as the uncorrelated indices to map Copernicus land cover classes at high spatial resolution (10 m) with acceptable accuracy.
From the smoothed time series, we derived phenological parameters such as start of season, end of season and length of season, greenup and senescence. While a direct link to ground phenology is challenging, we are able to map groups of similar phenological behavior, which is important for a more comprehensive characterization of vegetation.

Land surface phenology, the study of seasonal dynamics of vegetation phenology from satellite data, has emerged as an important scientific research focus in recent decades, as changes in LSP patterns have been considered as a key bio-indicator of climate change. LSP dynamics has generally been observed from the study of the dynamics of phenological metrics or phenometrics extracted from vegetation indices time series (VI). Phenometrics, such as the start of the growing season (SOS) (i.e., spring phenology) or the end of the growing season (EOS) (i.e., autumn phenology), could be considered as proxies of spring phenophases (e.g., leaf unfolding or flowering) or autumn phenophases (e. g. autumnal colouring of leaves or leaf falling).
Mediterranean ecosystems are especially vulnerable to effects of climate change and variations in climate conditions may be alter the numerous ecosystem services they provide. Therefore, LSP could improve the understanding on how Mediterranean ecosystems adapt to global changes. However, specific characteristics of Mediterranean ecosystems may make the extraction of phenometrics a challenging task. Thus, this literature review aims to provide a synthesis of the satellite data (i.e., sensors, VI) and methodologies (i.e., smoothing and phenometric extraction techniques) used in the LSP studies focused on Mediterranean ecosystems.
This literature review was performed using two electronic databases: Scopus and Web of Science. This focused on LSP articles based exclusively on analysing spatial or spatio-temporal patterns of phenometrics in the terrestrial ecosystems of the Mediterranean Basin. These studies should be published in peer-reviewed scientific journals between 2000 and 2019.
Although multiple satellite datasets and methodologies can be used for estimating phenometrics, there is no consensus regarding which of them is the most ideal. This could depend on the biological conditions of the ecosystems (e.g., dominant vegetation type, vegetation density, diversity of species, etc.) and the climate conditions (e.g., frequency of cloud cover, atmospheric aerosol content, etc.).
Mediterranean ecosystems are ecosystems with a wide diversity of plant species, concentring numerous endemic vegetative species. Hence, phenological signal in some Mediterranean areas may be influenced by the phenological cycles of multiple species. Several researchers focused on other temperature ecosystems reported the need to analyse LSP patterns at a high spatial resolution (e.g., Landsat or Sentinel data) in ecosystems with significant heterogeneity of plant species, such as Mediterranean ecosystems. However, most of LSP studies included in this review used Advanced Very High-Resolution Radiometer (AVHRR) and Moderate Resolution Imaging Spectroradiometer (MODIS) data for estimating phenometrics at a coarse spatial resolution (i.e., from 500 m to 8 km).
Mediterranean ecosystems are characterised by a medium-low vegetation density and the phenological signal may be altered by soil effects. Multiple LSP studies focused on Northern Hemisphere or based on a global scale excluded areas where the VI values were low (e.g., annual mean VI value < 0.1 or 0.2). Hence, an important extension of Mediterranean ecosystems was excluded in these studies. Normalised Difference Vegetation Index (NDVI) was the most-used VI in LSP studies (69%). This VI allowed the analysis of seasonal variability of plants, minimising cloud shadows or topographic effects. However, some studies reported that Enhanced Vegetation Index (EVI) (or EVI based on two bands; EVI2) may reduce the influence of soil background in ecosystems with sparse vegetation (e.g., Mediterranean ecosystems). However, only 16% of LSP studies used EVI or EVI2.
Both, soil and other factors influence (e.g., cloud cover, sensor disturbances, etc.) can be reduced from smoothing techniques. Savitzky-Golay algorithm and different logistic functions, such as piecewise logistic function or double logistic function, were the most used in LSP studies. Phenometrics were extracted from smoothed VI primarily using threshold-based methods (58%) and derivative-based methods (47%). Phenometric accuracy may be related to the smoothing algorithm and extraction technique. Hence, both the selection of a specific algorithm for denoising VI time series and the selection of an extraction technique of phenometrics are complicated tasks. Several studies assessed the performance of different smoothing techniques and extraction techniques over non-Mediterranean ecosystems. Due to the specific characteristics of Mediterranean ecosystems, it would be necessary to carry out a similar study in these ecosystems.

Earth system sciences investigate dynamic processes and cycles of materials which follow distinctive timings towards a better understanding of climate and biosphere interaction. Phenology of vegetation reveals specific temporal pattern for different vegetation types and is affected by abiotic environmental influences. Changing abiotic factors relate directly to changes in plant functional traits and affect plant phenology in consequence. Various models exist which describe plant phenology directly at the level of functional traits using in-situ characterization and have been validated against remotely sensed proxies at leaf and canopy level.
We examined high frequency time series of Vegetation Indices (VI) as a proxy for plant phenology obtained from autonomous, monitoring field spectrometer systems installed above dewberry (Rubus caesius) and pioneering vegetation in a river shore area over several months. An empirical regression model based on the linear combination of the abiotic factors photosynthetic active radiation (PAR), temperature, relative humidity, precipitation and time was constructed to describe the temporal pattern of VI in each site, respectively. In eliminating each of the factors separately from the linear combination, the influence of the environmental factors on the final model outcome was determined and compared with the measured phenology. Our results suggest a good agreement (R² ~ 0.9) between the complete model and the measurements in both sites. The eliminated factors affected the modelling outcome noticeably and in a very specific way. For example, precipitation revealed only a minor influence on the dewberry canopy model. Time exhibited the strongest influence on the empirical model followed by light and temperature to accurately follow the phenology in NDVI of dewberry. In the pioneering vegetation was time, followed by temperature, precipitation and relative humidity of the strongest influence on the modelling outcome. The most important variables for the complete models were time in the dewberry canopy, indicating mostly seasonal driven phenology, whereas precipitation together with temperature had the highest explanatory power in the pioneers, indicating mostly event-driven phenology (flooding events).
Thus, investigated temporally dense time series of VI derived from hyperspectral reflectance recorded over dewberry and pioneers revealed a specific phenological patter in each vegetation type which reflects the typical interaction between plant and habitat. Those time series of VI from RoX spectrometers showed a good potential towards integrating abiotic factors to explain plant phenological changes at a high temporal resolution. We were able to describe the specific phenology using a linear combination of abiotic factors and determined their individual influence on the plant phenology in two different canopies, respectively. Further research should investigate more closely other canopies and seasons towards enabling application scenarios in assessing canopy response to environmental factors or canopy development prediction. Furthermore, ground-based high-frequency time series of VI carry large potential to inform and enhance satellite observations, for example from the Sentinel family. The ground measurements open a way to bridge and integrate phenological insight across scales towards understanding vegetation phenology in response to the environment at a larger scale from satellites.

The Earth's biosphere and the phenology of vegetation are at the heart of climatic, economic and social concerns. Human activities have led to a significant degradation of ecosystem services (e.g. carbon sequestration, biodiversity, water quality, flood, and erosion regulation) provided by various extensive ecosystems such as forests, grasslands or crops.

A key parameter for relevant climate modeling, public policy implementations or commercial applications is the temporal resolution at which vegetation is observed. As a tool providing synoptic and regular coverage of Earth’s surfaces, satellite Earth Observation has been increasingly adopted, among others, for estimating biomass, yields, modeling different fluxes or detecting changes. Optical images have been historically used for vegetation monitoring, considering their efficient discrimination of phenomena related to photosynthetic activity.

To deal with missing data due to clouds, many interpolation strategies integrating one or more optical sensors have been developed. Most of these strategies are based on trend modelling that does not reflect the real evolution of the vegetation cover in many cases (sudden climatic impact, man-made effects). As a result, data that may be weeks or months apart are often interpolated on areas suffering from high cloud cover.

Copernicus Sentinels provide new opportunities and unprecedented observations for the monitoring of vegetation’s dynamics. In particular, concordant optical and SAR data sets provided by the Sentinel-1 and 2 satellites open the door to new multi-sensor methodologies aiming at the reconstruction of missing information.

Taking into account the still numerous non-cloudy observations provided by the Sentinel-2 satellites, a deep learning regression methodology, namely the Sentinels Regression for Vegetation Monitoring (SenRVM), has been developed. Its goal is the translation of SAR features acquired regardless of the climatic conditions into NDVI. The developed architecture integrates several deep learning architectures such as Multilayer Perceptron and Recurrent Neural Networks. The SenRVM regression strategy proposes the integration of auxiliary data such as climatic and topographic features. This allows accurate NDVI time series to be predicted by minimizing effects exogenous to the vegetation’s phenology through SAR acquisitions contextualization.

Object-oriented analysis of the results is carried out on large scale areas for various vegetation types with distinct phenologies (grasslands, crops and forests). The results are analyzed by taking into account spatial and temporal aspects or with an ablation study of the Network’s inputs. The proposed approach is further compared with traditional interpolation methods exploiting monomodal (Whittaker smoothing, linear weighted interpolation) or multimodal (Random Forest, Gaussian Regression Processes, single Multilayer Perceptron) features.

The potential of high-temporal NDVI time series obtained by the SenRVM method for several vegetation-related applications is subsequently illustrated. In particular, the interest of the obtained time series to observe the phenology and its associated parameters of the three main vegetation classes is presented.

To understand the vegetation status and evolution, it is necessary to use a multi-frequency Earth Observation approach. Visible/near-infrared indices are sensitive to green components of the vegetation as they are related to the photosynthetically active parts of the vegetation from widely used indices such as NDVI (Normalized Difference Vegetation Index), to FAPAR or the more innovative solar-induced fluorescence (SIF). FAPAR quantifies the fraction of the solar radiation absorbed by living leaves for the photosynthetic activity while SIF represents an emission of energy emanating from chlorophyll molecules that is a function of the activity of competing pathways for de-excitation (photochemistry versus nonphotochemical quenching). These indices can be used as proxy for the Gross Primary Production (GPP) but they saturate quickly even for moderate biomass values ( < 80 Mg/ha).

On the other hand, active microwave (MW) observations allow to map the Above Ground Biomass (AGB) distribution using the backscattered signal of Synthetic Aperture Radars (Bouvet et al. 2018). Active and passive MW observations also give access to the hydric state of the vegetation via the optical depth created by the water molecules contained in the plants/trees (hereafter VOD, Vegetation Optical Depth). Depending on the frequency, MW radiation is sensitive to the water content in distinct parts of the vegetation, from the leaves, to the branches and the trunk (with small water-containing elements being transparent for long wavelength radiation). Liu et al. (2015) showed that passive MW VOD with frequencies higher than 6 GHz can also be used to estimate AGB while Rodriguez-Fernandez et al. (2018) showed that passive L-band (1.4 GHz) VOD (L-VOD) is even more sensitive to AGB, without significant signs of saturation for high AGB values (300 – 400 Mg/h).

In this contribution we will discuss a panchromatic analysis of the vegetation using:
• Passive MW VOD at different frequency bands: SMOS L-VOD (Kerr et al. 2012), AMSR-2 C-VOD, X-VOD (van der Schalie et al. 2017)
• Active MW VOD at C-Band: ASCAT VOD (Vreugdenhil et al. 2016)
• Sentinel 3 FAPAR (Reyes Muñoz et al., this symposium)
• Sentinel 5P SIF (Guanter et al. 2021)
• ESA Climate Change Initiative AGB maps (Cartus et al., this symposium)

Three 500 x 500 km2 regions in the Iberian Peninsula, Northern Finland and Central Europe were studied using time series in the period from 2010 to 2021.

The VOD data from different sensors were compared showing a significant complementarity. Bias and temporal correlation maps of ASCAT and AMSR-2 VOD with respect to SMOS L-VOD show regions with positive values and regions with negative values. These maps were compared to land cover maps but no clear relationship was found. The temporal correlation of SIF and FAPAR was found to vary between the focus areas with values from R=0.59 in Northern Finland, to 0.71 in the Iberian Peninsula and 0.81 in Central Europe. FAPAR and SIF seasonal cycles are similar but FAPAR increases in early spring and SIF only in late May or June. A comparison between microwave and visible data was performed by comparing SMOS VOD and Sentinel 5P SIF. Temporal correlation maps show regions with negative values, in particular in the Iberian Peninsula.

This study is part of the ESA Land Carbon Constellation (LCC, Scholze et al., this symposium), whose goal is to constrain Carbon-cycle models using the assimilation of multi-frequency vegetation-related data (Kaminski et al. this symposium). The LCC project helps to pave the road to studies that will be done in the future with dedicated SIF observations by the ESA FLEX mission (Moreno, 2021) and high resolution AGB maps from the ESA Biomass mission (Le Toan et al. 2011).

References

Bouvet, A., Mermoz, S., Toan, T. L., Villard, L., et al..: An above-ground biomass map of African savannahs and woodlands at 25 m resolution derived from ALOS PALSAR, Remote Sens. Environ., 206, 156–173, 2018

Guanter, L., Bacour, C., Schneider, A., Aben, I., van Kempen, T. A., Maignan, F., ... & Zhang, Y. (2021). The TROPOSIF global sun-induced fluorescence dataset from the Sentinel-5P TROPOMI mission. Earth System Science Data, 13(11), 5423-5440.

Kerr, Y., Waldteufel, P., Richaume, P., et al.: The SMOS Soil Mois- ture Retrieval Algorithm, IEEE T. Geosci. Remote S., 50, 1384– 1403, 2012

Le Toan, T., Quegan, S., Davidson, M. W. J., et al.. (2011). The BIOMASS mission: Mapping global forest biomass to better understand the terrestrial carbon cycle. Remote sensing of environment, 115(11), 2850-2860

Liu, Y. Y., van Dijk, A. I. J. M., de Jeu, R. a. M., Canadell, J. G., Mc- Cabe, M. F., Evans, J. P., and Wang, G.: Recent reversal in loss of global terrestrial biomass: supplementary information, Nat. Clim. Change, 5, 1–5, 2015.

Moreno, J. F. (2021, July). The Fluorescence Explorer (FLEX) Mission: From Spectral Measurements to High-Level Science Products. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS (pp. 115-118). IEEE

Rodriguez-Fernandez, N. J., Mialon, A., Mermoz, S., et al.: “An evaluation of SMOS L-band vegetation optical depth (L-VOD) data sets: high sensitivity of L-VOD to above-ground biomass in Africa”. Biogeosci., vol. 15, no. 14, pp. 4627–4645, 2018

van der Schalie, R., de Jeu, R. A., Kerr, Y. H., et al.. (2017). The merging of radiative transfer based surface soil moisture data from SMOS and AMSR-E. Remote Sensing of Environment, 189, 180-193

Vreugdenhil, M.; Dorigo, W.A.; Wagner, W.; de Jeu, R.A.; Hahn, S.; van Marle, M.J. Analyzing the Vegetation Parameterization in the TU-Wien ASCAT Soil Moisture Retrieval. IEEE Trans. Geosci. Remote Sens. 2016, 54, 3513–3531

Land surface phenology (LSP), defined as the study of seasonal behaviour of phenological phases of plants observed from satellite imagery time series, has played an important role in monitoring the global terrestrial ecosystems dynamics. LSP approach is generally based on the study of seasonal patterns of phenological metrics or phenometrics derived from vegetation indices (VI) time series. These phenometrics (e. g., start of the growing season (SOS), end of the growing season (EOS)), ecologically-meaningful metrics considered as indicators of timing of spring and autumn phenophases, have been estimated from different VI and methodologies (i.e., smoothing and phenometric extraction techniques). However, there is no agreement in the scientific community on which of them are more suitable for studying LSP. The selection of a vegetation index and a methodology for estimating phenometrics depends on multiple factors: a) atmospheric factors (frequency of cloud cover, presence of atmospheric aerosols, gaseous absorbers); and b) ecosystem factors (dominant vegetation functional type, degree of vegetation cover, vegetation density, landscape heterogeneity). Although several studies have carried out inter-comparisons of VI and assessments of LSP estimation methodologies in multiple ecosystems, none of these have focused on Mediterranean ecosystems. The unique characteristics of Mediterranean ecosystems (e.g., heterogeneity of plant species, medium-low density of vegetation) can make LSP estimation a complex task.
Thus, this study aims to evaluate the differences in LSP estimations based on medium spatial resolution satellite data as a result of using different estimation methods (i.e., smoothing and phenometric extraction techniques) and VI. SOS and EOS were estimated from two VI (Normalised Difference Vegetation Index (NDVI) and a two-band Enhanced Vegetation Index (EVI2)), which were calculated from the MOD09Q1 product based on 8-day composites at a spatial resolution of 250 meters. Four smoothing algorithms were used for denoising raw time series (i.e., double logistic function (DL), asymmetric Gaussian function (AG), Savitzky-Golay filter (SG) and Discrete Fourier Transform (DFT)) and four techniques were applied for extracting phenometrics from the smoothed time series (i.e., first derivative-based method (DER), second derivative-based method (2DER), threshold method (TH) and inflection point method (INF)).
The mean difference between methods for estimating phenometrics was higher in EVI2 than in NDVI (32.7 days vs. 29.9 days (SOS), 28.3 days vs. 11.5 days (EOS)). The extraction technique had a greater influence on the difference between methods than the smoothing technique. The methods based on the techniques of extraction of TH, INF, and 2DER (except for DFT-2DER) were similar to each other, and these techniques can be considered as equivalent. The mean difference between the methods based on TH, INF and 2DER (excluding DFT-2DER) were 11.7 days for the estimation of SOS from NDVI and 13.7 days for the estimation of SOS from EVI2. However, the methods based on DER showed important differences in relation to the other methods (except for DFT-2der). The mean difference was 51.3 days for SOS estimation from NDVI and 60.5 days for SOS estimation from EVI2. In relation to the EOS, two very different behaviours related to the VI were observed. In the case of EVI2, the behaviour was similar to that previously described for SOS, with differences between the equivalent methods (TH, INF, 2DER) of 15.3 days, and with differences between the most disparate methods (DER with respect to the rest of methods) of 48.1 days. Surprisingly, the differences were low between methods for estimating EOS from NDVI.
In general, the patterns of differences between methods for SOS-NDVI estimation were very similar to the general patterns for all land covers, except for closed scrublands, where the difference between methods did not show any pattern. The differences between methods for estimating EOS-NDVI was random for all land covers. However, a pattern similar to that of SOS-NDVI is still observed in deciduous and mixed forests. The differences between methods for estimating EOS from EVI2 showed a pattern similar to that of SOS, but this was more attenuated for evergreen needleleaf forests, evergreen broadleaf forests and closed scrublands. Therefore, these results could suggest a lower robustness in the estimation of EOS in relation to SOS.

Smallholder farming accounts for roughly 80% of the available cropland in Africa. Food insecurity and diminished crop yield are chronic issues faced in Africa due to degraded soils, pest outbreaks and extreme weather events related to climate change. In spite of the importance and abundance of smallholder farming systems in Eastern Africa, their agricultural properties remain sparsely measured and understood.
In this regard, knowledge on crop phenology is highly relevant for a multitude of environmental and social topics, such as application of nutrients, management of land use, water distribution and irrigation. Furthermore, information on crop phenology is essential in combating food insecurity and understanding the impact of climate change on smallholder farming systems. Due to the volatile state of crops in Africa and projected negative trends to future yields, better adaptation and mitigation techniques are crucial to face current and future challenges in smallholder farming systems.
With this in mind, this research aims at (i) determining phenological variations at field-level in East African smallholder farming systems via a combination of datasets collected by multiple earth-observation missions and (ii) investigating the impact of regional climatological trends on the field-level variation by jointly analyzing the remote sensing record and meteorological observations. Thereby, several passive multispectral satellite datasets (Landsat 7, Landsat 8 and Sentinel-2) are combined and harmonized to produce a broader time-series and to minimize measurement gaps.
From these harmonized series, the Normalized Difference Vegetation Index (NDVI) was selected as a proxy to track farming field phenology through time. Furthermore, active Synthetic Aperture Radar (SAR) data from the Sentinel-1 mission is added to further increase the number of observations and to close remaining observational gaps in the series. SAR data are incorporated by applying a multivariate regression allowing to extract pseudo-NDVI values. This was achieved by expressing the Sentinel-2 NDVI as a function of VV and VH backscatter, as well as, local incidence angle and the normalized ratio between the two polarizations.
From the harmonized NDVI curves, phenological metrics (e.g. start, end, midpoint, duration and peak of the growing season) are computed for selected smallholder fields to obtain a continues depiction of the phenological cycles. Finally, climate indices are utilized to depict the climate fluctuations through time and to compare them to the phenological metrics. Most known examples of these indices include the El Niño Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD) and the Pacific Decadal Oscillation (PDO). The contribution will show first results on the effects of large-scale climate fluctuations on smallholder farming systems in East Africa.

Conventional manually derived phenological observations are usually labor-intensive. To overcome this, information regarding the phenology of vegetation can be drawn from Sentinel-2 times series. The high temporal availability of Sentinel-2 imagery for wide areas enables to derive phenology of different spatial locations. Switzerland as a rather small country (41,285 km2) comprises a great variety of landscapes, mostly caused by the complex topography (elevations between 193 and 4,634 m.a.s.l.) and subsequently even the same vegetation types show different phenological patterns. Especially for forests the knowledge about phenological patterns of tree species is essential to estimate and extend the viability of tree species.
In the present study, we are linking ground data of representative forest tree species from the Swiss National Forest Inventory with Sentinel-2 time series and analyzed the variability of the most common forest tree species in Switzerland in a spatial, temporal and spectral context. The required forest tree species information is obtained from 6,500 terrestrial plots from the Swiss National Forest Inventory distributed on a regular 1.4 km raster all over the country. We stratified the available tree species information into groups with similar conditions such as meteorology, elevation, aspect, slope, or biogeographical regions for comparing filled, continuous curves of Sentinel-2 data to each other. Identified information about start, end and length of seasons additionally serve as indicators of phenology. Furthermore, we compared spectral signatures interspecific and intraspecific.
Two research questions are to be answered in this work. Q1: How does the phenology derived from Sentinel-2 time series of a particular tree species vary through space and time in Switzerland? Our further interest is, if tree species are still showing unique features in the Sentinel-2 time series throughout the whole country to differentiate one particular tree species from another. That leads to our second research question. Q2: Which tree species are distinguishable by their Sentinel-2 spectra through space and time in Switzerland? The results have the potential to serve as new baseline for extensive phenological observations of forest tree species or can be used as benchmark for large-scale tree species classifications. The knowledge about phenological patterns may also be used to adjust forest management strategies in order to optimize for example timber availability, biodiversity richness or the carbon storage.

The new pan-European Copernicus High-Resolution Vegetation Phenology and Productivity (HR-VPP) service from Sentinel-2 provides, among other parameters, product layers on phenological dates (start, peak, end) and seasonal and total annual productivity at the spatial resolution of 10 m for the period 2017 to 2020. The product could be applied as input data in the assessment of the Sustainable Development Goal (SDG) indicator on the status of land degradation (15.3.1) at the national scale, for which the Finnish Environment Institute (SYKE) collects information for the reporting to Statistics Finland. Furthermore, phenology and productivity parameters could aid the mapping of habitats and ecosystems, and phenological dates are a required input to the modelling of carbon and nutrient cycles and biodiversity (e.g. insect phenology). The suitability of the product for the national use needs to be evaluated. Here, we i) compared the HR-VPP phenological dates against webcam observations and ii) tested the HR-VPP parameters as input data source for the mapping of natural habitats in Northern Lapland.
For the evaluation of the HR-VPP start, peak and end of season dates, we utilized the Monimet webcam network (Peltoniemi et al. 2018b) that was established for the monitoring of phenology and snow cover. Pre-processing of the webcam time series was carried out with the FMIPROT software (Tanis et al. 2018). Phenological dates were extracted for the period 2017 to 2020 from the webcam time series of the Green and Red Chromatic Coordinate following Peltoniemi et al. (2018a) and by using the Phenopix software (Filippa et al. 2016). First results indicated, a very good agreement of the HR-VPP start of season dates in wetland and deciduous forests with the webcam dates. We found lower correspondence for the peak and the end of season dates.
The mapping of natural habitats in Northern Lapland is carried out in the Finnish Ecosystem Observatory (FEO, www.feosuomi.fi/en, 2020-2024) and the Remote sensing of northern habitats project (https://www.metsa.fi/projekti/yla-lapin-kaukokartoitus/). Up-to-date and accurate information on the occurrence and status of habitat types ensure maintenance and restoration of conservation values of threatened habitats. This includes detection of possible changes and pressures in these areas (e.g. due to reindeer grazing), which may deteriorate the habitat quality. Since 2020, Metsähallitus and SYKE have been testing novel methods in updating the geometry and attributes of the habitats database (SAKTI) for an area of 30 000 km2 by using merely remote sensing data. For the prediction of habitat characteristics (inventory classes) and the Natura 2000 habitat types, the Random forest (RF) algorithm is used. Specifically, for the prediction of inventory classes, phenology and productivity metrics derived from monthly normalized difference vegetation index (NDVI), such as maximum, amplitude, sum as proxy for productivity, were found useful. Based on the statistical analysis of measured field data (2780 samples measured by Metsähallitus in summer 2020 and 2021), it was shown that HR-VPP seasonal productivity and amplitude describe well the increase in productivity level from bare rocks – lichen – moss to herb-rich habitat types, similar to the NDVI-based features. Thus, the HR-VPP parameters might also be useful in the monitoring of the degradation of herb-rich habitats with high biodiversity value in northern Lapland. However, the accuracy and importance of HR-VPP parameters in the predication of habitat types was lower than for the NDVI-based features and further investigations of the causes of the lower performance are needed.



References:

Filippa, G., Cremonese, E., Migliavacca, M., Galvagno, M., Forkel, M., Wingate, L., Tomelleri, E., Morra di Cella, U., & Richardson, A.D. (2016). Phenopix: A R package for image-based vegetation phenology. Agricultural and Forest Meteorology, 220, 141-150.
Peltoniemi, M., Aurela, M., Böttcher, K., Kolari, P., Loehr, J., Hokkanen, T., Karhu, J., Linkosalmi, M., Tanis, C.M., Metsämäki, S., Tuovinen, J.-P., Vesala, T., & Arslan, A.N. (2018a). Networked web-cameras monitor congruent seasonal development of birches with phenological field observations. Agricultural and Forest Meteorology, 249, 335-347.
Peltoniemi, M., Aurela, M., Böttcher, K., Kolari, P., Loehr, J., Karhu, J., Linkosalmi, M., Tanis, C.M., Tuovinen, J.P., & Arslan, A.N. (2018b). Webcam network and image database for studies of phenological changes of vegetation and snow cover in Finland, image time series from 2014 to 2016. Earth Syst. Sci. Data, 10, 173-184.
Tanis, C.M., Peltoniemi, M., Linkosalmi, M., Aurela, M., Böttcher, K., Manninen, T., & Arslan, A.N. (2018). A System for Acquisition, Processing and Visualization of Image Time Series from Multiple Camera Networks. Data, 3, 23.

Phenological shifts in forests and other vegetation types are observed in many parts of the world and mostly match with predictions based on climate simulations. However, weather station networks - and in particular phenological gardens - are generally sparse. The existing stations often record phenological events without proper coordination - and long-lasting historical records are restricted to a few places and particular vegetation types (e.g. some French vineyards record the start of the grape harvest since more than 600 years). Elevated and mountainous areas such as the Alps, are even less covered by in-situ data and this despite the possible effects of shifting phenologies on ecology and many other alpine ecoservices, as well as feedback mechanisms into the climate system. Likewise, climate simulations are often too coarse to cover the spatial variability and topography of environmentally important landscapes such as the Alps.
With an expected increase in air temperatures for the Alpine region, shifts in phenology are expected - however, no suitable monitoring system exists so far. The developed forest phenology service will address these issues and is one of six EO-based information services of the “ECO4ALPS – Alpine Regional Initiative (eo4alps) – Applications” of ESA. The service will use as inputs optical time series data from Sentinel-2 (2016-2020), Landsat-8 and Copernicus Forest HR layers to derive multi-annual trend information (e.g. earlier or delayed end-of-season) for two forest classes: conifers and deciduous forests. The phenology service will complement the sparse in-situ measurements and will provide full coverage information on land surface phenology (LSP) with focus on the less well studied autumn phenology. The service covers forested areas at deca-metric resolution without the need for spatial interpolations. Such an EO service is urgently needed for the entire Alpine region for climate change mitigation and policy formulation.
To derive the LSP, all datasets are processed into BoA (Level-2A). Cross- calibration of sensors is ensured as well as spatial co-registration. Using TIMESAT software, the end-of-season is detected for each year from NDVI time series using appropriate threshold values. The detected trends are analysed with respect to topography (in particular elevation and aspect) using an appropriate DEM. The outcome of the service is a detailed map at deca-metric resolution depicting areas with significant trends in forest end-of-season timings.

Understanding of vegetation phenology over the Earth surface, of its yearly variations and mid-term evolution, is a key element in the monitoring of climate change effects. Over last two decades, low resolution EO sensors such as MODIS Terra/Aqua, VGT on the SPOT then Proba-V missions, MERIS then now OLCI on the Sentinel-3 satellites have provided daily global acquisitions, which allowed to produce regular time series of vegetation parameters such as FAPAR or fCover, at global scale. However the corresponding spatial resolution is in general not compatible with the level of granularity of the Land Cover, except on very uniform landscapes, and this prevents separate characterisation of the phenology of the vegetation canopy types being mixed in low resolution pixels; disaggregation techniques have been applied but show their limit. Moreover, beyond the core description provided by the FAPAR and fCover parameters, a better understanding of the vegetation cycle in natural canopies would require as well a characterization of the dry (non-photosynthetic) part of this vegetation and its evolution along the season, and an insight into fresh foliage pigments composition and cycle.
Indeed we could access this level of detail with the Sentinel-2 (S2) data, thanks to both their spectral richness and spatial resolution. However obtaining regular time series of information from S2 data in order to feed phenology studies is indeed a challenging task, considering the global revisit performance achieved today with the two S2 satellites (5 days at Equator, with clouds).
The work presented here aims to bridge this gap. It is based on state-of-the-art techniques to best exploit each individual S2 image acquisition then on an innovative compositing approach to produce dense regular time series from irregular and partial observations.
The S2 images are processed from the L1C level using the Overland processor, a solution that performs both the atmospheric correction of the image and the generation of biophysical maps within a single step, using a coupled atmospheric and scene (vegetation) model. This produces altogether (1) surface reflectance, (2) atmospheric maps such as cloud optical depth, aerosol optical depth, and cast shadows, and (3) vegetation parameters such as FAPAR, fCover, fraction cover of non-photosynthetic vegetation (fNPV) and Canopy Chlorophyll Content (CCC). The role of the atmospheric maps, beyond their use in correcting reflectance down to bottom-of-atmosphere (BOA) level, is to serve as quality indicators for the further compositing steps. Cloud information is not simply used as a cloud mask but also to assess overall atmospheric transparency, combining cloud veils with aerosols. So for a given image, a pixel-wise compound quality weight is computed from the different quality indicators. Then to produce either the surface reflectance image or a vegetation parameter map at a given date, all S2 observations close to target date will be exploited, applying an additional weighting factor that qualifies the gap to target date. In this way, composite images or maps are generated by the weighted sum of all close observations. The range used around target date is fully adaptive, pixel-wise, and quality indicators, date span (range) and date shift, are generated together with the produced composite.
The whole process has been implemented on a Cloud-based infrastructure. The processing of yearly datasets of S2 source images can be launched in parallel, then all outputs are stored in the Geocube, a dedicated geo-spatial imagery database designed to store, get access and process time-series of spatial information. Geocube is an Open Software solution being developed with the support of CNES.
Results of this work will be presented, including striking videos produced over one year period in different regions of the world that help to catch, within a glimpse, all the variety of the vegetation dynamics; further steps and applications will be discussed.

Development of Earth observation data analytics allows to analyze and systematically extract diverse set of information from a variety of large datasets, including those observing and measuring the response of environmental and ecosystem processes. Vegetation indices are the basic information for phenological research, e.g. by analyzing the vegetation index time series to derive information, by assessing temporal statistics, or by estimating phenological metrics. The detection of land surface phenology at the landscape scale enables to investigate the spatio-temporal patterns of plant phenology and their relationship with environmental variability and climatic drivers. Along with the availability of high resolution satellite time series and proven methodologies to extract temporal information from satellite acquisitions, comes the need for procedures to generate high resolution Earth observation derived phenological metrics that could serve a wide range of applications, like monitor ecological status, environmental conditions, climate change impacts on ecosystems, cropland practices, and more accurately forecast crop yields. Additionally, vegetation phenology is highly sensitive to climate conditions and is a climate change fingerprint.
In this frame, here we present results from the applications of satellite derived high spatial resolution phenological metrics in the field of forest plant community mapping, crop mapping and for the identification of environmental factors driving plant phenology. The proposed procedure estimates phenological metrics using local curve fitting and local derivatives to identify phenophases, operating without thresholds or a priori information. The methodology temporally interpolates LAI time series, generated from Sentinel-2 MSI acquisitions, in order to homogenize seasonal trajectories of vegetation indices. LAI vegetation index estimated from satellites observations has been selected as source information to derive phenological metrics since it represents a biophysical parameter, and it is less affected by signal saturability in areas with high vegetation coverage. Removal of invalid pixels (e.g. clouds, topographic shadows) prior the analysis assumes a key role in reducing factors affecting the quality of observations, that typically results in reducing vegetation indices values. The selection of image dataset with rigorous cloud detection algorithm, the use of weighted smoothing procedure and the temporal image co-registration enhance vegetation indices time series integrity.
Error metrics calculated using PhenoCam ground observations resulted in a MAE of about 15 days for the various metrics analyzed, which is consistent with accuracy reported in other studies. Forest plant communities, together with their own typical floristic composition, show exclusive phenological dynamics recognizable by vegetation indices time series. Results show how both vegetation indices time series and land surface phenological metrics has been successfully used to classify plant communities and natural habitats for the Italian national territory. High spatial resolution smoothed time series and phenological metrics open up to the provision of novel temporal information about forest phenology anomalies, and useful monitoring system to scrutinize spatio-temporal patterns of forest disturbances, as demonstrated from the presented results of the investigation of a pest outbreak in central Italy. Moreover, detect phenological crop information has been used to map crop types for a small test area in central Italy. Finally, the investigated relations between selected phenological metrics and environmental factors are discussed.
This study highlights the importance of integrated data and methodologies to support the processes of vegetation recognition and monitoring activities. Earth observation derived phenological metrics represent a promising tool for a wide range of applications, as monitor terrestrial ecosystems, environmental conditions, and cropland practices.

Since 2011, the distribution extent of pelagic Sargassum algae has substantially increased and now covers the whole Tropical North Atlantic Ocean, with significant inter-annual variability. The ocean colour imagery has been used as the only alternative to monitor such a vast area. However, the detection is hampered by cloud masking, sunglint, coastal contamination and others phenomena. All together, they lead to false detections that cannot be discriminated with classic radiometric analysis, but may be overcome by considering the shape and the context of the detections. Here, we built a machine learning model based on spatial features to filter false detections. More specifically, Moderate-Resolution Imaging Spectroradiometer (MODIS, 1 km) data from Aqua and Terra satellites were used to generate daily map of Alternative Floating Algae Index (AFAI). Based on this radiometric index, Sargassum presence in the Tropical Atlantic North Ocean was inferred. For every Sargassum detections, five spatial indices were extracted for describing their shape and surrounding context and then used by a random forest binary classifier. Contextual features were most important in the classifier. Trained with a multi-annual (2016-2020) learning set, the classifier performs the filtering of daily false detections with an accuracy of 90\%. This leads to a reduction of detected Sargassum pixels of 50\% over the domain. The method provides reliable data while preserving high spatial and temporal resolutions (1 km, daily). The resulting distribution on 2016-2020 is consistent with the literature for seasonal and inter-annual fluctuations, with maximum coverage in 2018 and minimum in 2016. In particular, it retrieves the two areas of consolidation in the western and eastern part of the Tropical Atlantic Ocean associated with distinct temporal dynamics. This new dataset will be useful for understanding the drivers of Sargassum dynamics at fine and large scale and validate future models. For example, it may allow to track Sargassum aggregations over time and estimate Sargassum drift.

The marine environment supports a diverse range of sea life that is extremely important for global biodiversity and quality of life. Ocean spills have catastrophic consequences for marine and coastal environments, so rapid analysis and response is needed to monitor its evolution and minimise its environmental impact by controlling its spread. Satellite remote sensing has proven to be an effective tool for detecting and mapping marine pollutants in large areas, providing useful data for time-evolution modelling to track pollutants through space and time. Typically, optical sensors have been used to measure the impact of spills on soils and terrestrial vegetation, while microwave sensors, such as SAR, have monitored spills on the sea surface. Nevertheless, spill monitoring with optical sensors has improved in the last decade, partly driven by the proliferation of multi- and hyperspectral sensors on board satellites, which offer unprecedented spectral analysis of scenarios, and partly due to the development of artificial intelligence techniques. Optical sensors are a better choice than radar sensors for continuous monitoring due to their high repetition time.
There are several studies on high-precision marine pollution detection, both with optical and radar sensors, using machine learning techniques and neural networks. However, the mathematical simplicity of spectral indices makes them faster to process, an advantage especially when response time is crucial and when working with hyperspectral sensors with hundreds of bands. Spectral indices reduce the high dimensionality of the images by operating only on a reduced number of bands, thus being able to highlight a particular feature. However, most of the indices in the bibliography produce false positives due to sand in suspension on the coast which is identified as fuel in the sea. This is a major obstacle when spills occur in coastal areas, as it makes the initial study of the spill, its monitoring and its long-term impact more difficult. Thus, it is more complicated to measure the affected area and the thickness of the oil film at each point.
A novel spectral index for coastal oil spill detection (NDOI, Normalised Difference Oil Index) is proposed capable of identifying the spill without highlighting the suspended sand. To design the index, the spectral profile of different regions of interest (clouds, water, thin oil, thick oil and emulsion) from an AVIRIS image of the Deepwater Horizon catastrophe in the Gulf of Mexico is analysed. By looking for differences in the spectral behaviour of each region, bands of interest are selected highlighting the spill.
NDOI uses bands common to many optical satellites while staying in the VNIR range, facilitating its application on various platforms. The effectiveness of this index has been tested with sensors of different characteristics and a comparative study with other indexes in the literature and radar images validate it. Moreover, NDOI has been used to classify the pixels of several images using the kNN classifier, and an accuracy close to 95% has been obtained with an F1-score for the spill class of more than 80%. In addition, it has been shown to be able to distinguish between medium and high oil thickness.

Especially for the Baltic Sea, there is concern that cyanobacteria will benefit from climate change in decades to come (Karlberg and Wulff, 2013; Paerl and Otten, 2013; Visser et al., 2016). As one of the largest inland seas, the Baltic Sea is of great importance and acts in many ways as a burning glass for problems occurring in global oceans.
Here, the pressure concerning cyanobacteria is exceptionally high due to the following boundary conditions: Temperature have risen in the Baltic Sea relatively strong compared to the global ocean, while the growth and the nitrogen fixation of cyanobacteria are generally favoured by high temperatures. The increased stratification and the lowered viscosity of seawater due to higher temperatures might favour buoyant species, like specific cyanobacteria occurring in the Baltic Sea. Also the high levels of phosphor, as the second important limiting plant nutrient, promote the abundance of cyanobacteria, which can fix atmospheric nitrogen.
Additionally, disadvantage through an increase of cyanobacteria in the Baltic Sea is also exceptionally high: The Baltic Sea is already suffering from eutrophication and great effort has been made in order to reduce the input of nutrients and protect biodiversity. Cyanobacteria somehow counteract this effort, because of their excess nitrogen fixation and contribution to the formation and increase of dead zones.
A synoptic view on phytoplankton, with sufficient spatial and temporal coverage, is only possible from space. Phytoplankton is observed from Space through its main pigment chlorophyll-a by optical imagers.
The chlorophyll-a concentration is customarily derived in a two-step approach from the above water observations. The primary step, atmospheric correction, is sensor specific and often fails in complex waters. Especially in the Baltic Sea, with its high CDOM values, standard algorithm often fail, also due to a diminishing signal in the blue-green wavelength region.

We have developed a simple and fast algorithm for the retrieval of chlorophyll fluorescence (Fluorescence Peak height, FPH) and chlorophyll absorption (Absorption Peak Depth, APD) based on Level-1 radiances (with no need of atmospheric correction). Each of the two parameter and their combination are sensitive to chlorophyll concentration, primary production, functional types and layering of the algae. E.g. the difference of absorption and fluorescence is an indicator of cyanobacteria. This indicator allows the detection of surface blooms, also where the standard chl-a processors fail (The image shows the Gulf of Finland on the 12th of July, 2018, when it was covered with cyanobacteria. Row 2,3,4 show results of our retrieval, row 5 and 6 show OLCI L-2 results). As we have shown, chlorophyll fluorescence and chlorophyll absorption can be derived consistently from OLCI and its predecessor MERIS, providing global coverage and a total time window of nearly 20 years. Here, we show the potential of FPH and APD for the detection of cyanobacteria blooms in the Baltic Sea by comparing our results to insitu measurements.

Understanding ocean dynamics is vital to interpreting marine ecosystems' functioning and determining key processes affecting global climate and biodiversity. Changes in chemical composition, ocean warming, loss of biodiversity and several climate interactions can alter the dynamic equilibrium between ocean, land and atmosphere and between biotic and abiotic components in the Earth System. Such interactions can operate across multiple temporal and spatial scales and generate extreme conditions that impact ocean dynamics and ecosystem functioning.

Marine heatwave (MHW) is one such extreme climatic event that can have devastating impacts on the oceanic ecosystem. Irregular, abrupt, but persistent changes in sea surface temperature (SST) have been noticed in recent years, with cascading effects on different components of marine ecosystems. Recent research has shown that the rapid increase in MHW frequency and duration trends in the last four decades directly affects global climate change. Therefore, it has become crucial to understand the behaviour of marine heatwaves in the spatio-temporal domain, which can unveil the future influence of climate change on marine ecosystems. We present a method to disentangle the relationships between different ocean properties and the temporal dynamics and spatial patterns of MHW in the global ocean. Deep learning architectures are used to analyze these relationships and provide a method to derive early warming indicators of future MHWs.

As an initial hypothesis, we have considered the time, location, duration and intensity of both extreme warm and cold events as the primary features to develop a graph-based deep learning approach, Graph Neural Networks (GNN). This network can share information across nodes through edges and understand the importance of edge connections between nodes. The idea is to consider MHWs and cold events as nodes for the graph, where primary features of the MHWs are defined as the feature vector for each node. Since there is no prior knowledge to determine the relationship between MHWs (typically are edges in graphs), an attention mechanism is being used to calculate a pairwise correlation between nodes and helps consider only the virtual nodes for the knowledge base of edges.

Eventually, the model outputs the final weight of the edges, which defines the relationship between existing MHWs and cold events. This output will help us understand the correlation of the rise in temperature between different locations and periods and can provide the knowledge to predict future dynamics of MHWs.

Coral reefs protect coastlines from storms and erosion, support tourism, provide habitat for fishes, and consequently replenish fish stocks. Particularly, it has been estimated that over half a billion people depend on reefs for food, income, and protection [1]. The structure of a coral reef indicates an underwater ecosystem characterized by diverse corals colonies forming a three-dimensional and visible structure. Coral reef structure have been studied and classified at various scales. At every scale, the object of focus is different. At microscale, a single polyp is studied at microscale, whereas a coral colony is considered when the object of focus is classified by species or shape [2]. Further, the zones within a coral reef structure are considered when these are classified based on benthic or geomorphic properties, and finally a group of coral reef formations are considered when the reef region is classified based on geological formation (e.g., atoll, fringing reef, barrier reef), at macroscale.

Up to now, limited research on mesoscale has been published. At mesoscale, the object of focus is a group of coral colonies or a single coral reef formation. These structures are often found in isolated areas of the ocean and can be studied using in-situ data. However, the in-situ inspections are challenging in terms of logistics, availability of trained personnel, and funding. Remote sensing techniques overcome these problems and show the potential for gaining knowledge on the coral reef structures worldwide [3,4].

In this study we investigate the radar and optical images and environmental products from the ESA’s Sentinel- Missions (i.e., Sentinel-1/2/3) to relate coral reef structure shapes to the Sentinel-derived parameters. The coral reef structure shapes are obtained from the Allen Coral Atlas [5], where worldwide coral reef structures are mapped. From the Allen Coral Atlas bathymetric and benthic maps of 5-meter resolution, the data is aggregated to fit the size of a reef structure and is transferred to a binary (reef/no reef) map where the shape of the coral reef structure can be extracted. Every shape is parametrized with geometry-based descriptors and is delineated by a polygon with associated geographical coordinates.

In a next step the Sentinel-Data are trained on a local area of interest where the conditions are very well known. Then a simple machine learning algorithm is applied to extend the area of interest to a wider region where a correlation with the Sentinel-derived parameters is expected. The main goal of this study is to predict a general coral reef structure shape for a previously unseen region of the earth using the trained machine learning algorithm and high-resolution satellite data.

The study is focusing on two parts: Firstly, to understand whether there are recurring patterns in the shape of the coral reef structure in a given region of interest. Secondly, to discover relationships between the shape of the coral reef structure (extracted from a reef/no reef classification map) and the environmental and physical variables acting on it (extracted from high-resolution satellite data).

Reference

[1] The Prince of Wales’ International Sustainability Unit (ISU) and United Nations Environment Programme and International Coral Reef Initiative (ICRI). The Coral Reef Economy: The business case for investment in the protection, preservation and enhancement of coral reef health. Tech. rep. 2018, pp. 36.
[2] Zawada, K. J., Dornelas, M., & Madin, J. S. (2019). Quantifying coral morphology. Coral Reefs, 38(6), 1281-1292.
[3] Hedley, J. D., Roelfsema, C. M., Chollett, I., Harborne, A. R., Heron, S. F., Weeks, S., ... & Mumby, P. J. (2016). Remote sensing of coral reefs for monitoring and management: a review. Remote Sensing, 8(2), 118.
[4] Purkis, S. J., Kohler, K. E., Riegl, B. M., & Rohmann, S. O. (2007). The statistics of natural shapes in modern coral reef landscapes. The Journal of Geology, 115(5), 493-508.
[5] B. Lyons, M., M. Roelfsema, C., V. Kennedy, E., M. Kovacs, E., Borrego‐Acevedo, R., Markey, K., ... & J. Murray, N. (2020). Mapping the world's coral reefs using a global multiscale earth observation framework. Remote Sensing in Ecology and Conservation, 6(4), 557-568.

According to research published in Frontiers in Ecology and the Environment, seagrass ecosystems globally consume carbon up to 35 times faster than rainforests, and lock this carbon into seabed sediments. In addition, having an abundance of seagrass in our coastal waters encourages biodiverse marine habitats in otherwise barren environments, acting as a nursery for young fish and other marine animals to create a rich and diverse ecosystem. Despite this, they are increasingly under threat from multiple sources including coastal development, pollution, and damage from human activity in the marine environment.

With seagrass ecosystems in decline, conservationists need a rapid understanding of the distribution, health, and coverage of seagrass over time. In the UK, current techniques require in-situ mapping or manual inspection of drone imagery, a time consuming and laborious task. However, it is through the increasing availability of Earth Observation (EO) data combined the processing power of cloud computing that the opportunity arises to develop new services that facilitate the rapid planning of seagrass conservation activities and re-planting; policy planning in coastal areas to inhibit activities causing seagrass decline; and more targeted research into these precious marine environments.

The development of such a service has been the product of ongoing collaboration between CGI and the environmental charity Project Seagrass to characterise seagrass distribution in UK coastal waters. The service is supported by CGI’s state-of-the-art cloud-based data processing platform, GeoData360, to provide reliable and repeatable processing across all 31,368km of UK coastline. The platform provides systematic access to Copernicus Sentinel data, to which it applies atmospheric corrections and a CGI developed seagrass mapping algorithm. The service is run in parallel across hundreds of Sentinel-2 products to generate data across the whole UK, reliably and continuously monitoring vast lengths of coastline. GeoData360 publishes this data for easy access to organisations such as Project Seagrass enabling them to focus their resources and feed the data into downstream applications, either through pre-planning or use in the field.

The service itself generates two data products from the Sentinel-2 inputs, preserving the spatial resolution of this data at 10m. These products are a Seagrass Percentage Coverage layer to indicate the spatial distribution of intertidal seagrass, and a Seagrass Biomass layer to help with estimates of carbon sequestration. In conjunction with a characterisation of the seagrass spectral response, a habitat suitability model is applied remove the coastal areas that provide inadequate conditions for seagrass growth and therefore help to reduce commission errors.

The algorithm continues to evolve in finding new and more reliable ways of detecting and mapping seagrass ecosystems and hence reduce problematic areas of commission, particularly with respect to detecting the similar spectral response of macroalgae. Future areas of investigation for algorithm improvement include spatio-temporal synthesis of Sentinel products to dynamically map intertidal zones, and investigating statistical methods of classification for more sophisticated seagrass detection. A platform such as GeoData360 provides the framework, data and processing power to test and deploy experimental services and rapidly converge on a solution that is performant at large scale.

The Copernicus Marine service (CMEMS) in operation today routinely delivers information about the Green Ocean based on satellite and in situ data combined with numerical models. The aim is to provide users with “best estimate” representations of the state of marine ecosystems and biogeochemical indicators of interest. A key strategic evolution at Copernicus 2 horizon will be to consolidate the service with more robust information about product uncertainties, whether in real time, in delayed mode (reanalyses) and in forecast mode with a few days of lead time. In that perspective, the transition to probabilistic analysis and prediction methodologies is a necessary step, e.g. to provide more actionable information to help in decision-making and management of marine ecosystems.

In the frame of the H2020 SEAMLESS project, ensemble generation methods are being developed with the aim to improve the service through better data assimilation / inversion methods. A stochastic version of the NEMO-PISCES model has been developed and implemented in a global ocean configuration at ¼° inherited from the CMEMS global Monitoring and Forecasting Centre. The stochastic NEMO-PISCES coupled model is based on parameterizations introduced by Garnier et al. (2016) and further used by Santana-Falcon et al. (2020) to assimilate satellite ocean colour in the North Atlantic.

The generation of a 40-member ensemble has been produced in the global configuration using a similar setup. The assumed uncertainty sources originate from (i) 7 critical biogeochemical model parameters of the PISCES formulation; (ii) sub-grid scale effects associated to the ¼°, eddy-permitting resolution, and (iii) the location of ocean mesoscale structures and associated advective/diffusive fluxes (Garnier et al., 2016; Leroux et al., 2021). Before activating the stochastic model, a spin-up has been produced using the deterministic NEMO-PISCES model run from 01.01.2017 to 22.12.2018 and initial conditions from a BIOMER run. The stochastic ensemble simulation was then initialized on 22 December 2018, and further integrated to cover the full 2019 year using unperturbed ERA5 atmospheric forcing. The resulting 40-member ensemble represents a probabilistic view of the 2019 seasonal cycle in the global and North Atlantic ocean.

The computational burden of the ensemble production requires a large CPU/storage capacity and an adapted strategy for diagnostics. The ensemble is analysed in terms of spread, median, min and max distributions (i) at basin-scale, with a focus on the North Atlantic sector using ensemble outputs every 5 days ; (ii) in 4 selected regional sectors (Western Mediterranean, Bermuda, North-East Atlantic, Eastern upwelling system) using daily ensemble outputs ; in 1D reference sites (BATS, BOUSSOLE). The analysis is done on model state variables related to surface chlorophyll concentration, as well as on a variety of targeted indicators (e.g. NPP, phenology, trophic efficiency).

In order to evaluate the relevance of the ensemble pdfs with respect to observed data, verification statistics have been produced using the EnsScores library (Brasseur et al., 2021) to check the consistency against satellite products. In particular, we use daily L4 ocean colour products from the CMEMS catalogue which combine measurements from Sentinel-3 satellites. The computed metrics include rank histograms, CRPS (decomposed into reliability and resolution skill scores) and RCRV.

In this poster, we will present a synthesis of the ensemble scores obtained in the different regions, highlighting situations where the prior ensemble is consistent with uncertainty hypotheses made in the stochastic NEMO-PISCES model. Further, we will show how to take into account irreducible uncertainties in the verification data products to compute the scores. We will discuss the sensitivity of the computed metrics against these uncertainties, underlying the importance of properly accounting for error propagation in the CMEMS TAC production chains. We will finally explore a new 4D Bayesian inversion scheme aimed at delivering probabilistic analyses and predictions with a few days of lead time.

References
Brasseur P., Brankart J.M., Popov M, (2021). D3.1 Code for ensemble evaluation in prototype. Deliverable report of project H2020 SEAMLESS (grant No 101004032.), https://doi.org/10.5281/zenodo.5554808
Garnier F., Brankart J.-M., Brasseur P. and Cosme E., 2016: Stochastic parameterizations of biogeochemical uncertainties in a 1/4° NEMO/PISCES model for probabilistic comparisons with ocean color data, J. Mar. Systems, 155, 59-72, https://doi.org/10.1016/j.jmarsys.2015.10.012
Santana-Falcon Y., Brasseur P., Brankart J.M. and Garnier F., 2020: Assimilation of chlorophyll data into a stochastic ensemble simulation for the North Atlantic ocean, Ocean Sci., 16, 1297–1315, 2020, https://doi.org/10.5194/os-16-1297-2020
Leroux S., Brankart J.M., Albert A., Molines J.M., Brodeau L., Penduff P., Lesommer J. and Brasseur P. (2021). D7.2 Ensemble quantification of short-term predictability of ocean fine-scale dynamics. Deliverable report of project H2020 IMMERSE (grant No 821926), https://doi.org/10.5281/zenodo.4570158

TechWorks Marine are working on a project, funded by the European Space Agency (ESA), to investigate if high Escherichia coli (E. coli) levels in coastal waters can be detected and predicted via a combination of Earth Observation (EO) data, in-situ measurements, and machine learning models. In the long term, this activity aims to provide actionable, quantified, and timely information in the form of an E. coli Alert Data Service for clients (e.g., environmental agencies and local authorities) on the likelihood of E. coli contamination in coastal waters. On the timescale of this activity, the aim is to determine the viability of such a service and to establish a framework for the development of a commercial service.

E. coli infection can cause serious symptoms, particularly in children. Bathing water users near locations where wastewater is discharged into the sea are at risk. In the case of Ireland, city and county councils are responsible for alerting local communities when an increase in E. coli is measured locally and are obliged to restrict access to local bathing areas when the water has been contaminated. However, current measurements are made, at best, on a weekly basis at a limited number of locations and it takes approximately 48 hours to obtain the results from an individual measurement, during which the local community may be at risk. This activity aims to develop a service to alert clients of potential high risk E. coli events, particularly during summer months, enabling timely and pre-emptive action to be taken.

To achieve this objective, TechWorks Marine are examining historical and current data to determine if relationships can be established between in-situ measurements of E. coli, coincident measurements of a range of biophysical ocean parameters (e.g., salinity, pH, chlorophyll-a), and multi-sensor satellite observations (e.g., optical, SAR, and thermal) of wastewater release events (e.g., from wastewater treatment plants). This multi-sensor approach will increase the accuracy of determining biophysical signatures for wastewater events, but also mitigate common issues encountered when using a single sensor approach.

Customised EO processing chains are being developed to derive biophysical parameters (e.g., turbidity, chlorophyll-a, CDOM) from satellite imagery of wastewater events. A data fusion approach is being developed to model E. coli concentrations and investigate the dynamics and relationships of E. coli levels with environmental stressors. A range of machine learning algorithms, such as gradient boosting, random forests, and recurrent neural networks are being evaluated, along with different feature and hyperparameter combinations, for their ability to model E. coli concentrations.

In addition to the technical objectives, the project aims to demonstrate the importance of routine coastal water quality monitoring and highlight the advantages of utilising satellite EO data and machine learning techniques for both environmental and societal protection.

Rivers have been identified as a major input pathway for plastic debris entering the ocean. However, much uncertainty remains over the spatial and temporal distribution of plastic pollution and the scale of fluxes entering the marine environment from river systems. As part of the SIMPLER (SensIng Marine Plastic Litter using Earth observation in River outflows) project, this work aims to develop and validate an Earth Observation (EO) approach to quantify the microplastic flux rates from rivers into coastal waters. The current methods available to quantify and monitor microplastic fluxes are based on localised, time and resource intensive in-situ sampling, laboratory analysis and modelling approaches. Previous EO approaches to monitor plastic pollution have focused primarily on floating macroplastic aggregation detection in the open ocean and coastal waters. Microplastic concentrations are often not sufficiently high to change the water surface optical reflectance signal, and even at high concentrations the strong absorbance of near-infrared and shortwave-infrared light in the water’s surface precludes the detection of a microplastic signal. River mouths present additional challenges because the waters are often highly turbid and optically complex, making it more difficult to accurately detect the concentration of water constituent components. In light of these challenges, one possible approach to quantify microplastic fluxes from rivers is to use an indirect detection method where water constituent concentrations derived from medium resolution optical satellite imagery are used as proxies to estimate microplastic concentrations. This study will explore how satellite-derived water quality metrics, such as suspended particulate matter (SPM), compare to in situ measurements of microplastic concentration (200µm>D>5mm) in the river Tamar, UK. This relationship will be used to map microplastic proxy concentration and, with river gauge data, a microplastic flux estimate for this river system will be calculated. The methodology for the microplastic flux quantification will be presented with the assumptions it makes and the advantages and limitations of the approach will be discussed.

Coastal areas are ecosystems of significant importance and sensitivity and play a very crucial role in the economy (i.e. tourism, fisheries and aquaculture). However, their importance is often overlooked, as being overexploited or subjected to intense environmental pressures. It is a fact that almost every year many coastal ecosystems are threatened by oil spills or petroleum products. The sources of pollution associated with these spills are often of natural origins or human-induced activities, e.g. land transport of oil from tankers, trucks and pipelines (oil from accidents and leaks that can end up in coastal environments) as well as illegal ship operations and offshore oil rigs (Alpers et al. 2016, Konstianoy et al., 2018).

Spills of oil and petroleum products in the coastal environment can have serious environmental and economic consequences (Atlas and Hazen 2011, Perrons, 2013). Therefore, the necessity of environmental monitoring of these leaks is indisputable as it is of great importance for public safety and environmental protection. Detection of oil spills is also important for estimating the potential spread and movement from the source to adjacent coastal areas (Mishra and Balagi, 2016). Legislative frameworks, policies and various monitoring systems have been developed in many countries to prevent and deal with oil spills. However, in most of the cases this does not involve exploitation of satellite imagery.

Appropriate monitoring techniques can facilitate leak recovery by providing timely oil spill detection as well as oil characteristics/properties, estimating/calculating the size of an oil spill and predicting its displacement (Robbe and Hengstermann, 2006). The main advantage of satellite remote sensing is that it provides both large and small scale monitoring (Dassenakis et al. 2011) while allowing 24-hour coverage, making it easier for decision makers to respond promptly and take appropriate action. In particular, satellites equipped with Synthetic Aperture Radar (SAR) instruments are key-tools for monitoring oil spills - as they provide data regardless of the time of day and weather conditions.
In this work, the detection of oil spills was carried out for the monitoring of the coastal environment of the Prinos - Kavala basin, North Aegean Sea, Greece, using satellite data and Geographic Information Systems. The area of the Prinos - Kavala basin, is the only area in Greece where offshore oil production activity takes place. Moreover, around the coastal areas there is intense touristic and fishing activity, while at the same time it consists of a set of protected areas. Copernicus Sentinel-1 satellite data was utilized and a semi-automatic methodology was developed for the detection of oil using the open-source ESA SNAP toolbox, while post-processing combining auxiliary geospatial datasets (coastal water surfaces, faults, Natura 2000 sites etc.) aimed for the generation of analysis ready maps to support decision making.
This study was funded by Hellenic Petroleum S.A. through Aristotle University (AUTh) Research Committee.

In recent years, the sea-level rise caused by climate change has resulted in inundation of low-lying coastal areas. India is among the top ten countries affected by natural disasters, particularly those caused by hydrological and meteorological factors (Thattai D V et al., 2017). Indian coastlines are also susceptible to recurring storm surges and localized unforeseen flash floods. People living in low-lying areas of the country are exposed to disease transmission caused by contamination of water bodies connected with natural catastrophes. To protect and control developmental activities in the coastal areas, the government of India has categorized coastlines under the coastal regulation zone act notification, 1991. In spite of such laws, the alteration of coastal zones is continuing throughout the country. Andhra Pradesh, an Indian coastal state along the Bay of Bengal is severely impacted by regular cyclones. According to the state cyclone warning center, 54 cyclones have affected and damaged the coastal districts of Andhra Pradesh since 1971, with some of the districts bearing the brunt of the damages. The coastal region of the state is densely populated with around 34.2 million people (2011 Census) and natural disasters cause significant disturbance to people’s lives. With an estimated area of 404 sq km (ISFR,2019), mangroves in Andhra Pradesh contribute to coastline protection in the state. The growth of industry, aquaculture, and salt pans near mangrove areas, release of domestic and industrial pollutants, and the felling of mangrove forests, have all led to considerable destruction of these delicate ecosystems. The state’s mangrove forest once covered 582 sq km (Ramasubramanian et al., 2004), but it has fallen to 404 sq km in 2019.

In this study, we have mapped mangroves of Andhra Pradesh using the multispectral images of Sentinel 2 satellites from the European Space Agency. The mangrove patches in the state were delineated using Random Forest classification algorithm in the Sentinel Applications Platform (SNAP, Version 8.0) and a mangrove atlas for Andhra Pradesh was developed. About 41 mangrove patches were identified along the eight coastal districts. The classification accuracy was estimated with the ground control points (GCPs) collected through extensive field surveys in all the 41 mangrove patches in the state and from the high-resolution Google Earth image software. Approximately one-third of the reference data (a total of 150 GCPs), homogeneously distributed, were utilized to assess the accuracy. The classified images were subjected to an accuracy assessment by measuring the Cohen’s Kappa coefficient (K), and we obtained a value of K = 0.87.
The total coastal length of the state is approximately 975 km and it was found that 253 km of the coastline is protected by the 41 dense mangrove patches. The loss of mangroves in the state increases risk of flooding and inundation of the coastal areas, and mangrove restoration is vital in vulnerable areas. Protection from natural hazards, and reducing the risk of epidemic diseases associated with cyclones and floods in the state can only be achieved by regular monitoring and restoring the coastal structures that help to combat these natural disasters. Mangroves are one of the blue carbon ecosystems on the coast, and the vast areas of mangroves in the state would help to reduce the effect of anthropogenic carbon-dioxide released into the atmosphere by sequestering them. Additionally, estimating and understanding the coverage and health of mangroves in the state would help to achieve the UNFCCC's "REDD+" strategies and "blue carbon" mission.

Keywords: Mangroves, Water quality, Conseravation, Mapping, Sentinel-2

References:
ISFR. (2019). India state of forest report.
Ramasubramanian, R., & Ravishankar, T. (2004). Mangrove Forest Restoration in Andhra Pradesh, India. MS Swaminathan Research Foundation, Chennai, MSSRF/MA/04/13, pp8.
Thattai, D. V., Sathyanathan, R., Dinesh, R., & Kumar, L. H. (2017, July). Natural disaster management in India with focus on floods and cyclones. In IOP conference series: earth and environmental science (Vol. 80, No. 1, p. 012054). IOP Publishing.

The occurrence of Ostreopsis. ovata (O. ovata) Harmful Algal Blooms (HAB) has increased during the last years, also in the coasts of Ligurian Sea. Exposure to the toxins produced by O. ovata can cause bio-intoxication in humans, and also suffering or mortality in marine benthic communities (Accoroni at al., 2016). O. ovata can be harmful threating coastal food web and fisheries. The cause of the increase of bloom occurrence during the last years in temperate areas such as the Mediterranean Sea is subject to multiple interpretations, in some cases controversial, invoking climate change and pollution. Understanding the environmental conditions which can trigger the bloom, can therefore be important as an aid to protect marine biodiversity but for the tourism sector.
The most widely used techniques for algal-bloom detection are based on standard data from ocean color satellites and concern extensive algal blooms, covering sea area large enough to be directly observed by space instruments with the standard ocean color sensors, (Blondeau – Patissier et al., 2014). These techniques, as they rely on satellites not suitable for coastal observations, so they are inapplicable to the monitoring of O. Ovata blooms, which usually occur in small areas close to the coastline. In some recent works, the use of those satellites data as proxy to observe phenomena occurring at a small scale and very close to the coast, was explored. Specifically, the environmental large scale drivers most relevant for O. ovata growth and bloom, are estimated by the available proxies. However, as for satellite data use in coastal areas, and for small scale phenomena such as algal bloom detection, it is now possible to test ESA Sentinel 3 OLCI ocean color data, as the ocean color measurements performed by this satellite have the main focus on coastal zone monitoring (Sentinel-3 OLCI Marine User Handbook, 2021).
Understanding the role of coastal sea drivers in biological phenomena is, however, complicated, because of the complexity of the relationships between the physical, biogeochemical and biologic parameters. The environmental and ecological drivers involved in O. ovata growth and bloom dynamics are modulated by complex interactions among biotic (intra and inter-specific relationships) and abiotic (meteo-marine) drivers, and the chain of factors triggering the bloom is still object of research (Asnaghi et al., 2012). Although O. ovata bloom is a complex phenomenon, our hypothesis is that specific meteo-marine factors at large scale, obtainable by satellite and model, are able to explain a significant part of the phenomenon variability. In particular, the possibility of monitoring O. ovata HAB triggering conditions by satellite and model data could help to better and sustainably manage the environment hosting this species.
For this reason, this study starts with a short review of the most important factors involved in O. ovata bloom triggering dynamics, and we hypothesize that a reduced number of meteo-marine factors, namely water temperature and remixing, are able to explain a significant part of the phenomenon variability. Satellite and model data obtained from Copernicus service, namely Sea Surface Temperature (SST) and Significant Wave Height (SWH), were tested as proxies of these local factors.
Statistical analyses were first conducted on the relationships between SST and SWH and O. ovata blooms cumulated over relatively long time periods (one to few months). The relationships found confirm the importance of the drivers in regulating the bloom, but are of limited utility to predict the phenomenon.
A simplified conceptual model to predict O. ovata occurrence based on these products and derived also from previous works but calibrated with local measurements is then developed and tested using data taken in two study areas along the Tuscany coast (Central Italy) over an eight-year period (2012-2019). This model, corroborated by the results of the statistical analyses, was finally applied to yield a synthetic indicator informative on O. ovata abundance. Although O. ovata growth and bloom are expected to be sensitive to a large number of drivers, the indicator is able to account for 35% of their variance for the two study areas, with the exception of a single case. The analyses indicate that the use of SST from satellite and SWH from local models is effective in explaining the main behaviour of O. ovata in two study sites in the Ligurian Sea, according to a simplified conceptual model.
We are also exploring the possibility of using ocean color data, as the new Sentinel 3 OLCI sensor is designed also for coastal observations, and compared it with a Copernicus multisatellite product. In particular, the trends of various ocean color products (chlorophyll a, turbidity) are being evaluated and used to detect favorable conditions for O. ovata bloom.

References:
Accoroni S., Totti C., The toxic benthic dinoflagellates of the genus Ostreopsis in temperate areas: a review, Advances in Oceanography and Limnology, 7(1) (2016); pp. 1-15
Asnaghi V., Bertolotto B., Giussani V., Mangialajo L.., Hewitt J.., Thrush S., Moretto P..,. Castellano M, Rossi A., Povero. P., Cattaneo-Vietti R., Chiantore M., Interannual variability in Ostreopsis ovata bloom dynamic along Genoa coast (North-western Mediterranean): a preliminary modeling approach, Cryptogamie, Algologie, 33 (2) (2012), pp. 181-189
Blondeau-Patissier D, Gower JFR, Dekker AG, Phinn SR, Brando VE, A review of ocean color remote sensing methods and statistical techniques for the detection, mapping and analysis of phytoplankton blooms in coastal and open oceans, Progress in Oceanography, 123, (2014), pp.23-144
Sentinel-3 OLCI Marine User Handbook, EUM/OPS-SEN3/MAN/17/907205, v2G e-signed, 12 March 2021

It has been estimated that nearly half of the pollution at sea caused by crude oil (and other refined products) results from international maritime traffic. In the last decade there were 63 oil spills of greater than 7 tonnes, resulting in 196,000 tonnes of oil lost into our oceans. We are all familiar with the images from large oil spills, such as the more recent disaster in Mauritius in 2020. Large oil spill disasters garner great media attention and highlight to the public the devastation caused to our ocean and coastal eco-system. However, it is estimated that large oil spills account for only 10% of global marine oil pollution. In fact, approximately 35% comes from more discreet tanker traffic and other shipping operations. 45% comes from industrial effluents, atmosphere and drilling rigs. [World Ocean Review 4, Maribus 2010] This large and mostly hidden contribution to marine pollution is something Earth Observation can play a key role to tackle.

Traditional methods that agencies have used to detect and monitor pollution events at sea primarily involve aerial imagery and video capture from helicopters or light aircraft. Whilst aerial imagery and video capture using this application is an excellent way of monitoring and tracking the evolution of oil spills, it has many disadvantages. Obtaining worthwhile results requires detailed preparation before take-off and careful interpretation of the information gathered. In addition, whilst spot checks are possible across wider regions using this method, it is expensive and practically very difficult to monitor large expanses of the ocean, undoubtedly not being effective enough to discourage illegal oil dumping activities. Its application is more suited to deployment of already known events. In comparison, EO data offers significant opportunity to support persistent monitoring of the oceans due to its frequent revisit time and wide ranging coverage and would prove a valuable tool to authorities and a deterrent to those carrying out these activities. More specifically, Synthetic Aperture Radar (SAR) Imagery provides the necessary characteristics to identify oil pollution thanks to its ability to operate independently from daylight and cloud coverage.

Obtaining the imagery is just one element of providing an effective solution. Interpretation of imagery is another one; currently it is time consuming and impractical over a large area with numerous images available per day. Therefore an automated approach to oil pollution detection is necessary. Airbus Connected Intelligence UK is developing a method of automating this detection process, to deploy at scale and provide alerts and reports to agencies and authorities enabling action to be taken. The method is based on a U-Net convolutional neural network approach. Any such approach is reliant on the availability of good quality training data. Fortunately, we have been able to leverage historical data we have available from our GlobalSeeps® service, a wold leading oil seepage information product used by the oil exploration industry. It is a database of offshore oil slicks identified by trained analysts from SAR Imagery. Whilst focussing on natural oil slicks for customers in the oil industry, it also contains a number of man-made slicks, delineated by a trained analyst. In addition to this, we are curating a training dataset of more specific examples of oil pollution, to train the algorithm to detect man made slicks and not just (or even at all) natural ones.

Whilst our solution aims to tackle specifically the pollution ranging from the small, discreet and common crime of discharging oily waste into the ocean, there is reason to expect the algorithm can be re-trained to detect a variety of types of oil pollution.

This concept works towards the United Nation Sustainable Development Goals 14 (SDG 14); aiming to prevent and reduce marine pollution and instilling the sustainable management and protection of marine and coastal ecosystems.

Preliminary outcomes have shown that the proposed solution can identify oil slicks in SAR imagery. Using Sentinel-1 imagery and training data sourced from just the GlobalSeeps® database, a simplistic U-Net model has been trained as a proof-of-concept. Figure 1 shows a selection of validation chips reserved to test the model performance. The initial results show that an automated delineation of an oil slick using deep learning is achievable. We look forward to presenting in the conference further results from additional developments towards a fully automated solution integrating a model which has been trained on more data with more scenario variance for better real-world performance.

Figure 1: Validation Sentinel-1 image chip examples of the U-Net model output which has been trained on Sentinel-1 imagery. From left to right: SAR image chip (seep image); manual delineation (Mask), U-Net model output (Segmentation); and binary threshold output (Threshold = 0.5)

Compound events, i.e., the co-occurrence of multiple extreme events such as marine heat waves (MHW) in combination with other physical and biogeochemical anomalies, can have a strong impact on marine ecosystems and their functioning, with severe cascading effects that propagate throughout the oceanic food web. Compound events are generally associated to multiple natural (e.g., ENSO) and anthropogenic causes (e.g., ocean warming, deoxygenation), which interact at different spatial and temporal scales. Yet, key factors that can lead to the emergence of compound events are not well known. Within the project “deteCtion and threAts of maRinE Heat waves – CAREHeat”, in the framework of the the ESA’s Ocean Health initiative, we exploit a multitude of physical (e.g., SST) and biogeochemical data (eg. Productivity, Nutrient fluxes) to:
• Identify and characterize the co-occurrence of MHW associated with other physical (e.g., stratification) and biogeochemical (eg: ocean deoxygenation, nutrient decline, ocean acidification) features.
• Provide frequency and intensity maps of the regional distribution of compound events
• Assess the impact of these compound events on marine primary and secondary productivity
• Identify the key drivers of the identified compound events.
Here we will present the proposed method, which will allow to improve our understanding on the factors triggering compound events and on their impact on marine ecosystems.

Different sources of sea level data often refer to various vertical reference surfaces (e.g., a (quasi-)geoid, the mean sea level, or another tidal datum). This lack of consistency hampers proper understanding and management of marine areas (e.g., marine spatial planning and engineering, impact of climate change). An important source of sea level data is provided by hydrodynamic models (HDM). HDMs offer sea level data at high spatiotemporal resolution. The models typically lack, however, a well-defined vertical reference. As such it is not straightforward to express the model-derived sea levels in a 3D coordinate system. Often, the reference surface is believed to be coinciding with the mean sea level (MSL). However, since the model dynamics assume zero horizontal gravity components this interpretation is not correct. In fact, the reference surface coincides with an equipotential surface of the Earth's gravity field. This implies that the model-derived water levels can be reduced to the geoid by applying a bias correction. In this study we aim to compute this bias by means of geoid-referenced tide gauge data. After that, we will compare the vertically referenced HDM output to both tide gauge and altimeter data.
Our case study area is the Baltic Sea (northern Europe) where a high-resolution geoid model (NKG2015) is available, as well as a dense network of geoid-referenced TGs (BSCD2000). In this study, 73 TG stations from nine countries along the Baltic Sea coastline were used to shift the reference surface of the HDM down as much as the median of the difference between HDM and TG records at the stations. The vertical reference bias of the HDM with respect to geoid referenced TGs in the same permanent tide system (mean-tide system) is roughly 10.8 cm with a standard deviation of 2.5 cm for the period of January 2017 to June 2019. After reducing the zero level of the HDM to the geoid, we show the performance of the HDM to simulate the dynamic topography by a frequency analysis at the location of the TGs. Moreover, the offshore model dynamic topography is examined by utilizing the ESA funded Baltic+ SEAL project Sentinel-3A along-track altimetry data (that has been specifically adapted for coastal and sea ice conditions). The results also demonstrate the temporal and spatial discrepancies and inconsistencies among the data sets. For instance, a large discrepancy between HDM and SA data in the eastern part of the Gulf of Finland is observed that probably points to problems in the geoid model; in the area we lack sufficient data to compute and validate the geoid.

The Copernicus Polar Ice and Snow Topography Altimeter (CRISTAL) mission, planned to be launched in 2027 will incorporate a dual Ku/Ka band interferometric altimeter with specific transmission pulse sequences designed to enhance the performances over sea and land ice. The open burst mode will enable the generation of Fully Focussed products over sea ice with snow depth retrievals derived from the Ku/Ka range differences instead of taking them from external auxiliary data. In the same way as in CryoSat-2, the closed burst interferometic mode over land ice will allow the generation of swath elevations for the full Greenland and Antarctica, improving the current coverage of the CryoSat-2 swath products that are only produced in the ice margins.

At this stage of the mission design, phase B2/C/D, the expected performances need to be evaluated against the requirements to verify the effectiveness of the mission configuration and assess its compliance.

In this framework, an end-to-end validation environment has been designed. It is composed of the System and Instrument Simulator (SIS), the Ground Processor Prototype (GPP) and the Performance assessment tool (PAT).

Following the validation plan defined during the first stage of the project, the SIS will be in charge of generating datasets for the different scenarios that are foreseen to be of interest for the mission performance assessment (e.g. point targets, sea ice with different snow properties, ice sheet with small slope and uniform snow and ice characteristics, glaciers with different size, slope and orientations, ocean tracks with different SWH and wind conditions, river and lakes for specific size and geometry).

In order to be able to generate such a broad set of data, the SIS is able to simulate orbit and attitude, the scene characteristics according the specific scenarios required and radar echoes following the different instrument design parameters.

The GPP will process the simulated data using different processing chains to ensure the compliance of the functional and performance requirements. It is composed among others of Level1 Calibration chains, Level1 Low Rate chains (LR-RMC, LR Over-Sampled and the conventional LR), Level1 Delay Doppler chain, Level1 Fully Focussed chain, Level2 retrackers module (compilation of different retrackers tailored for the different thematic surfaces), Level2 Geophysical corrections and retrievals (translating the information from the retrackers into sea ice, land ice, ocean and inland waters measurements).

The PAT is in charge of closing the end-to-end chain: it will cross check each of the geophysical parameters generated by the GPP against the corresponding requirement, starting from the knowledge of the simulated parameters, assessing and validating the end to end performance chain.

This presentation will give an overview of the expected performances of the CRISTAL mission based on the end-to-end validation activity carried out in this project.

Continuous monitoring of sea ice thickness is needed to constrain weather forecasts, plan shipping routes and to understand climate-driven changes in the cryosphere. At present, ice thickness measurements from the CryoSat-2 satellite Ku-band radar altimeter include an assumption of a fixed scattering horizon within the snow. However, scattering of the microwave radiation within the snow is one of the main sources of error in sea ice thickness retrievals, and this is sensitive to the roughness of dielectric boundaries as well as snow microstructure. The candidate ESA CRISTAL mission will incorporate a Ku-band and Ka-band sensor for deriving both snow depth and sea ice thickness and will be more sensitive to snow microstructure at the higher frequency. It is therefore imperative to understand microwave scattering within snow on sea ice better to develop future snow and sea-ice thickness retrieval algorithms from dual-frequency measurements.

Microwave scattering within the snow-sea ice column can be simulated with the Snow Microwave Radiative Transfer (SMRT) model. SMRT was initially designed to investigate and understand the role of snow microstructure in scattering in the X-Ka band range for both passive and active applications but has now been extended to incorporate Brown’s radar theory to simulate altimeter waveforms. This has been validated over Antarctica but has yet to be tested over sea ice. Here, we use data from ground campaigns linked to Operation Ice Bridge in 2016 and ESA CryoVex 2017 to demonstrate the sensitivity of the waveforms to observed variability in snow depth, density and microstructure. For a site with small variability in depth, microstructural rather than density differences accounted for the difference in Ku-band waveform.

Microstructural properties for these campaigns were derived from snow micropenetrometer measurements, reprocessed to correct for the incompatibility between instrument hardware and software but in the absence of independent microstructural information to verify accuracy. A new field campaign will take place in April 2022 in Cambridge Bay, Nunavut, Canada as part of the Altimetric Ku-Band Radar Observations Simulated with SMRT (AKROSS) project to evaluate SMRT for snow on sea ice. Field measurements will include a range of microstructural measurements from several techniques including x-ray tomography of casted snow samples. The roughness of air-snow, snow-sea ice and any crust boundaries within the snow will be measured with photogrammetry and used to parameterize rough interfaces within SMRT.

Here we present an evaluation of SMRT against airborne and satellite altimeter waveforms for the 2016 and 2017 field campaigns. We discuss the sensitivity of SMRT to snow microstructure, interface roughness and depth both at Ku- and Ka-band and the implications for sea ice thickness retrievals. Finally, we present preliminary data from the AKROSS field campaign, which will be a valuable resource for the design of future CRISTAL retrieval algorithms.

For the Cryosat-2 mission precision orbit determination (POD) is performed with the DORIS system and satellite laser ranging. We obtain a radial consistency of better than 1.3 cm which is a significant term in the performance budget of the altimeter system. Yet the method requires that there are accurate models to describe the accelerations acting on the satellite. The importance of dynamic models increases when the tracking data is not continuously available; due to the 700 km altitude of CryoSat-2 gaps in the IDS tracking network happen more often than for other altimeter systems like the Sentinel-6.

CryoSat-2 POD is affected by time variable gravity (TVG) modelling which depends on the availability of GRACE / GRACE-FO monthly series of gravity solutions that started in February 2002 well before CryoSat-2 was launched in April 2010. TVG solutions from GRACE are available until June 2017 when the mission came to an end; from this date on there is an outage until June 2018 when the GRACE-FO solutions become available.

In this talk we will show that, despite the interruption and despite quality differences between GRACE and GRACE-FO, a more realistic TVG model can be developed than by relying on the GRACE TVG modeling alone. The new TVG model strategy replaces the version developed from GRACE. Tracking residuals improve and our conclusion is that there is an optimal strategy for the POD of CryoSat-2.

Despite global warming, Antarctic sea ice expanded during most of the first four decades of satellite observations. However, in 2016 the Antarctic sea ice area plummeted, in a change far outside the range of previously observed variability. Neither the increasing trend nor the rapid decline are authentically simulated by climate models, casting doubt on their ability to represent associated processes including Southern Ocean heat and carbon uptake, melting of the Antarctic Ice Sheet, and many other aspects of the Southern Hemisphere climate.

DEFIANT is a £4M NERC-funded consortium to investigate the drivers and effects of fluctuations in sea ice in the Antarctic Ocean. The project will generate a new mechanistic understanding of the drivers and impacts of Antarctic sea ice variability, including the dramatic decline in 2016.

To achieve this overarching goal, we will launch an ambitious, internationally-leveraged field programme to provide the first comprehensive, systematic, year-round measurements of atmospheric and oceanic conditions at the Antarctic air sea-ice interface. The project will involve the collection of new measurements from subsurface gliders, autonomous submersibles, remote ice mass balance buoys, in situ stations, ice strengthened ships, fixed-wing aircraft, helicopter, and satellite platforms. We will use these observations to create and validate new physical models of sea ice snow loading and satellite retrievals of sea ice thickness, making use of CryoVex, CryoSat-2 and ICESat-2 radar and laser altimetry.

These retrievals and measurements will in turn be used to constrain a hierarchy of numerical models and apply them to understand the processes controlling the historical decadal sea ice expansion and 2016 decline. Finally, we will assess the short-term and decadal consequences of Antarctic sea ice variability. Through this interlinked programme of observations, model development and model evaluation, DEFIANT will deliver a step-change in our understanding of the Antarctic sea ice system.

The presentation will introduce the objectives of the DEFIANT project and present the results of the first field campaign to the Weddell Sea in Austral summer/spring 2022.

Snow depth and density are required to estimate sea ice thickness from satellite altimetry but must be assumed by using a climatology or model simulations based on reanalysis data. However, climatologies and modelled snow parameters lack temporal and spatial variability or convey substantial uncertainties. Obtaining snow depth from state of the Art or future satellite missions has the potential to reduce this gap.
The planned dual-frequency Copernicus Polar Ice and Snow Topography Altimeter (CRISTAL) satellite by the European Space Agency (ESA) with a launch planned in 2027 will have the capability to simultaneously perform range measurements in Ku/Ka band frequencies. This approach has the potential to sample the snow-ice and air-snow interfaces to derive snow depth on sea ice.
Future snow depth retrieval algorithms based on dual-frequency altimetry need to be calibrated and validated with in-situ data, and in-depth information about the snowpack is crucial to characterize penetration into the snow layer in both frequencies. Airborne campaigns (e.g., ESA CryoVEx) solely or in combination with ground measurements (e.g., snow probe surveys) have been carried out previously to validate satellite altimetry. However, a seamless link from ground observations via airborne surveys to satellite observations is limited by differences in range, spatial coverage, footprint size, sampling, as well as differences in sensors. Ground in-situ data seldom cover any thin or deformed ice areas. Moreover, also information about snow density on Arctic sea-ice is sparse. Current climatologies are mostly based on measurements on soviet drifting stations until the 1990’s. Therefore, new methods to retrieve accurate but spatially representative measurements of snow depth and density are highly required, especially considering the fast-changing Arctic climate.
Here we present a measurement system to fill the gap between ground measurements and aircraft observations of snow on sea ice for satellite validation. The presented unmanned aerial vehicle system carries an ultra-wideband radar and can obtain information about the snow layer on sea ice to complement future satellite calibration/validation activities, including measurements of snow depth and potentially snow density. In addition, it can also support future research studies and field work on sea ice that benefits from this mobile and flexible system.

DTU Space has carried out a large number of satellite-validation airborne campaigns over sea- and land ice since 1993, both in the Arctic and also a few in Antarctica. The ESA CryoVEx campaigns (along with associated EU and national campaigns) have provided a unique dataset for insight into sea- and land ice properties, such as penetration depths, sampling strategies, and surface roughness at different scales. The primary instruments flown include scanning lidar and Ku-band radar. Since Fall 2016 the instrument setup also includes a Ka-band radar, to exploit the dual-frequency concept for future Copernicus expansion mission, CRISTAL. Such long-term validation data series are key to form a baseline for inter-calibration and validation of multiple altimeter missions, and is ranking high as a candidate for future FRMs (Fiducial Reference Measurements).
The main objectives of the campaigns have been to understand retrieval of CryoSat-2 over sea- and land ice, and also inter-calibration with other satellite missions such as Sentinel-3, SARAL/AltiKa,ICESat-2, and upcoming - CRYO2ICE - in Spring 2022. The campaigns have frequently been coordinated with international teams of partners such as AWI’s EM-bird ice thickness measurements and NASA’s Operation IceBridge and ground team activities observing on-ice field observation of snow and ice properties. These collocated and contemporary measurements provide the opportunity to compare snow and ice parameters on scales from local to regional with the aim of improving the accuracy and benefit of the global satellite mission data
Here, we present an overview and the main learnings from the CryoVEx campaigns both in terms of past, present and future. We also present the most recent activities related to CRYO2ICE with preliminary results and discuss how we can maximize the outcome of this campaign and use past surveys to improve the baseline of validation observations.

The Copernicus Sentinel-3 Surface Topography Mission (STM) provides extremely valuable surface elevation information over inland waters, sea ice and land ice, thanks to its Synthetic Aperture Radar (SAR) altimeter which retrieves high-resolution along-track elevation measurements, and to its orbit that covers relatively high-latitudes in polar regions. The ESA St3TART project aims to generalize the concept of Fiducial Reference Measurements (FRMs) for the Copernicus Sentinel-3 STM and to collect and distribute FRM data for the validation of the satellite mission over inland waters, sea ice and land ice.

This presentation will focus on the current status of the work made within the St3TART sea ice theme, where the main objectives are to define the framework of FRMs to support a proper validation of the Copernicus Sentinel-3 STM Land sea ice products and associated geophysical measurements, i.e. the surface type classification over sea ice (leads/floes), the radar sea ice freeboard and the sea ice thickness. This implicitly includes knowledge of the snow depth, the speed of propagation of the radar wave in snow (assuming that the radar wave passes through the snow), and of densities of water, ice and snow.

We will present existing methods and lessons learnt from previous reference campaigns to identify suitable solutions in terms of sensors, validation sites, sampling strategy, sensor stability, uncertainty budget. Based on these, the sea ice team has planned a St3TART campaign in sea ice covered regions in spring 2022 coordinated with ESA CryoVEx/CRYO2ICE 2022 airborne campaign. The St3TART campaign will combine well tested techniques, i.e. a similar airborne setup as recent ESA CryoVEx campaigns with lidar and dual-frequency radar altimeters, as well as new novel techniques such as lidar equipped drones and drifting buoys, which in combination will provide a unique combination of data beneficial for the validation of the Copernicus Sentinel-3 sea ice products. We will in terms of the campaign setup present the full cycle of properly characterized and traceable to standards and/or community best practices of FRMs, using CryoVEx airborne measurements, as an example and provide highlights from the st3TART campaign.

The aim is to conclude the presentation by presenting an initial roadmap for the provision of Copernicus Sentinel-3 FRM for sea ice, considering the most relevant and cost-effective methods to be maintained, supported as far as possible or implemented, and identifying the parameters that are missing or are insufficiently covered or precise over the entire end-to-end duration of a satellite mission. This roadmap can potentially be used as a baseline for other present and future satellite altimetry missions, e.g. the Copernicus expansion mission CRISTAL.

Satellite laser and radar freeboard observations are integral to the estimation, modelling and monitoring of sea ice thickness and snow depth. It is commonly assumed that Ku-band radar altimeters penetrate through to the snow-ice interface, while Ka-band laser altimeters reflect off the air-snow interface, with snow depth assumed to be the difference between the two. Here, we investigate the relationship between freeboard retrievals and snow accumulation. We produce radar freeboards from CryoSat-2 and Sentinel-3 at 50km daily resolution using a recently developed Gaussian Process optimal interpolation scheme. We then compare these to snow accumulation in a physics-based, Lagrangian snow evolution model driven by atmospheric reanalysis. We show that a statistically significant relationship exists between radar freeboard and snow accumulation, in both regional analyses and analyses separated by ice age. We observe a positive correlation between radar freeboard and snow accumulation for the altimeters which suggests that, in the period after snowfall, radar pulses are not penetrating through to the ice surface as commonly assumed. Instead, the dominant scattering surface is located somewhere between the air/snow and snow/ice interfaces. While radar scattering from above the snow-ice interface has been observed in in-situ and airborne campaigns, our study presents the first satellite-based evidence for this unwanted phenomenon at a pan-Arctic and daily resolution (i.e. at sub-synoptic scales).

If anomalous scattering of Ku-band radar waves is not correctly accounted for, sea ice thickness returns may be significantly increased as a result of anomalous scattering above the snow-ice interface. However, sea ice thickness estimate biases have not been systemically high, which we argue is a result of masking by other uncertainties caused by waveform retracking and surface roughness. To investigate this masking, we test the sensitivity of radar-freeboard estimation to snow using two different retrackers and show that significant differences exist. Comparing a traditional (empirical) threshold retracker with a direct numerical simulation (physical) retracker, we show that radar freeboards that are retracked in a way that accounts for large-scale surface roughness are significantly less sensitive to the role of snow accumulation. We find in particular that these results hold for both Ku (CryoSat-2, Sentinel-3) and Ka (AltiKa) band altimeters and for LRM or SAR modes of acquisition. We finally compare our results with ICESat-2 (collaboration with Cryo2Ice project) and discuss our results in line with recent snow on sea ice product developments (in collaboration with the Polar+ Snow on Sea Ice project).

Meltwater forms at the base of the Antarctic Ice Sheet due to geothermal heat flux (GHF) and basal frictional dissipation. Despite the relatively small volume, this water has a profound effect on ice-sheet stability, controlling the dynamics of the ice sheet and the interaction of the ice sheet with the ocean. However, observations of subglacial melting and hydrology in Antarctica are limited. Here we use numerical modelling to primarily assess subglacial melt rates and hydrology within the Amery Ice Shelf catchment. Total subglacial melting is 6.5 Gt yr$^{-1}$, over 50% larger than previous estimates. Variation in total melting of $\pm 7$\% is due to uncertain estimates of GHF. This meltwater provides an extra 8\% flux of freshwater to the ocean in addition to contributions from the ice shelf. GHF and basal dissipation contribute equally to the total melt rate, but basal dissipation is an order of magnitude larger beneath ice streams. Remote-sensing observations, from CryoSat-2, constrain subglacial hydrology modelling, which shows a network of subglacial channels that link subglacial lakes and trigger isolated areas of sub-ice-shelf melting close to the grounding line. Building upon this Amery case study, we expand our analysis to quantify subglacial melt rates and hydrology beneath the entire Antarctic Ice Sheet.

The Polar Ice and Snow Topography Mission (CRISTAL) is an Earth observation mission developed by ESA, whose main objective is to perform quantitative measurements and monitoring of the variability of Arctic and Southern Ocean sea-ice thickness and its snow depth.
This work presents the different strategies for the Cristal Mission, flying in a non-synchronous orbit at around 685 km altitude, with an accurately controlled ground-track. This study addresses the design and computation of possible resonant orbits with the NASA ICESat-2 Mission.

The orbit resonance concept was already successfully implemented for Cryosat-2 flying in resonance with the same NASA Icesat-2. The orbit of Crysat-2 was marginally tuned such that the delta-longitude of Ascending Node between Cryosat-2 and ICESat-2 is kept constant after 19 Cryosat-2 revolutions and 20 ICESat-2 revolutions. This will create a rotating pattern of collocated observations in space and quasi simultaneous in time.

For the Cristal mission, different resonance orbit concepts have been studied:

Cristal R0, maintaining the Cristal reference orbit, and only optimising the difference in Right Ascension with ICESat-2. This is the reference case, tuning the Right Ascension of Cristal is possible as the Cristal – ICESat-2 envisaged campaign will be immediately after Cristal launch.

Cristal R1 which has exactly 2 (integer) orbits less per 3 (integer) Julian days compared to ICESat-2, maintaining 92.0 degrees inclination for Cristal. This will result in colocations fixed in longitude.

Cristal R2 which has exactly 2 (integer) orbits less per 3 (integer) Julian days compared to ICESat-2, with 92.2 degrees inclination for Cristal to have an identical orbital plane precession. This will result in colocations fixed in longitude with a constant delta observation time.

Cristal R3 which has N-1 orbits for N ICESat-2 orbits, maintaining 92.0 degrees inclination for Cristal - the concept selected for Cryosat-2 & ICESat-2. This concept results in colocations fixed in latitude.

Cristal R4 which has the same repeat cycle as ICESat-2 (91 days), but with less orbits. Finally, this concept results in colocations fixed in location – all over the globe.

Each of these resonant orbit concepts is studied to detail when co-linear observations with small time differences can be expected.

Radiative transfer modelling is an essential part of the Earth Observation measurement data processing and control chain. Radiative transfer simulation tools are notably used for Level-1 data calibration and Level-2 geophysical product validation. Verifying calibration with an accuracy below 3% can currently not be done reliably with common 1D radiative transfer models: achieving high accuracy levels requires a 3D radiative transfer model to relax the “smooth, flat, homogeneous Earth assumption” implicitly made in 1D models and account for the structural effects of complex geometry (e.g. cloud, terrain or vegetation).

While all these features are available from specialised software packages separately, few software packages currently integrate them, none of which is free software to our knowledge. The Eradiate radiative transfer model [1] is an attempt at addressing this need and was thought as a way to conveniently combine the advances made by all subcommunities (e.g. atmosphere, land, cryosphere, ocean colour, etc.) contributing to the development of radiative transfer modelling for Earth observation applications. It is also designed as a platform for experimenting with radiative transfer simulation algorithms and data.

The development of Eradiate started in 2019 with the goal of producing by 2022 an open-source radiative transfer simulation platform sufficiently accurate for usage in calibration/validation activities. It is written in C++ and Python using the Mitsuba 2 physically based rendering library [2], which provides powerful tools to assemble radiative transfer simulations accounting for the aforementioned physical phenomena. Eradiate integrates in modern scientific analysis workflows and offers a modern and comprehensive interface suitable for usage in interactive computing environments. Finally, Eradiate ships sourced, tested and traceable algorithms and data.

In this presentation, we will see how the Eradiate radiative transfer model can be relevant in the context of the radiometric calibration of optical imaging passive sensors, including uncertainty propagation when simulating satellite data. We will also review the current development progress and plans for this simulation platform.

References:

[1] Eradiate Website: http://www.eradiate.eu/
[2] Nimier-David, M., D. Vicini, T. Zeltner, and W. Jakob. “Mitsuba 2: A Retargetable Forward and Inverse Renderer”. ACM Transactions on Graphics 38, nᵒ 6 (2019): 203:1-203:17. https://doi.org/10.1145/3355089.3356498.

Recently, inter-comparison activities have been conducted for the Sentinel-3 OLCI and Sentinel-2 MSI sensors (A versus B unit). We present the motivations, methodologies and outcomes of these activities and investigate possible approaches for future missions.
The OLCI intercomparison relied mostly on data acquired during the tandem phase (Lamquin et al. 2020a), which allowed a direct comparison of the radiometry of the sensors after correction of geometric and spectral effects. After aligning the geometries and correcting for spectral response differences, the measurements of the two sensors can be directly compared across the whole field of view. Scenes are classified by radiometric types (cloud, desert, ocean, land) and statistics are compute for each class. This allows checking for possible non-linear or spectral effects. The method highlighted a bias of approximately 2%, with a nearly linear spectral variation. Validation tests showed that a radiometric alignment of the OLCI-A on OLCI-B could significantly improve the consistency of Level-2 products (Lamquin et al. 2020b, Hammond et al. 2020).

Since no tandem phase was implemented for Sentinel-2, the intercalibration of the MSI sensors relies on standard vicarious methods and comparisons with external references (using double ratios). A small bias of 1.1% between the two sensors was observed in the VISNIR domain (see e.g. Alhammoud et al. 2021). Results of inter-band methods such as the Deep Convective Cloud (DCC) method are consistent with a spectrally constant bias. On the basis of these results, the Sentinel-2 MPC proposes to harmonize the sensors for the new Collection 1 reprocessing currently in preparation. MSI-A would be use as the reference as the sensor is generally considered more accurate than MSI-B.

While current methodologies can support the operational inter-calibration of similar sensors, further improvements may lead to inter-calibrations within the Sentinel family to improve the overall inter-operability. The future inter-calibration strategy will rely on:
• Tandem operations for satellites of the same family whenever possible
• Improved vicarious methods using reference targets such as DCC (see Lamquin et al. 2020c) or the Moon (see Neneman et al. 2020)
• Reference calibration missions such as TRUTHS (Fox et al 2003)
• Future hyperspectral radiometer networks such as Hypernets (Goyens et al. 2018)

References

N. Fox, J. Aiken, J.J. Barnett, X. Briottet, R. Carvell, C. Frohlich, S.B. Groom, O. Hagolle, J.D. Haigh, H.H. Kieffer, J. Lean, D.B. Pollock, T. Quinn, M.C.W. Sandford, M. Schaepman, K.P. Shine, W.K. Schmutz, P.M. Teillet, K.J. Thome, M.M. Verstraete, E. Zalewski, Traceable radiometry underpinning terrestrial- and helio-studies (TRUTHS), Advances in Space Research, Volume 32, Issue 11, 2003, Pages 2253-2261, ISSN 0273-1177, https://doi.org/10.1016/S0273-1177(03)90551-5.

C. Goyens, K. Ruddick and J. Kuusk, "Spectral Requirements for the Development of a New Hyperspectral Radiometer Integrated in Automated Networks - the Hypernets Sensor," 2018 9th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2018, pp. 1-5, doi: 10.1109/WHISPERS.2018.8747259.

Lamquin, N.; Clerc, S.; Bourg, L.; Donlon, C. OLCI A/B Tandem Phase Analysis, Part 1: Level 1 Homogenisation and Harmonisation. Remote Sens. 2020, 12, 1804, https://doi.org/10.3390/rs12111804

Neneman, M.; Wagner, S.; Bourg, L.; Blanot, L.; Bouvet, M.; Adriaensen, S.; Nieke, J. Use of Moon Observations for Characterization of Sentinel-3B Ocean and Land Color Instrument. Remote Sens. 2020, 12, 2543. https://doi.org/10.3390/rs12162543

Clerc, S.; Donlon, C.; Borde, F.; Lamquin, N.; Hunt, S.E.; Smith, D.; McMillan, M.; Mittaz, J.; Woolliams, E.; Hammond, M.; Banks, C.; Moreau, T.; Picard, B.; Raynal, M.; Rieu, P.; Guérou, A. Benefits and Lessons Learned from the Sentinel-3 Tandem Phase. Remote Sens. 2020, 12, 2668. https://doi.org/10.3390/rs12172668

Lamquin, N.; Bourg, L.; Clerc, S.; Donlon, C. OLCI A/B Tandem Phase Analysis, Part 3: Post-Tandem Monitoring of Cross-Calibration from Statistics of Deep Convective Clouds Observations. Remote Sens. 2020, 12, 3105. https://doi.org/10.3390/rs12183105

Lamquin, N.; Déru, A.; Clerc, S.; Bourg, L.; Donlon, C. OLCI A/B Tandem Phase Analysis, Part 2: Benefits of Sensors Harmonisation for Level 2 Products. Remote Sens. 2020, 12, 2702. https://doi.org/10.3390/rs12172702

B. Alhammoud, C. Quang, V. Boccia and R. Q. Iannone, "Assessment of Copernicus Sentinel-2 Constellation After Five Years In-Orbit: Level-1C User-Products," 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, 2021, pp. 7759-7762, doi: 10.1109/IGARSS47720.2021.9554851.

In order to deliver high quality, traceable, multi-sensor optical satellite measurements, there is a need for common reference sources and dedicated intercalibration and bias assessment activities. The goal of the Ground to Space CALibration Experiment (G-SCALE) was to demonstrate the use of convex mirrors as a radiometric and spatial calibration and validation (cal/val) technology for Earth Observation (EO) assets in comparison with traditional diffuse targets operating at multiple altitudes and spatial scales. The experiment was carried out at the Rochester Institute of Technology’s 177-acre Tait Preserve in Penfield, NY, USA on July 23, 2021. The G-SCALE represents a unique, international collaboration between commercial, academic, and government entities for the purpose of evaluating a novel method to improve vicarious cal/val for EO. This abstract provides an overview of the experiment and acquired data sets, with details of analysis to be provided in future updates.
Traditional ground-based calibration and validation of airborne or spaceborne assets relies on diffuse reflectance surface targets. Targets suitable for spectroradiometric cal/val activities are typically spatially large, level, and homogenous and are void of any strong directional (i.e. specular or hotspot) or spectral domain features. Permanent targets such as large natural features (eg. Railroad Valley Playa), permanent manmade structures (e.g. concrete and asphalt surfaces) or specialized temporary cal/val materials (eg. tarps, Permaflect® panels) are commonly employed for airborne and satellite sensors (Joseph, 2005). For ultra-high resolution sensors such as those used with UAV platforms, grey-scale Spectralon® panels become a cost efficient, robust, solution (Hakala et al., 2018). In the spatial domain, high contrast targets like coastlines, bridges, airport runways, or purpose-built contrast arrays are used to characterize system performance (Languille et al., 2015).
The novel SPecular Array Radiometric Calibration (SPARC) method employs convex mirrors to relay an image of the solar disk to the sensor under test (Schiller and Silny, 2010). This approach employs mirrors or mirror-arrays to generate a point source with NIST-traceable absolute radiance for near-simultaneous vicarious calibration of hyper- and multi-spectral sensors in the solar reflective (VNIR/SWIR) spectral region. In addition to spectroradiometric assessment of the imagery, the sub-pixel point source nature of the mirror lends itself to a direct estimate of the Sampled Point Spread Function (SPSF), the full 2D Modulation Transfer Function (MTF), ground resolving distances and various post-processing artifacts. The technique is scalable over a wide range of spatial sampling resolutions, from sub-meter (UAV) to meters (manned aircraft) to tens of meter pixel footprints (Landsat 8/9, Sentinel 2, and PRISMA).
For the G-SCALE experiment, numerous reference and test targets suitable to the ground resolution of single or multiple imaging assets were deployed. These included SPARC mirrors, Pseudo-Lambertian Permaflect® reflectance standards, colored and grey-scale tarps, spectral unmixing targets, and spatial contrast arrays, as well as instrumentation for radiometric measurements. In addition to the mirrors deployed in the primary target area, two floating mirror targets were deployed on the water surface located at the northern end of the test area. These targets were specifically configured to provide low-radiance calibration points for dark-target, low signal-to-noise remote sensing applications such as water bodies. During overflights, near-simultaneous ground-truth data were acquired with a SVC HR-1024i field spectrometer (350 to 2500 nm) of the diffuse surface targets relative to a Spectralon® reference panel of known reflectance as a function of SZA (Soffer et al. 2019). Variability of downwelling solar irradiance levels (diffuse and direct) were monitored with using an ASD Field-Spec 4 configured with a cosine receptor (350 to 2500 nm).
Two Maxar owned and operated EO satellites were tasked for observation of the Tait Preserve during G-SCALE. GeoEye-1 launched on September 6, 2008 in a sun-synchronous orbit at an altitude of 770 km. The sensor geospatial accuracy is less than 5 m CE90 without ground control. A nadir view provides 0.46 and 1.84 m resolution for the panchromatic and multi-spectral bands, respectively. GeoEye-1 has four multispectral bands: blue (450-510 nm), green (510-580 nm), red (655-690 nm) and near-infrared (780-920 nm). WorldView-2 launched on October 8, 2009 in a similar orbit as GeoEye-1. With identical geospatial accuracy and nadir view resolutions to GeoEye-1. WorldView-2 has eight multispectral bands: coastal (400-450 nm), blue (450-510 nm), green (510-580 nm), yellow (585-625 nm), red (630-690 nm), red edge (705-745 nm), near-infrared 1 (770-895 nm) and near-infrared 2 (860-1040 nm). A total of three satellite images were collected of the Tait Preserve target area. GeoEye-1 collected images at view angles of 25.9° at 15:56:29 GMT (SZA = 28.6°) and 31.0° at 15:56:47 (SZA = 28.6°) resulting in nominal GSDs of 0.479 m for pan and 1.920 m for the multi-band, and 0.513 m pan 2.056 m multi respectively. WorldView-2 collected a single image at a view angle 31.7°at 15:52:53 (SZA = 29.1°), resulting in nominal GSDs of 0.582 m for the pan and 2.326 m for the multi-band.
Mounted aboard a Twin-Otter aircraft, two co-mounted and complimentary airborne hyperspectral systems were deployed by the National Research Council Canada (NRC) in support of G-SCALE over the Tait Preserve: the Compact Airborne Spectrographic Imager (CASI-1500) covering the Visible and Near Infrared (VNIR) spectral range from 372 nm to 1062 nm and the Shortwave Airborne Infrared Spectrographic Imager (SASI-600), covering the 957 nm to 2442 nm spectral region, both with a field-of-view of ±20°. CASI imagery is composed of 1496 spatial pixels in 288 channels (native mode). Programmability built into the system allows the spectral resolution to be reduced by summing channels allowing faster frame rates to be employed, resulting in smaller along-track spatial spacing. A total of 7 identical imaging overpass were made of the target site at an altitude of 793 ± 4 m above ground (water surface) level between 15:39:21 (SZA = 31.1°) and 16:09:21 GMT (SZA = 27.3°) with a ground speed of 44 ± 2 m/s. Four of the CASI passes were performed using a sum x 3 configuration (96 channel) with an integration time (IT) of 17 ms to allow an assessment of the repeatability and spatial aliasing effect within the process under evaluation (1st, 2nd, 6th, and 7th flight lines). In addition, imagery at three additional integration times (24, 40, and 60 ms) was collected in order to assess the impact of pixel GSD on the radiometric performance of the sensors with respect to SPARC calibration targets resulting in nominal native cross-track resolutions of 0.38 m for all images with along-track resolutions of 0.75, 1.08, 1.77 and 2.68 m for the four employed ITs. SASI imagery is composed of 600 spatial pixels and 100 channels operating with a fixed frame rate (66.6 fps – 16 seconds per frames) but with programmable IT in order to optimize the recorded signal levels. Imagery was acquired with ITs of 1.5, 2.0, 2.5, 3.0 and 3.5 ms) resulting in native cross-track resolution of 0.92 m and along-track spacing of 0.71 m. The overpasses were timed in coordination with the Maxar satellite overpass.
Two primary RIT UAVs were flown as underflight systems to the air-borne and space-based platforms. The first is called the MX1-UAV. The MX1 multi-modal UAV is a modified DJI Matrice 600 and contains five sensors that collect data simultaneously. The platform contains a Headwall Nano Hyperspectral VNIR sensor (400 to 1000 nm, 270 bands), a Velodyne VLP16 LiDAR sensor, Tamarisk640 LWIR (8-14um) sensor, Mako G419 5-megapixel RBG sensor and a Micasense Red Edge Multispectral sensor. The IMU/GPS on-board the MX1 is an Applanix APX15. A second UAV, called the SWIR-UAV, was flown in tandem to the MX1. This platform contained a Headwall Hyperspectral SWIR sensor (900 – 2500 nm, 267 bands) and Applanix APX15 with 1-3 cm accuracy. A third UAV (DJI Mavic 2 Enterprise) containing a 12 megapixel CMOS camera was subsequently flown to provide hi-resolution contextual RGB imagery. Data from the two hyperspectral sensors as well as the hi-resolution RGB photos were processed to provide geo-referenced imagery that was then mosaicked over the entire field campaign site.

Inter-calibration is one of well-known calibration methods monitoring the measurement performance after the launch of a satellite. The method could be highly effective when the number of collocation data is sufficient to fully cover the measurement characteristics of each sensor. The satellite sensors onboard an identical satellite has an advantage in this regard, considering that the sensors observe earth with an identical observation angle, a difficult factor to be matched with for ray-matching. The sensors onboard the Geostationary Korean Multi-Purpose Satellite-2A/B (GK-2A/B) can fully exploit the advantage as the sub-nadir longitude of the satellites are identical to 128°E. The Atmospheric Monitoring Imager (AMI) onboard GK-2A measures radiances at visible and infrared spectral channels for atmospheric monitoring. The AMI has been continuously monitored and inter-calibrated with the reference sensors onboard low earth orbit satellite sensors of the Global Space-based Inter-Calibration System (GSICS), which could be a useful transition reference for the sensors onboard GK-2B. The Geostationary Environment Monitoring Spectrometer (GEMS) and the Geostationary Ocean Color Imager-2 (GOCI-2) onboard GK-2B provide hyperspectral radiance spectrum in 300-500 nm and the broadband channel radiances from ultraviolet to near infrared spectral region, respectively for environmental and ocean monitoring. With the measurements of the three sensors, in this study, the availability of inter-calibration in the ultraviolet-visible spectral region is evaluated with the collocation thresholds in terms of spatial, time and spectral response. The channels of AMI and GOCI-2 are selected with the condition whether the spectral response function could be covered by a GEMS spectrum. The different spatial response of each sensor is matched by applying spatial averaging with the grid size of 0.1 degree. The comparison results show that the three sensors have high correlation around 0.98 by eliminating cloud edges, which has high spatial and temporal variability causing significant collocation error. Standard deviation of the AMI reflectance at 470 nm is used to evaluate the time and spatial homogeneity of a scene with the empirical threshold of 0.05. The interesting finding is that the shorter the wavelength is, the more affected the radiances are with the solar zenith angle dependence caused by Rayleigh scattering, and this causes high time-dependent biases in the ultraviolet-visible spectral region. With further update reducing the biases, the approach is expected to provide a means of monitoring and quantifying the long-term trend of the sensors during the operation.

The PROBA-V mission has ensured global vegetation monitoring from October 2013 until October 2021. It provides data continuity with the monitoring of SPOT-VGT (since 1998), and managed to do so with a much smaller satellite (less than 1m3). By using three compact multispectral cameras, it offers comparable coverage and spectral bands, with strongly improved spatial detail. It provides global land coverage daily (every two days near the equator) at 1/3 km (BLUE, RED and NIR) and 2/3 km (SWIR) resolution. Additionally, every 5 days global coverage with 100 m resolution (200m for SWIR) is achieved.

To ensure excellent in-orbit radiometric and geometric performance, a dedicated software system, called 'Image Quality Center (IQC)' was developed and deployed at the VITO premises. After successfully supporting the commissioning phase, the IQC has been continuously monitoring and adjusting the calibrated status of the instrument in order to maintain optimal performance. A suite of complementary Cal/Val geometric and radiometric correction routines were applied on a regular basis to ensure the proper calibration of PROBA-V instrument throughout its lifetime.

Due to the absence of on-board calibration devices, radiometric in-flight performance monitoring and calibration relies fully on vicarious calibration methods. The radiometric calibration facility contains, among other things, the OSCAR (Optical Sensor CAlibration with simulated Radiance) tools which exploit the reflected radiance over bright desert surfaces, Deep Convective Clouds (DCC), Ocean Rayleigh scattering and the Moon.

To maximise PROBA-V product’s geolocation accuracy, a continuous and highly autonomous in-flight geometrical calibration/validation software has been implemented in the mission’s User Segment (US) to estimate and monitor the Exterior Orientation parameters (boresight angles) and Interior Orientation deformations (CCD line of sight vectors) of each camera of the PROBA-V Vegetation Instrument (VI) on a regular basis. The geometrical parameters are estimated using a large number of globally distributed Ground Control Points (GCP), a priori extracted from the Landsat Global Land Survey GLS 2010 cloud-free dataset. The estimated geometrical calibration parameters are then used to update the Instrument Calibration Parameter (ICP) file that is used by the US Processing Facility.

The overall PROBA-V in-flight calibration approach was successful in ensuring outstanding and temporally stable radiometric and geometric performances throughout the full mission lifetime despite the relatively small satellite platform and the lack of on-board calibration devices.
In this paper, details of the radiometric and geometric calibration activities during the operational phase will be addressed. The stability and trend analysis of the instrument radiometric behaviour (bias, temperature dependencies, dark current variation) will be presented.

The Meteosat Third Generation (MTG) Programme is being realised through the well-established and successful Cooperation between EUMETSAT and ESA. It will ensure the continuity with, and enhancement of, operational meteorological (and climate) data from Geostationary Orbit as currently provided by the Meteosat Second Generation (MSG) system. The last of which MSG4/MET11 has been successfully launched, commissioned in 2015 and is in full operation mode at 0o longitude since end of 2018.

The industrial Prime Contractor for the Space segment is Thales Alenia Space France (TAS-F) with a core team consortium including OHB-Bremen (Germany) and OHB-Munich (Germany). This contract includes the provision of six satellites, four Imaging satellites (MTG-I) and two sounding satellites (MTG-S), which will ensure a total operational life of the MTG system in excess of 20 years. TAS-F is also responsible for the MTG Imager development, namely the Flexible Combined Imager (FCI).

The FCI is an improved successor of the SEVIRI on board MSG. It is designed to provide images of the Earth every 2.5 to 10 minutes in 16 spectral channels between 0.44 and 13.3 μm, with a ground resolution ranging of 0.5 km, 1 km and 2 km. Following the completion of its Qualification Review in July 2021, the FCI has undergone its testing phases. The Structural and Thermal Models completed in 2019, the Engineering Models in performed in 2020 and the proto-flight (PFM) tests is successfully completed in July 2021. The end to end radiometric, spectral and image quality performance evaluation indicates that the FCI confidently meets most of its stringent mission requirements. The integration and testing onto the satellite PFM (MTG-I1) has been successful which allowed the testing at satellite level to be carried out.

This paper will present the FCI instrument development status. The presentation will provide the test results and the overview of the measured performances, including the in-flight predictions for the FCI in terms of image quality, spectral and radiometric performances.

The Copernicus Sentinel-2 mission comprises a constellation of two polar-orbiting satellites placed in the same sun-synchronous orbit, phased at 180° to each other. It aims at monitoring variability in land surface conditions, thanks to an optical push-broom multispectral instrument (MSI). With a wide swath width (290 km) and high revisit time (10 days at the equator with one satellite, and 5 days with 2 satellites under cloud-free conditions which results in 2-3 days at mid-latitudes) it sup-ports monitoring of Earth's surface changes.

The ground segment is composed of two main parts. The Flight Operations Segment monitors and controls the Sentinel-2 spacecraft, ensures uplink of the mission plan and performs orbit control maintenance and collision avoidance. The Payload Data Ground Segment (PDGS) ensures the measurement data acquisition, processing, archiving and dissemination to the final users. In addition, it is responsible for performing conflict-free mission planning according to a pre-defined operational scenario, and ensures the quality of data products and performance of the space-borne sensors by continuous monitoring, calibration and validation activities, guaranteeing the overall performance of the mission. This last task is responsibility of the Mission Performance Centre (MPC).

In order to improve the radiometric stability of the payload the mission performs regular Sun calibrations. In the frame of the development of the recurrent models (S2C and S2D), an additional Moon calibration was introduced. This calibration requires a slew of the satellite around the flight velocity axis during Sun eclipses, de-facto letting it fly for a brief period almost upside-down. This activity has been implemented without changes in the satellite design, but taking advantage of the envelope of the already existing capabilities.

In coordination with the different shareholders, ESA plans to implement the Moon calibration as extension of the mission planning activity. The FOS will be in charge to identify the opportunities for the calibration and validate the sequence for the slew, while the MPC will trigger the actual execution of the operation based on the radiometric measurement needs.

This presentation will describe the objectives of the Moon Calibration, the planned implementation, the processing and integration of the results as well as the scientific benefit of this additional activity.

As part of the Copernicus programme of the European Commission (EC), the European Space Agency (ESA) operates the Sentinel-2 missions that acquire high spatial resolution optical imagery. The Multispectral Instrument (MSI) on board of Sentinel-2 A and B platforms measures the Earth's reflected radiance in 13 spectral bands from VNIR to SWIR.
The radiometric consistency of the image time series is ensured by some specific performance requirements such as the radiometric stability, routinely monitored by the Sentinel-2 Mission Performance Centre (S2MPC). This presentation provides a description of the radiometric calibration activities, performed by the S2MPC including several Expert Support Laboratories (ESL).

The radiometric calibration is based on the exploitation of the on-board sun diffuser images (for relative gains and absolute radiometric calibration) and images acquired over ocean at night (for dark signal calibration).
The on-board sun diffuser is a full-field/full-pupil diffuser called the Calibration and Shutter Mechanism (CSM). It is integrated to MSI to ensure the radiometric calibration in order to guarantee high quality radiometric performance. This on-board calibration device reflects the sun light to the sensor after reflection by the diffuser.

For the VNIR bands, the absolute calibration coefficients are stable, while for the SWIR bands (especially for the band B10), there is a decrease trend which is currently up to −1.5% in one year. However, the monthly updates of these coefficients, as well as the MSI decontamination operation every year, compensate properly for this trend and the sensor sensitivity is completely retrieved.
The temporal evolution of the equalisation coefficients is monitored too with the monthly sun-diffuser acquisitions. A new set of relative gains calibration parameters is generated after each sun-diffuser acquisition. The time variation of relative gains is weak for the VNIR bands: typically, maximal variations do not exceed 0.2 to 0.4% between two monthly assessments. Variations are larger for the SWIR bands with a change of gain coefficients up to 3% for some pixels and with an inter-pixel variation more pronounced for these bands than for the VNIR bands.

The instrument response non-uniformity is assessed through the Fixed Pattern Noise (FPN) which quantifies local non-uniformities in the response of physical detector pixels across the swath. For all the VNIR bands, the maximal FPN value is clearly below the specified limits (0.2% for all bands) typically by an order of magnitude. For the SWIR bands, the maximal FPN value can be sometimes higher than the specified limits (0.35% for the band B10, 0.2% for the band B11). However, the FPN values exceed the limit only for a few pixels as the values of FPN for the 98% quantile are significantly below the specified limit, even for these bands which are more affected by artefacts on the sun-diffuser BRDF characterization than the VNIR bands.
Nevertheless, the monthly update of the equalization coefficients resets the FPN to zero and ensures that the requirements are met until the next calibration.

The Signal-to-Noise Ratio (SNR) is well inside the requirement (with more than 20% margin) for all bands and appears to be very stable.

The latest radiometric calibration results highlight the good quality of the mission products in terms of radiometry. The current performances for MSI-A and MSI-B compared to mission product quality requirements will be presented, including a focus on the radiometric noise.

Since 1999, Radiation Transfer Model Intercomparison (RAMI) proposes a mechanism to benchmark Radiative Model (RT) models designed to simulate the transfer of radiation at or near the Earth's terrestrial surface, i.e., in plant canopies and over soil surfaces (Pinty et. al., 2001). RAMI operates in consecutive phases, each one aiming at re-assessing the capability, performance and agreement of the latest generation of radiative transfer (RT) models.

This in turn, leads to model enhancements and further developments that benefit the RT modelling community as a whole. The fourth phase of RAMI (RAMI-IV) included a new set of architectural scenarios conveniently subdivided into "abstract" and "actual" canopies (Widlowski et. al., 2013, 2015). In 2020, the fifth phase of RAMI (RAMI-V) proposes the same canopies as RAMI-IV with two new architectural canopies. We increase the number of experiments by including more spectral bands (such as the ones of the Copernicus Sentinel-3 Ocean Land Color Imager (OLCI), Sentinel-2 Multispectral Instrument (MSI) and the Moderate Resolution Imaging Spectroradiometer (MODIS) sensors, covering the solar spectrum from 0.4 to 2.2 micron). The actual canopies were based on either detailed inventories of existing vegetation sites but also on highly detailed 3D terrestrial laser scanning (TLS) data to generate virtual forests (Calders et. al., 2018).

Bidirectional Reflectance Factors (BRFs), albedo and fluxes, as well as the response of ground-based measuring devices such as the Tracing Radiation and Architecture of Canopies (TRAC) or Digital Hemispherical Photography (DHP) are the main measurements to be simulated.

In addition to the traditional RT models, ‘emulators’ were welcome to participate for the abstract canopies experiments. Most of the simulation have to be performed in either direct or diffuse illumination conditions. The sun geometries are taken from real satellite geometries in different seasons. All participants were invited to provide computations times for assessing the benefit of such approach versus reliable results.

This paper presents preliminary analysis of the RAMI-V results.

References
Pinty, B., N. Gobron, J.-L. Widlowski , S. A. W. Gerstl, M. M. Verstraete, M. Antunes, C. Bacour, F. Gascon,
J.-P. Gastellu, N. Goel, S. Jacquemoud, P. North, W. Qin, and R. Thompson (2001), 'Radiation Transfer Model
Intercomparison (RAMI) Exercise', Journal of Geophysical Research, 106, D11, 11,937-11,956, DOI:
10.1029/2000JD900493.
Widlowski J.-L., Pinty B., Lopatka M., Atzberger C., Buzica D., Chelle M., Disney M., Gastellu-Etchegorry J.
P., Gerboles M., Gobron N., Grau E., Huang H., Kallel A., Kobayashi H., Lewis P., Qin W., Schlerf M.,
Stuckens J. and Xie D. (2013), 'The fourth Radiation Transfer Model Intercomparison (RAMI-IV): Proficiency
Testing of Canopy Reflectance Models with ISO-13528.', Journal of Geophysical Research - Atmospheres, 118,
D09111, 13.DOI: http://dx.doi.org/10.1002/jgrd.50497.
Widlowski J.-L., Mio C., Disney M., Adams J., Andredakis I., Atzberger C., Brennan J., Busetto L., Chelle M.,
Ceccherini G., Colombo R., Côté J-F., Eenmäe A., Essery R., Gastellu-Etchegorry J. P., Gobron N., Grau E.,
Haverd V., Homolová L., Huang H., Hunt L., Kobayashi H., Koetz B., Kuusk A., Kuusk J., Lang M., Lewis P.,
Lovell J. L., Malenovsky Z., Meroni M., Morsdorf F., Mõttus M., Ni-Meister W., Pinty B., Rautiainen M.,
Schlerf M., Somers B., Stuckens J., Verstraete M. M., Yang W., Zhao F. and Zenone T. (2015), 'The fourth
phase of the radiative transfer model intercomparison (RAMI) exercise: Actual canopy scenarios and conformity
testing.', Remote Sensing of Environment, 169:, 418-437.
Calders, K., Origo, N., Burt, A., Disney, M. Nightingale, J., Raumonen, P., Akerblom, M, Malhi, Y. and Lewis,
P. (2018) Virtual forest for radiative transfer modelling: realistic stand reconstruction from terrestrial LiDAR.
Remote Sensing, 10(6):1-15

PICSCAR initiative starts in 2015 following a recommendation of the CEOS Infrared and Visible Optical Sensors Subgroup (CEOS-IVOS) to prioritize research on Pseudo Invariants Calibration Sites (PICS) and their usage for the benefit of the EO community. A list of PICS has been identified by Cosnefroy et al., 1996 for the calibration of medium/coarse spatial resolution instruments in the solar spectral range (400–2500 nm), leading to the creation of the CNES SADE database (Cabot, 1997; 1999), which evolutions have been retraced by Henry et al., 2008.
20 sites have been identified, located in the Saharan desert and in the Arabian Peninsula. Six of them have since been endorsed by the CEOS/WGCV/IVOS as reference Calibration/Validation test sites for the high temporal stability of their surface optical properties combined with a high spatial homogeneity.
In the frame of this CEOS activity, a roadmap has been set up at IVOS 28 in March 2017 to continue to collect information on the selected PICS. Priorities have been given to 6 sub subjects that need to be known to evaluate the long-term stability of an instrument and to facilitate inter-comparison of multiple instruments. These are the BRDF characterization, the spectral characterization, the atmosphere properties, the temporal stability, the combination of multiple sites calibration results and the need to revisiting the sites.
A PICSCAR portal (https://picscar.magellium.com/) has been set up to gather the information.
Thanks to data collected over the Libya4 site for AATSR, AVHRR, MERIS, MISR, MODIS, PROBAV, VGT2, S3A/OLCI, S3A/SLSTR, PARASOL, VIIRS, FY3, L8, S2A, S2B/MSI, HYPERION, the site stability has been monitored on 10 years of sensor time series thanks to CNES BRDF model based derived from PARASOL. Spectral variability has been assessed using HYPERION data. Atmosphere has been characterized thanks to ERA Interim products. Site usage for intercalibration monitoring assessed for S2A/MSI and L8/OLI. This has led to the operational monitoring the sensor radiometry established now from 3 years. CNES and SDSU provides us update of L8 versus S2A intercalibration results every 6 months. Results are published on the portal.
In this presentation we will give an insight of the PICSCAR update: extension to other selected sites, update of data collection for some sensors, assessment of the spectral behaviour of the sites using PRISMA data…
References:
Cosnefroy, H., M. Leroy, X. Briottet (1996) Selection and Characterization of Saharian and Arabian desert sites for the calibration of optical satellite sensors. Remote Sensing of Environment, 58, 101– 114
Cabot F. (1997) A proposal for the development of a repository for inflight calibration of optical sensors over terrestrial targets. Proceeding of earth observing systems II. SPIE, vol. 3117
Cabot F., O. Hagolle, C. Ruffel, P. Henry (1999) Remote Sensing Data Repository for In-Flight Calibration of Optical Sensors over Terrestrial Targets, Proceedings of SPIE, 3750, Denver
Patrice Henry, Claire Tinel, Sophie Lacherade, Bertrand Fougnie, Denis Blumstein, - CNES GSICS WG Meeting – Feb 2008 – Claire Tinel / CNES Characterization and use of desert sites as in-flight calibration targets

Planet currently operates a fleet of more than 200 satellites. SkySats, which comprise 21 satellites of the overall fleet, are Planet’s high resolution satellites which have 5 spectral bands - red, green, blue, near infrared and panchromatic. The SkySats were launched over a period of seven years beginning in 2013. These sub-meter resolution satellites can image scenes in a range of viewing angles, create 3D scene composites, take videos even in regions traditionally difficult to observe due to low satellite capacity and have high intra-day revisit rate capabilities. Consistency and interoperability between all of Planet’s satellites is essential. To attain this consistency, all calibrations across Planet’s fleets will be based on gathering a dataset of near simultaneous crossovers with Sentinel-2 as the reference satellite. In order to achieve optimal radiometric calibration, SkySats image Pseudo Invariant Calibration (PIC) sites and a set of further calibration sites that have been characterized by Hyperion imagery. The SkySats are tasked daily to generate numerous near-simultaneous crossovers of these sites with Sentinel-2. Calibration crossovers with Sentinel-2 are limited to these well characterized sites as the spectral responses of SkySats are quite different from Sentinel-2 and therefore a spectral band adjustment factor (SBAF) must be applied before comparison. Radiometric calibration of SkySat imagery using Sentinel-2 poses unique challenges as it has relatively wide spectral bands, frequently takes images off-nadir and there is a large difference in resolution between the satellites. The first SkySats, 1-15, are at a 450 km altitude ~98° sun-synchronous orbit (similar to the PlanetScope Dove satellites). The most recently launched SkySats, 16-21, are at a 53° inclined orbit at a lower altitude; this was done to increase the intra-day revisit rate of the constellation. The variable crossing times of these satellites, though, yields a large variation in sun angles and local crossing times. The wider variation of sun elevation angles and possibilities for off-nadir imaging in the SkySat fleet create complications when trying to calibrate the radiometry of the sensors. The large differences in footprints between the satellites also meant that there were some particular challenges when trying to generate calibration data. This presentation will discuss some of these issues as well as how we managed to overcome them to get an accurate radiometric calibration. We also will present work on characterizing the variation in uncertainties when sun-angle and view-angle are changed. In addition, the SkySat payload consists of three individual sensors. Previous calibration campaigns used RadCalNet data to calibrate the SkySat and so, due to the small footprint of the RadCalNet sites, could only characterize the center sensor. The other two sensors were only calibrated in relation to the center sensor. By using crossovers with Sentinel-2, we now have the opportunity to provide absolute calibration to all three sensors independently and we will present results on this. Planet has now unified its use of RadCalNet data as an independent and traceable radiometric validation data source for both the reporting of Dove and SkySat radiometric uncertainties. We will present some of the latest results of SkySat crossovers over the RadCalNet sites.

The Sentinel-3 SLSTR instrument is primarily designed to measure Sea and Land Surface Temperatures for metrological and climate research applications. As well as spectral bands in the thermal infrared for measurement of surface temperatures, the radiometer is equipped with channels in the visible to short wavelength infrared range, primarily for daytime cloud screening and scene classification. However, the VIS-SWIR channels are used in the retrievals of Fire Radiative Power (FRP) and the land applications when used in synergy with the OLCI instrument data. Furthermore, when combined with the dual view capability of the radiometer, the data from the VIS-SWIR channels are used for measurements of aerosol and cloud properties not possible with a single view instrument.

The use of the VIS-SWIR channels in Sentinel-3 level-2 data products, demands that they be radiometrically calibrated to standards traceable to SI. Demonstrating direct traceability to SI is not so straightforward as the L1 processing has many inputs, and several assumptions are made about the instrument model. Analysis of the radiometric model used in the L1 processing shows that there are several key factors that affect the radiometric performances, for instance:

• Calibration of the diffuser-based calibrator
• Long term stability of the calibrator
• Response non-linearity
• Ground-orbit changes – e.g., where there are differences between the pre-launch test conditions and flight operations.

The approach that has been followed for SLSTR has been to perform a pre-launch calibration test campaign at instrument level, and to verify the radiometric performance by comparisons with other similar satellite sensors over stable reference scenes, or by analysis of sun-glint scenes. In this paper we present the results of the calibration analysis from the first 5 years of the Sentinel-3A and B missions. The results show the differences of the SLSTR A and B measured top-of-atmosphere reflectances relative to the reference sensors through direct or indirect comparisons, the long-term trend stability of the radiometric calibration and an assessment of the non-linearity of the VIS/SWIR channels.

Activities for the SLSTR-C and D units are in progress at the time of writing, and lessons from the model-A and B units are being applied to improve the pre-launch calibration activities.

The radiometric calibration of the Multispectral Instrument (MSI) on board of Sentinel-2 A and B platforms is based on the exploitation of the sun diffuser acquisitions. This on-board device is a full-field/full-pupil diffuser, called the Calibration and Shutter Mechanism (CSM), which is used for relative gains and absolute radiometric calibration. It reflects the sun light to the sensor after reflection by the diffuser.
In order to assess the absolute gains per band and to evaluate the change of relative gain coefficients per pixel and band (13 bands from VNIR to SWIR, 12 detectors per bands, 1296 or 2592 pixels per detector), a comparison between the measurements and the simulated observations is done. The accuracy of the simulation is a key factor in the precision of the radiometric calibration. It depends on the knowledge of the incoming solar irradiance for the different spectral bands and on the accuracy of the sun-diffuser BRDF characterization. The latest was estimated, for MSI-A and B, from on-ground test bench measurements for different illumination and viewing angles, and contains the sun-diffuser’ reflectance for each pixel (per band and detector). These measurements were interpolated for a constant viewing angle per pixel (constant for sun-diffuser observations) to provide a set of reflectance nodes for illumination angles: Solar Zenith Angles (SZA) and Solar Azimuth Angles (SAA).

The temporal evolution of the relative gain coefficients is monitored monthly, based on sun-diffuser acquisitions. For VNIR bands, the stability of these coefficients is expected to be very good. However, the monthly monitoring has shown seasonal variations of the relative gain coefficients, which can be clearly correlated to the variation of the SAA angle. Indeed, the BRDF artefacts can induce a temporal change in the pixel equalization, mainly at the edge of the field of view (i.e., for the extremum detectors). For these pixels, the equalization coefficients can change during the year by about 0.8% for the VNIR bands, and more for the SWIR bands. That leads to a non-expected time variation of the radiometric reflectance measured from the concerned pixels and to a seasonal non-uniformity in the field of view.

The first objective of the presented study is to characterize the BRDF artefacts for each pixels and bands. Then, from this characterization, a correction scheme is introduced to improve the stability of the radiometric calibration results from one calibration sequence to the next one. Finally, the possibility to directly correct the Sun-diffuser BRDF model is analysed, to allow an improvement of the calibration results by simply substituting the sun-diffuser BRDF configuration file without adjusting the radiometric calibration code.

The remote sensing community and other users in the public or private sector are interested in reliable, highly accurate and fast orthorectification of data for various applications. This presentation addresses the prototype of a service for automatic orthorectification of high-resolution (HR) and very high-resolution (VHR) optical satellite imagery that meets these needs. The service is designed to process large amounts of data from different sensors and to use ancillary reference data that is automatically obtained from dedicated, freely available web databases. The service is specially prepared to operate in a cloud-based environment.

The developed prototype of the orthorectification service is divided into two main parts:
- the web application (including the user interface (UI) segment), and
- the processing module (fundamental computation).

The service web application part manages the communication between the user and the processing module. All processes are interconnected and work in the backend of the system. The user interacts with the system only through the website where the input data and additional/optional parameters are selected. The user must upload the input images in a ZIP or TAR file format. The processing service then unpacks the files and sends them to the processing system. Once processing is complete the user automatically receives a download link to the finished product via email.

The core of the application is the processing module, which is tasked with creating orthoimages. The module uses auxiliary data from various geodatabases on the Internet, which are maintained and periodically updated by different institutions. To obtain accurate results, the processing module requires specific ancillary data:
- road network data (e.g. OpenStreetMap (OSM) data),
- digital terrain model (DTM) (e.g. Shuttle Radar Topography Mission (SRTM) data, airborne laser scanning (ALS) data),
- orthophotos or orthoimage data (e.g. Copernicus data, ALS data).

In most cases, road network data is retrieved from the OSM database. Elevation data can be accessed from different sources depending on the geographic location of the image. Wherever possible, the module uses a high resolution DTM based on freely available ALS data. In regions where no open source ALS data is available, the module uses SRTM data. Orthophotos proved to be more difficult to obtain and are therefore replaced by ALS intensity data where available.

The processing module was designed and implemented in a modular way – it consists of 6 modules. Each module represents a distinct processing part, which constitutes a specific topic. It also forms a separate part of the code that functions relatively independently but is connected to other modules.

The modules of the prototype processing module are:
• MODULE 1: Input data preparation
This module is mainly concerned with extracting information from the metadata file. It uses a parser to automatically extract the basic image parameters and rational polynomial coefficients (RPCs) needed for the geometric corrections. The parameters that are not in the correct coordinate system or have different units than those required are transformed and stored. The module also prepares a single raster file of the satellite image: It is unpacked, read in and mosaicked if necessary; finally, all bands are stacked into a single image.
• MODULE 2: Ancillary data preparation
The module deals with the automatic download and preparation of ancillary data. It consists of three different parts, each part focusing on a specific product: DTM, roads and reference orthoimages. The most complex part is the preparation of the DTM: if high-resolution ALS data is not available for the target area of interest (AOI), a global DTM from SRTM data is used instead. The road data is downloaded and extracted from the OSM database and used to create the road distance map. Creating the reference image is trivial if orthoimages are available for download. Since this is rarely the case, the intensity image derived from the ALS point cloud is used instead. If no ALS data is available, the accuracy of the matching may suffer.
• MODULE 3: Road detection
Road detection depends on the resolution of the input satellite image. When processing satellite images with a resolution of less than 1.5 m, the detection algorithm uses morphological image filtering with the “top-hat” operator of a predefined size that detects bright road line candidate segments. For VHR satellite images, the road detection algorithm uses watershed segmentation with decision tree classification of segments based on their geometric properties. This is a better choice than a "top-hat" since the width of the road is preserved. Both methods produce a binary mask of candidate roads.
• MODULE 4: Image matching
The automatic ground control point (GCP) extraction algorithm performs a matching of the roads detected on the satellite image with the rasterized digital vector roads or road distance data. The initial transformation of the detected binary road mask is performed in two successive affine transformations. The main GCP extraction works in three steps. The first step of the matching performs a robust estimation of the coarse tile translation parameters by minimizing the correlation-based distance between the satellite road tiles and the road distance image. The second step localizes potential GCPs and performs local minimization of matching distances. The third step is the super-fine tuning to the reference orthoimages or the intensity image. The output of this sub-module are GCPs equipped with pairs of image pixel coordinates and reference system coordinates, and attributes describing their quality.
• MODULE 5: Geometric modelling
The rational function geometric model (RFM) is used for the geometric modelling. RPCs are usually accurate but must be bias-corrected first. The correction is performed using the extracted GCPs. The GCPs are introduced into the model and the obtained residuals of each individual GCP are used to calculate the coefficients of two first order polynomials, which are later used to correct the pixel positions during orthorectification. The algorithm also uses two gross error detection techniques. The module supports the Universal Transverse Mercator (UTM) projection and user selected national coordinate reference systems.
• MODULE 6: Orthorectification
The orthorectification of images is performed with the indirect method. This method starts in the object space (orthoimage) and projects the pixels onto the input image. Relief displacement is corrected pixel by pixel by calculating the correct position of the pixel within the orthoimage using a DTM of the selected AOI and the estimated parameters calculated during geometric modelling. The grey value of the pixel is determined by the bilinear resampling of the neighbouring pixels on the input image.

The test AOIs were selected based on the availability of ancillary data in web databases and the availability of the optical satellite data from accessible sources. Based on these sources we selected AOIs in Slovenia, Denmark, Germany, the Netherlands, Scotland, Sweden, Mexico and Australia.

We assessed the accuracy and robustness of the procedure to handle different types of images with different imaging modes and spatial resolutions. Initially, the GCPs were generated without reference orthophoto data. In this case, only OSM roads were used as reference data to extract GCPs, and the processing was stopped at the fine matching step (thus skipping the superfine matching). The obtained results mostly show good accuracies for multispectral (MS) images with 2 m resolution. These are usually around 2 pixels. The results for panchromatic (PAN) images mostly range from around 3 pixels to more than 5 pixels. However, OSM roads proved to be an unreliable data source with variable quality and the accuracy achieved was mainly due to the inaccurate OSM layer.

When reference orthophotos were available, they were used for superfine matching, and the results were very good (usually below 2 pixels). When ALS intensity data is used as a reference instead, the results rarely improved. The most likely reason for this discrepancy is the difference in spectral information, because intensity data comes from a different type of sensor than the satellite images. Detection of reflected laser light is more dependent on the incidence angle and other factors and the resulting intensity data raster is more difficult to match with the optical image data. For this reason, superfine matching between these two types of data is rather difficult.

At this stage of development, the service supports processing of images in JPEG2000 or TIFF formats, which are common de facto standards. The procedure can process MS and PAN images with a resolution of up to 0.3 m. The service is already functional but needs some improvements before it can enter the production phase. The possible foreseen upgrades include:
- improve the security of the web application (e.g. hash for passwords, securing the APIs),
- enable the loading of more than one satellite image of the same area in one project,
- speed-up the whole processing (e.g. include GPU processing where possible),
- improve robustness of the orthoimage generation,
- update the description page of the project and better present the outputs, and
- add storage of user’s usage history.

Estimating the spectral properties of vegetation signals from spaceborne imagers has been a central task in remote sensing for more than half a century, however every mission has particular engineering constraints and therefore pressure for innovation. One such mission under development is TreeView: a smallsat mission that will measure vegetation reflectance at a sub 2.5m ground sampling distance with moderate spectral (8-band) and temporal resolution. The project is constrained by a tight budget (£15M) and a requirement for multispectral measurement of individual tree canopies with monthly repeat imaging across the UK and other target sites.

Core to the mission is the development of a data processing pipeline that allows reliable recovery of vegetation signals throughout the temperate spring-summer-autumn period. Retrieval of the quantity-of-interest, surface reflectance, is hampered by the spatial and temporal variation in surface structure and atmosphere, as well as the inherent noise properties of the imaging system. Coupled with this are the hardware engineering constraints, limiting the additional instruments that can be carried onboard. Many previous Earth observation multispectral missions have Short Wave Infrared (SWIR) bands that are integral to atmospheric correction routines (e.g. Sentinel 2, LandSat, etc.), however this is not available in the TreeView mission. To meet these challenges, we considered a range of approaches for different geographic regions over which the availability of complementary data (e.g. meteorological data, other satellite imagery datasets, etc.) varies widely.

In this talk we present a novel processing pipeline for moderate spatial and spectral resolution satellite imagery, with an emphasis on missions with only Visible and Very Near Infrared (VNIR) spectral coverage. We discuss the engineering considerations of the processing system and explore results from our validation study using both Sentinel 2 and simulated TreeView imagery across a range of urban, natural and semi-natural landscapes. Finally, we consider the implications for hardware specification and the deployment on different modern computing infrastructure.

In the frame of PICSCAR initiative, an exercise of intercomparison of intercalibration of OLI/L8 and MSI/S2A sensors over Libya 4 small site started three years ago. Three teams contribute to this exercise, and agreements were found between CNES, SDSU, and PICSCAR in 2018 to share the results. In 2019, this exercise moved towards an operational monitoring whose results are updated every 6 months and posted on the PICSCAR portal (https://picscar.magellium.com/)
Intercalibration results are provided for a period of 6 months and compared using the same metrics: mean and standard deviation for each semester since the launch of Sentinel 2 for the 7 common spectral bands, and the mean and standard deviation of the full period.
The portal offers the capability to display comparisons for selected bands, for selected periods, mean results over all dates and Tables. It allows to keep a track of the results and share them with scientific community. And finally, to understand how the underlying assumptions to each of the methods generates the differences between results.
All results obtained between July 2015 and December 2021 will be presented. The results, from these three teams, were derived individually using different techniques, however, the calibration results do agree within 1% for all 7 bands.
References
B. Berthelot and P. Henry, Monitoring the Intercalibration of L8/OLI with S2A/MSI over Libya4 PICS in the frame of PICSCAR CEOS/IVOS initiative, 4th Sentinel-2 validation team meeting, 15-17 March 2021
Morakot Kaewmanee, Esad Micijevic, Dennis Helder, Md Obaidul Haque, Julia Barsi, Radiometric Comparison of Sentinel 2A, Sentinel 2B and Landsat-8: Lifetime Trending, Cross Calibration and Absolute Calibration Assessment Over the Libya 4 PICS, JACIE Workshop 2018, College Park, Sep 17-19, 2018. https://calval.cr.usgs.gov/apps/sites/default/files/jacie/S2AS2BOLIJACIE2018v.1.6.pdf.
Lacherade, S., Fougnie, B., Henry,P.,& Gamet, P.(2013). Cross calibration over desert sites: Description, methodology, and operational implementation. IEEE Transactions on Geoscience and Remote Sensing, 51, 10981113, doi:10.1109/TGRS.2012.2227061

In this paper, the condition for maximum daily global coverage achieved by multiple independent wide swath missions is investigated. Spacecraft from different earth observations monitoring systems flying at different altitudes can form a multi-mission constellation to contribute with data acquisitions for research and applications, maximizing the scientific and coverage outcome with different satellites joint observations in a multi-mission constellation. Accordingly, the most important orbit design requirement, either in altitude or in phasing, would be to accomplish a daily global coverage with all the satellites in orbit, while allowing for regular cross-validation of the various systems.

The preliminary analysis of possible orbit definitions in a multi-altitude constellation will be presented. Besides the daily global coverage, the systems need also to be cross-validated to ensure a homogeneous data quality, which means collocations between the satellites to perform quasi-simultaneous observations over calibration areas.

Different constellations are studied. This work addresses the constellation design and computation of the metrics by qualifying how well the two, apparently conflicting, objectives (daily global coverage and sufficient quasi simultaneous observations) can be met from an orbit geometric perspective by a constellation of different missions.

BRDF estimations using a Kalman-Filtering approach and related normalisations of S2 L2 and L1C reflectances: radiometric comparisons with the RadCalNet and Landsat 8 measurements.

Saulquin B.(1), Savinaud M.(1)
(1) CS-Group, Toulouse, France

Calibration and inter-comparisons using different sensors require the estimation of the Bidirectional Reflectance Distribution Function (BRDF), i.e. the reflectance properties of the surface. Once estimated, the BRDF parameters are used to normalise the surface reflectances, i.e. positioning the sun and the sensor in a common reference allowing thus a ‘direct’ comparison with either the in-situ or with another sensor.

The inversion of the BRDF parameters is nevertheless often corrupted by the noise introduced by the atmospheric corrections. This typically leads to non-realistic time and space-varying BRDF coefficients. To address this inversion issue, we use here an approach initiated in the scope of the S2MPC based onto the Kalman filtering (KF).

We estimate the satellite-derived BRDF parameters starting from the Sentinel-2 (S2) Level 2 (L2) at 20m resolution and the Roujean models. Our BRDF parameter estimates are thus consistent in the space and time, thanks to the KF approach.
The derived S2 nadir-viewing normalised L2 reflectances show clearly less uncertainties compared with the RadCalNet ones and the non-normalised S2 L2 reflectances . We compare also our results with the ones obtained using the NASA BRDF parameters, i.e. the MODIS MCD43A1 product.

The satellite-derived surface BRDF parameters are then used in addition with the atmospheric conditions as input to radiative transfer simulations to obtain simulated TOA reflectances in all the geometries. The S2 L1C reflectances are then normalised (using the ratio of the TOA simulations) regarding the Radcalnet nadir-viewing convention. The S2 nadir-viewing normalised L1C reflectances show clearly less uncertainties compared with the TOA RadCalNet ones.

Finally, we generalise our L1C normalisation approach for the inter-sensor CAL/VAL applications. For this, we normalise two spectrally adjusted time series of S2 L1C to fit the Landsat 8 over bare soils and small vegetation. We show using the two time-series that the biases and the uncertainties between the two sensors L1C time series are reduced.

Our BRDF estimation, associated with reliable uncertainties, and the derived normalisations for both the L1C and L2A may be seen as a generalisation of the work performed by Bouvet (2014) with a time-varying BRDF model. The generation of reliable maps of BRDF parameters at 20m resolution derived using S2 could hence participate constructively to the European Space Agency’s (ESA) strategy to create continuous time series and increase the inter-sensors interoperability in the optical domain.


References
Saulquin, B. (2021). BRDF Estimations and Normalizations of Sentinel 2 Level 2 Data Using a Kalman-Filtering Approach and Comparisons with RadCalNet Measurements. Remote Sensing, 13(17), 3373.
https://www.mdpi.com/2072-4292/13/17/3373/htm
Bouvet, M. (2014). Radiometric comparison of multispectral imagers over a pseudo-invariant calibration site using a reference radiometric model. Remote sensing of environment, 140, 141-154.

ESA is building on almost 18 years of continuous Fiducial Reference Measurements (FRM) from UK-funded shipborne radiometers by establishing a service to provide historic and ongoing FRM measurements to the wider sea surface temperature (SST) community through an International SST FRM Radiometer Network (ships4sst). The ships4sst is open for partners around the world, currently compromising of partners from the UK (University of Southampton, Rutherford Appleton Laboratory, Space ConneXions), Denmark (Danish Meteorological Institute) and France (Ifremer) and will not only collect shipborne radiometer data but also use the data to validate satellite SST products.
The service provides not only FRM SST measurements, put also includes a long term data archive of the FRM datasets at Ifremer where the data will be stored in the ships4sst netCDF L2R format and provides a validation service based on the ESA felyx match-up database (MDB) hosted at EUMETSAT. The ships4sst data is freely available to anyone as are the validation results. At ships4sst we organise and participate in regular inter-comparisons at the National Physics Laboratory (NPL) in the UK and the National Institute of Standards and Technology (NIST) in the USA to ensure not only the SI (International System of Units) traceability of our remeasurements but also the validity of the per SST value uncertainties.
To demonstrate the value of the FRM SST service we will first show some examples of the ships4sst data around the world and second show the most recent validation results from SLSTR A and B from the ships4sst network regions. This will not only demonstrate hat SLSTR A and B are preforming to specification and at least as well as their predecessor AATSR, but also show a potential route for SI-traceability for SLSTR SST measurements. And finally we will show results from SLSTR A and B during the Sentinel-3 tandem phase using a number of comparison tools, including triple collocations between he ships4sst FRM data and the SLSTR units on Sentinel 3 A and B.

The OLCI (Ocean and Land Colour Imager) instrument on board the sentinel 3 satellites is a quasi-autonomous, self-contained, visible push-broom imaging spectrometer. It is a highly sensitive instrument, delivering multi-channel wide-swath optical measurements for ocean and land surfaces in 21 bands between 400nm and 1020nm. It consists of five identical modules, arranged in a fan shape, to cover the full range of view of 69°. The five OLCI cameras are each equipped with a CCD that is composed of 740 columns in the spatial dimension across the camera field of view and 520 rows in the spectral dimension. The bands are created by binning several rows (having a sampling distance of ~1.25nm and full width at half maximum of ~1.8nm). OLCI’s Spectral Response Functions (SRF) have been determined on ground and verified during the commissioning phases. One noteworthy spectral feature is a small spectral response drift of the CCD rows across the camera field of view, which is in the order of up to 1nm. The poster/presentation describes the additional temporal evolution of the spectral responses of the CCD rows occurring in-flight, and a model to quantify the temporal evolution.
OLCI’s spectral characteristics are regularly monitored in-flight using so called spectral campaigns. The campaignes use the programming capability of OLCI to define 45 bands around stable spectral features, to characterize the spectral dispersion of each camera system with respect to the spectral and the spatial (across track) dimension. Simulations of OLCI measurements in the 45 bands are optimized for best agreement with the spectral features, as a function of assumed bandwidth and band center wavelength of an individual CCD element. Depending on the used spectral feature the achieved accuracy for the centre wavelength is in the order of 0.1-0.2 nm, the precision (repeatability) is better than 0.05 nm. Various different calibration sequences have been performed in the commissioning phases. The so-called S09 campaign is regularly performed on OLCI A/B approximately three times per year. In S09 the 45 bands are grouped around the atmospheric oxygen absorption band at 760 nm and around distinct solar Fraunhofer lines at 485nm, 656 nm and 854 nm. The same few hundred frames are acquired at the same orbit cycle (number 24), belonging to Libyan desert.
The spectral characterisation is applied to all CCD-columns (across track), for all spectral features and for all cameras. The regularity of the spectral campaigns allows a precise quantification of the temporal evolution of the spectral response for each individual CCD-element on each camera CCD, at least for the investigated spectral features. It emerges that all cameras show a tiny but distinct evolution. Both, the single CCD-row bandwidth and the across track variability (‘smile’) of the central wavelength remain almost constant for all cameras of OLCI A and B, but the central wavelengths of all pixels move almost homogenously with a decreasing rate. Since launch, four of the five cameras of OLCI A and B respectively, have drifted up to 0.3 nm towards longer wavelengths. One camera (camera 5 for OLCI A and camera 3 for OLCI B) has drifted by 0.3 nm to shorter wavelengths.
Based on these observations and under the assumption that they hold for the full spectral range of each camera, the characteristics of all OLCI bands (which are binned CCD-rows) have been recalculated for each date of an inflight spectral characterization. The methodology used, is exactly the same as it has been used to establish the band characteristics of OLCI-A at the end of the commissioning phase and of OLCI-B using preflight; the only difference is the spectral position of the CCD elements according to the inflight spectral characterization. Eventually, the resulting band characteristics of all S09 campaigns form a temporal model with the orbit-numbers as the sampling points. The model is published in form of look up tables. They enable the calculation of the best estimate of centre wavelength and bandwidth for all bands and all pixels at a given orbit number.
We will present details on the compilation and the usage of these look up tables.

Spectral-based models extracted from laboratory reflectance across the 400–2500 nm spectral region to predict soil attributes may not be applicable to soil spectra acquired in the field. This is because laboratory sampling procedures disturb the natural soil surface's status. This gap is critical for a practical utilization of the many Soils Spectral Libraries (SSLs) available today worldwide for remote sensing applications. This issue was investigated by using the soil surface-dependent property namely Water-Infiltration Rate (WIR) to demonstrate how traditional SSLs are not adequate to model soil surface-dependent properties. To this end, we created a dataset with 114 samples collected from six fields with varying textures located in three different Mediterranean countries (Israel, Greece, Italy). For each sample, the WIR and the field spectral reflectance were measured simultaneously using adequate instruments: The SoilPRO® assembly connected to an ASD spectrometer and a minidisk infiltrometer (METER Group Inc., Pullman, WA, USA). Using the field and laboratory spectral datasets, the results showed that the WIR property is better predicted by field vs. laboratory measurements (R2 = 0.92 and 0.56, respectively). Based on this observation we then developed a transfer function (TF) to predict the field spectral response from the laboratory spectral data. Adopting the TF-processed dataset considerably improved the WIR prediction accuracy using laboratory information (from R2 = 0.56 to 0.76) and demonstrated that it is possible to find a TF that will permit an adaptation of laboratory SSL to predict the field’s soil surface condition. It was concluded that the SSLs can be better used to predict soil surface-dependent properties (e.g., WIR) that are affected by the soil seals, salinity, dust, and aggregation. The approach presented in this study paves the way to adapt the existing large SSLs available today for remote sensing of soil surface properties from one hand and for validating aerial/satellite remote sensing reflectance output on the other.

The H2020/HYPERNETS project is preparing the next generation of hyperspectral radiometric validation instruments, systems and networks for validation of water and land surface reflectance derived from satellite missions. To this aim, a new automated hyperspectral radiometer, the HYPSTAR®, has been designed and is being tested to provide fine spectral resolution (3nm FWHM for the range 350-1100nm) and high quality measurements.
Three Italian sites are being established as part of the HYPERNETS water validation network: a prototype HYPSTAR® system was installed on the Acqua Alta Oceanographic Tower in May 2021; other two will be installed in Lake Garda in December 2021, and close to the Lampedusa Island in the spring of 2022.
The Aqua Alta Oceanographic Tower is located in the North Adriatic Sea, 15 km off the coast of the Venice lagoon. The waters range from oligotrophic to mesotrophic depending on the local circulation. The water types vary from clear waters to moderately turbid waters due to wind and wave driven re-suspension or coastal currents, as the site is not reached directly by the river plumes under flood conditions.
Lake Garda is a large deep Italian lake characterized by clear oligo/meso- trophic waters and coastal areas colonized by submerged macrophyte beds. The HYPERNETS site will be located in the south-western basin on the “Sesarole” navigational pylon close to Manerba.
The Lampedusa Oceanographic Observatory is an elastic beacon located about 5 kilometers South West of the island of Lampedusa, south of the Strait of Sicily. The ocean depth at the location is 74 m in blue oligotrophic waters.
The site characterization in terms of existing infrastructure and data environmental characteristics will be presented and the site preparation and installation will be described. An overview of the HYPSTAR® data acquired at the Italian sites will be followed by a matchup analyses of the available in situ data with water reflectance data from various satellites, including Copernicus missions Sentinel-2 and Sentinel-3 and the ASI/PRISMA hyperspectral satellite data.

Land Surface Temperature (LST) is an essential climate variable (ECV) which yields critical information about the Earth’s radiative energy budget and helps to constrain climate models, as well as providing information about temperature changes in remote regions. Satellite LST datasets are required to have a spatial resolution of < 1 km and a measurement uncertainty of < 1 K to meet WMO Global Climate Observing System (GCOS) requirements. The Sea and Land Surface Temperature Radiometer (SLSTR) aboard the Sentinel-3 satellites A (launched in 2016) and B (launched in 2018) are capable of producing such datasets, but require rigorous ground-based validation to confirm this.

Absolute validation of satellite LST is only possible via comparisons with in-situ observations of thermal radiation. A well-established suite of measurement sites exist as part of long-term monitoring networks (e.g. ARM, SURFRAD), and are routinely used for validation of SLSTR within the Sentinel-3 Mission Performance Centre (S3MPC) and will be for its successor the Optical Mission Performance Cluster (OPT-MPC). These sites do not however fully account for all possible biomes on the Earth’s surface, as mosaic vegetation and broadleaf deciduous forests are not represented by existing measurement sites. Therefore, to increase the scope of Sentinel-3 validation requires the deployment of new measurement sites, in addition to the comparisons with existing monitoring networks.

The “Copernicus Space Component Validation for Land Surface Temperature, Aerosol Optical Depth and Water Vapour Sentinel-3 Products” (LAW) project is performing an extensive and systematic validation of Sentinel-3 datasets against ground-based observations through calibration, instrument deployment and subsequent validation matchups for 5 new LST observation sites in previously unobserved biomes:

- KIT forest (Germany): closed broadleaved deciduous forest
- Svartberget (Sweden): open needleleaved deciduous or evergreen forest
- Hyytiälä (Finland): closed to open mixed broadleaved and needle leaved forest
- Robson Creek (Australia): closed to open (more than 15 %) broadleaved evergreen and/or semi-deciduous forest
- Puéchabon (France): sparse vegetation


This presentation will cover the progress made in deploying the new stations, initial comparisons between Sentinel-3 and ground-based data, and potential consequences for refining the Sentinel-3 retrieval algorithm based on these analyses. It will also include results from operational validation of SLSTR made within S3MPC and OPT-MPC.

Since March 2021, the Sentinel-2 Global Reference Image (GRI) is used to improve the geolocation of Sentinel-2 images. Simultaneously, a new Digital Elevation Model, the Copernicus DEM, was introduced in the orthorectification step. In this presentation, we provide for the first time the assessments of the geometric performance of the geometrically refined images.

The Sentinel-2 GRI is a database of single-band (red band B04) L1B products (in sensor geometry). The viewing models of the L1B products have been improved by a global spatio-triangulation using a set Ground Control Points. For each new L1B product, the geometric refinement algorithm finds overlapping L1B products from the same repeat orbit in the GRI and aligns them to a common geometry. Tie-points are found between the GRI products and the current product, and shifts are computed for each matching tie-point. A refinement of the viewing model is computed to minimize these shifts. The correction applied are first order polynomials for the roll, pitch and yaw angles, and a constant homothety correction. The improved viewing model is then used for the orthorectification step at L1C.

The performance assessment relies on two different approaches: comparison with independent ground control points, and a posteriori comparison with projected GRI images. Both assessments show the very significant improvement of the geometric uncertainty after geometric refinement, both in terms of absolute geolocation (from 12 m to < 7 m) and multi-temporal relative geolocation (from 11 m to < 6 m). We also report a slightly better performance for Sentinel-2B than Sentinel-2A and investigate possible explanations.

The presentation will also analyze from a statistical point of view the corrections estimated by the geometric refinement process. The profiles of the corrections are noticeably consistent from orbit to orbit and across satellite units, most probably due to deterministic thermo-elastic effects.

However efficient and reliable, the refinement algorithm cannot work in all situations. For example, when very cloudy products with few tie-points are refined, small shift estimation inaccuracies or false matches can lead to an incorrect refinement. To detect such cases, a set of refinement quality indicators are computed and a “fall-back” to the unrefined product is triggered whenever a flag is raised. We will present some statistics on the occurrences of such fall-back cases.

In 2021 the University of Zurich (UZH) carried out a Europe-wide airborne hyperspectral imaging campaign in cooperation with NASA’s Jet Propulsion Laboratory (JPL) on behalf of the European Space Agency (ESA) and NASA to support the upcoming CHIME and SBG satellite missions. After a planning phase of two years the sensor of choice, namely AVIRIS-NG, was operated from mid-May until the beginning of August out of Switzerland under challenging operational conditions due to the COVID-19 pandemic, unfavorable weather situations and other constraints.
To support the two future satellite missions, different targets were acquired to image a wide range of different land surface types like forests, peatland, mines, coastal regions, and lakes. The targets were spread all over Europe, from Scotland to Spain to Romania.
Amongst the most stringent conditions was the necessity to coordinate an AVIRIS-NG imaging mission with a PRISMA overpass in a predefined time window. Given the weather conditions in Europe this summer, it was a real challenge to coordinate the missions with the field teams sent to collect spectroradiometric in-situ data which is being used for calibration and validation (CAL/VAL) purposes of the airborne data. The measured in-situ data is stored in the in-house spectral data base SPECCHIO, from which it is retrieved in the CAL/VAL process. The entire coordination of the airborne imaging campaign was executed from the home base in Switzerland, where a mission operation control center (MOCC) was set up. In mission control five days at a time were planned. This was done using an automated MATLAB script in which the weather information, PRISMA overpasses and other constraints were processed into a target decision matrix indicating which sites appeared to have the best probability of being successfully imaged. After this automated pre-selection of targets, a more detailed weather analysis was carried out before the mission targets were communicated to the principal investigators of the respective mission site.
After completion of the campaign, the data was processed by NASA/JPL and shipped to UZH where the CAL/VAL analysis was carried out using an in-house CAL/VAL tool. The latter compares the in-situ data with the airborne data at reflectance level in an automated way and computes descriptive statistics. The data acquired during these missions is made accessible to a broader user community on the campaign web page https://ares-observatory.ch/esa_chime_mission_2021/.

The hyperspectral instrument "DLR Earth Sensing Imaging Spectrometer" (DESIS) is a VNIR sensor
on-board of the International Space Station (ISS) and operational since October 2019.

DESIS acquires images of Earth on user request with a swath of about 30 km and 235 bands with a Full Width at Half Maximum (FWHM) of 3.5 nm in the spectral range of 400 to 1000 nm.

The DESIS Ground Segment L2A processor corrects the at sensor received terrestrial reflection of the incident solar radiation from the effect of the atmospheric constituents.
Implemented within the L2A processor, the PACO atmospheric correction software processes ortho-rectified Top-Of-Atmosphere (TOA) radiance scenes and generates the Bottom-Of-Atmosphere (BOA) ground reflectance spectral image cube, together with pixel-classification masks, Aerosol Optical Thickness (AOT at 550 nm) and Water Vapor (WV) maps.

In this contribution we present the validation of the DESIS atmospheric (AOT550 and WV) and spectral products (BOA reflectance) using independent in-situ measurements.
The aerosol optical thickness and water vapor is compared with
the Aerosol Robotic Network (AERONET) measurements and
the surface reflectance is validated with the Radiometric Calibration Network (RadCalNet) data.

The validation results are expressed in terms of uncertainty, which we propose to be used as the DESIS error estimation.

The uncertainty function is provided for water vapor and surface reflectance. The lack of scenes with high AOT values allows only to determine a global RMSE for all available values of AOT.

The current shortage of diverse ground surface in-situ measurements publicly available also makes it impossible to provide a complete and random study of the ground reflectance uncertainty.
Within the data available (RadCalNet) an uncertainty of the ground reflectance is provided for a limited range of observation conditions.

The hyperspectral nature of the DESIS sensor allows the estimation of the water vapor uncertainty by using the 820 nm absorption region in the Atmospheric Pre-corrected Differential Algorithm (APDA). Results show a lower uncertainty in hyperspectral data when using several bands in the algorithm and with a shorter wavelength distance between them.

The hyperspectral sampling range starting at 400 nm also opens the possibility to define new non-linear models to express the uncertainty of the ground reflectance, rather than a linear dependency with the surface reflectance. The DESIS BOA validation results show that the approach of a linear dependency is not sufficient to represent the influence of the AOT uncertainty in the BOA uncertainty.

Validation is a necessary means for quantifying the accuracy of satellite data and their uncertainties by comparison with reference data. For water applications, AENONET-OC provided the first widely applied validation base for water remote sensing analysis, and a broad validation community was established. For land application, the RADCALNET has shown its effectiveness in calibrating and validating optical satellite data so far; however, it has limited spectral resolution directional reflectance (e.g., the RADCALENT system provides only 10 nm spectral resolution nadir viewing reflectance). In other words, the current validation data provided by the RADCALNET are only multispectral and limited in viewing geometries and only a few locations for land remote sensing reflectance.
Land-Hypstar® is a new low-cost developed system that provides high spectral resolution directional reflectance validation data for current and upcoming optical orbital missions with Visible-Near-infrared (VNIR) and Short-Wave-infrared (SWIR) spectral bands. Land-Hypstar® can collect data from 380nm till 1020 in VNIR and 1000-1600 nm, in the SWIR spectral range with 5 to 10 nm spectral resolution.
The main objective of this study is to present the initial results from the Land-Hypstar® on land products by comparing the Land-Hypstar® data versus surface reflectance extracted from several orbital multispectral sensors such as Sentinel-2, Sentinel-3, Landsat-8, and MODIS. The first Land-Hypstar® has installed in the middle of an agricultural field, Demmin, Germany, and is collecting surface reflectance from dynamic surfaces like soil and vegetation. The Land-Hypstar® data are collected daily from 10 am – 5 pm UTC every 30 minutes automatically. Satellite data products are downloaded directly from the data providers' repository, for instance, Sentinel-2 and Sentinel-3 downloaded from the ESA Sentinels Scientific Datahub (https://scihub.copernicus.eu/), and Landsat-8 and MODIS downloaded from Earth Explorer (https://earthexplorer.usgs.gov/). These images corresponded to Level-1C (L1C) radiometrically and geometrically corrected Top Of Atmosphere (TOA) products. Afterward, all satellite products were atmospherically corrected to the Bottom Of Atmosphere (BOA) level using a different processor (e.g., Sen2Cor, iCor, ACOLITE, and LaSRC). Finally, the matchups between surface reflectances from different processors and Hypstar® data were generated. The initial results from matchups indicate that surface reflectance from multi-mission satellites is close to in-situ data from Hypstar®.

Quality assessment of surface reflectances provided to end-users through L2A processors is a prerequisite for most optical remote sensing quantitative applications. In an attempt to cross-compare these processors, the Atmospheric Correction Inter-comparison Exercises, ACIX I and the later ACIX II, jointly organized by ESA and NASA, mostly rely on a reference surface reflectance dataset computed with the combined use of AERONET atmospheric observation and the 6S radiative transfer code, and marginally on in-situ observations. Although such approach allow for a large number of reference sites located worldwide (127 sites in ACIX-II), these reference reflectances remain indirect estimates, and may suffer from methodological bias. In a continuous effort to provide high-quality observations of surface reflectances for the calibration of optical sensors and the validation of the MAJA (MACCS-ATCOR Joint Algorithm) cloud screening and atmospheric correction processor, the French Space Agency (CNES) together with the Center for the Study of the Biosphere from Space (CESBIO) implemented the Robotic Station for Atmosphere and Surface characterization (ROSAS) methodology. ROSAS forms a consistent framework of both in-situ instrument and processing software dedicated to optical sensors calibration and validation, with a specific focus on surface reflectance and bi-directional reflectance function (BRDF) retrieval. Initially deployed in an homogeneous landscape composed of grass and pebbles in the South of France (La Crau site) in 1997, ROSAS benefited from a second site in an arid desert in Namibia (Gobabeb site) in 2017 within a collaboration with ESA. Both stations provide data to RadCalNet CEOS network. With their overall land cover homogeneity and low cloud coverage, these two sites are perfect matches for calibration purposes, but offer a rather limited range of surface reflectance values, either on seasonal or an inter-annual basis. In order to address this need for a wider range of surface reflectances values for L2A validation purposes, CESBIO and CNES jointly deployed and integrated in 2021 a new ROSAS station over a farmland area belonging to the Purpan Engineering School, in the South-Western France (Fr-Lam experimental site). Located within a 23 hectares field cultivated with a 2-years crop rotation, this site offers a strong inter-seasonal and inter-annual variability of land cover and canopy structure, ranging from bare soil to full winter wheat and maize growing and senescent stages. Besides, the nearby forest brings opportunities to assess adjacency effect correction algorithms. After some highlights on the limitations of indirect estimates of surface reflectances used in ACIX for validation purposes, this poster presents a detailed view of the new Fr-Lam ROSAS station, as well as observed surface reflectances, BRDF and a validation of MAJA L2A products over the 2021 maize growing season.

The Optical Mission Performance Cluster (OPT-MPC), one of the Copernicus Mission Performance Clusters set-up by the European Space Agency, is the follow-on of the activities carried out in the S2 MPC and the optical part of the S3 MPC. Instead of being organized per mission, this second generation of MPCs is structured around the type of sensors.

The novel approach has several advantages:
- This allows to exploit the synergies across the Sentinel missions which have similarities in terms of instrument, acquisitions and spectrum.
- The experts from the two missions can reinforce their expertise by benefiting from knowledge and shared experience of the other missions.
- Validation activities can naturally be shared, considering that the same in-situ and FRM data benefit to the validation of S2 and S3 products
- Beyond the needs of S2 and optical S3 missions, setting-up a common working frame, procedures and processes, sharing results, will faciliate the inclusion of future Copernicus missions of the same type, such as CHIME and LSTM.

Despite this new high-level organisation, the scope of activities to be done in the OPT-MPC does not differ from the MPC activities carried out since the launch of S2A and S3A and the main objectives are still:
- The assessment of performance of the S2 and S3 optical missions and core products delivered by the ground segment.
- The quality control of products (off-line detection of anomalies)
- The maintenance of the processing baselines, by updating as necessary the calibration and/or the processing chains, to meet the mission requirements of each mission.
- The evolution which would improve the quality of products or even would add important information layers in the product (e.g., the uncertainties)
- The support during the C and D units commissioning phases.

The OPT-MPC relies on two important components, which are strongly linked together: the operational component and the expertise component.

The operational component relies on a specific architecture which host the Expert Collaborative Environment (ECE) and provides key functionalities:
- To access to Sentinel products and in-situ data
- To prepare and verify new calibration or new algorithm evolution
- To process Test Data Sets (TDS) and perform cal/val reprocessing
- To share information and results between the OPT MPC experts.

As part of the operational component, the operators perform the offline quality control of S2 and S3 products, they record all the events of the missions and they provide support to all experts (to prepare a new calibration, to do an in-depth analysis in case of anomalies).
The expertise component is composed of all the organisations (industry, research lab, university, technical centre) providing a team of highly-qualified experts who are involved in calibration and validation activities, the assessment of the mission performance and of the products (e.g., analysis of trends, ad hoc analysis of anomalies, etc.)

Thus, after a few months of operations, we will briefly present all the activities done by the OPT-MPC which include:
- The QC activities done by the operational component
- The Monitoring and assessment of instrument performances, and the reporting documents which summarize the analysis done
- The calibration and validation activities done on L1 and L2 products
- The maintenance and evolutions made on the L2 operational processing chains (IPFs)
- The support made to the PDGS coordination desk
- Information (reports, Sentinel Online, …) prepared by the OPT-MPC to help the users.

Improving the modeling of adjacency effects for MAJA high resolution atmospheric corrections

Authors: M. Lassalle¹, M. Moulana², D. Ramon², J. Colin¹, O.Hagolle¹,³

¹ CESBIO, UMR 5126 (CNES/CNRS/INRAE/IRD/Université Toulouse III), Toulouse, France
² HYGEOS, Lille, France
³ CNES, Toulouse, France

emails :
M. Lassalle : micael.lassalle@cesbio.cnes.fr
M. Moulana : mm@hygeos.com
D. Ramon : dr@hygeos.com
J. Colin : jerome.colin@cesbio.cnes.fr
O. Hagolle : olivier.hagolle@cnes.fr

Abstract:

The use of optical time series of images such as those provided by Sentinel-2 or LANDSAT requires several pre-processing steps to provide temporally consistent surface reflectances : detection of clouds and their shadows, aerosol optical thickness and water vapour estimation, atmospheric effect corrections. One of the atmospheric effects, called « adjacency effect », is due to scattering by aerosols and molecules and leads to somewhat blurred images recorded by satellites at the top of the atmosphere. In some cases, adjacency effects account for the same order of magnitude as the surface reflectances. Consequently, most atmospheric correction processors should include a correction of adjacency effects. The MAJA processor [1], developed by CNES and CESBIO, with some modules provided by DLR, already includes an adjacency effect correction module. It is based on a convolution made with a Gaussian kernel, independent of the observation geometric conditions and of the state of the atmosphere, which is a coarse approximation.

The objective of our study is to provide a better modeling of the adjacency effect and to implement it within MAJA. For some tests we decided to use the model provided by Miesch et al. [2], which brings additional complexity as compared to the 5S approach [3] since it relies on two independent point spread functions that respectively correspond to diffuse transmission in the atmosphere and to successive scattering between earth and atmosphere, where 5S only uses one point spread function. The 3D radiative transfer code SMART-G [4] is used to perform the reference simulations. SMART-G code (based on Monte Carlo statistical method) uses GPU resources to dramatically reduce simulation times, while providing a sufficient amount of photons to compute this second order effect.

One simulation with SMART-G directly provides the point spread function corresponding to diffuse transmission in the atmosphere, while a second simulation provides the reflectance implied in successive scattering between earth and atmosphere. From these simulations, we can identify all the parameters of the Miesch et al. model to estimate the second point spread function.

With the two available point-spread functions, Miesch et al. provide an iterative algorithm which retrieves the surface reflectance for a given top of atmosphere reflectance, which can be compared to the simulations provided by SMART-G. In order to use this algorithm in operational conditions within MAJA, both point spread functions have to be given as analytical functions depending on all parameters (observation geometric conditions and state of the atmosphere). This poster aims at presenting both point spread analytical functions and their dependence on the geometrical and atmospheric conditions, and provides the results of Miesch et al. algorithm to retrieve surface reflectances from given top of atmosphere reflectances.


Keywords: atmospheric corrections, adjacency effects, atmospheric point spread functions, top of atmosphere reflectance, surface reflectance


[1] O. Hagolle, M. Huc, C. Desjardins, S. Auer, & R. Richter. (2017). MAJA Algorithm Theoretical Basis Document (1.0). Zenodo. https://doi.org/10.5281/zenodo.1209633
[2] C. Miesch, L. Poutier, V. Achard, X. Briottet, « Direct and inverse radiative transfer solutions for visible and near-infrared hyperspectral imagery », EEE Transactions on Geoscience and Remote Sensing 43(7):1552 – 1562, DOI:10.1109/TGRS.2005.847793 (2005)
[3] E. Vermote, J-C. Roger, L. Remer, « Atmospheric correction of visible to middle-infrared EOS-MODIS data over land surfaces: Background, operational algorithm and validation », Journal of Geophysical Research Atmospheres 102(D14):17131-17141, DOI:10.1029/97jd00201 (1997)
[4] D. Ramon, F. Steinmetz, D. Jolivet, M. Compiègne, R. Frouin, « Modeling polarized radiative transfer in the ocean-atmosphere system with the GPU-accelerated SMART-G Monte Carlo code », Journal of Quantitative Spectroscopy and Radiative Transfer
Volumes 222–223, Pages 89-107, (2019).

Cloud screening in satellite retrievals is often the most critical part for obtaining reliable aerosol optical depth (AOD) product. In case cloud screening is not strict enough, cloud contamination may cause unrealistic AOD increase, while in case it is too strict, some of cloud-free pixels is wrongly rejected. Thus, the evaluation of cloud screening is an important component in the AOD product validation. It helps a product developer to reveal conditions when cloud screening is not working properly.
Automatic globally distributed Aerosol Robotic Network (AERONET) for systematic ground-based sun photometer measurements of aerosol optical depth provide measurements of natural and anthropogenic aerosol loading, which is important in many local and regional studies as well as in global change research investigations. The AERONET Version 3 (V3) algorithm provides fully automatic cloud screening and instrument anomaly quality controls.
Aerosol Optical depth provided by AERONET (aAOD) is widely used for validation of satellite AOD (sAOD) products. For satellite AOD validation, pair of collocated satellite retrieved and measured on the ground AOD values should exist. For evaluation of the cloud screening in the satellite retrievals, two cases should be considered: for collocated in time and space pairs, sAOD is not retrieved and aAOD is measured (case 1) or sAOD is retrieved and aAOD is not measured (case 2). If sAOD and aAOD fit to case 1, the conclusion can be done that satellite cloud screening is too strict; if sAOD and aAOD fit to case 2, that means that satellite cloud screening is too relaxed and satellite product may include cloud contaminated pixels.
The proposed approach has been applied for evaluation of cloud screening in SY_2 retrieval algorithm. We analysed collocations of SY_2 AOD product with AERONET stations for the period of 01.2020-09.2021 and revealed cases when one of two products, either sAOD or aAOD, was not available. By collocations we intend 11*11 satellite pixels areas around AERONET stations in ±30min time window around satellite overpass.
A thorough check of flags provided in the SY_2 AOD product was performed to recognise and exclude from the analysis cases, when sAOD was not provided for other than cloud contamination reasons. S3A operational anomalies have been considered.
ACKNOWLEDGEMENTS
This work is supported by the ESA/Copernicus LAW project.

The Copernicus program is a European initiative for the implementation of information services dealing with environment and security, mainly based on observation data received from Earth Observation (EO) satellites. In the frame of this program, ESA launched the Copernicus Sentinel-2 and Sentinel-3 optical imaging satellites. These satellites deliver a new generation of optical data products designed to directly feed downstream services mainly related to land and ocean monitoring, emergency management and security. To ensure the highest quality of Earth Observation data from the Copernicus Sentinel-2 and Sentinel-3 optical missions, ESA set up the Optical Mission Performance Centre (OPT-MPC) supported by several Expert Support Laboratories (ESL). The OPT-MPC brings together the experience of the former separated S2 and S3 Mission Performance Centers.
The main objective of the working group on atmospheric correction within OPT-MPC is to employ synergies across the missions on surface reflectance (SR), aerosol optical thickness (AOT) and water vapor (WV) validation. Harmonization of validation practices, approaches, metrics, terms and definitions will support inter-operability among missions. This can also result in a guideline for good validation practice for other optical remote sensing missions in the Copernicus program and worldwide. Eventually it will give recommendations for algorithm improvements.
This presentation reports first results of this process and will show some examples for consistent validation of Copernicus Sentinel-2 and Sentinel-3 optical missions. It invites for discussion about validation methodologies and standards in the community.

Atmospheric aerosols play a crucial role in climate change and represent one of the largest uncertainties in the understanding of the climate systems. Validation of satellite products is an essential activity. Without adequate validation the level 2 retrieval methods, product algorithms, and geophysical parameters derived from satellite measurements, satellite products cannot be used with confidence.
ESA/Copernicus LAW project is focused on the validation of SY-2 AOD products retrieved from the SLSTR and OLCI instruments onboard Sentinel-3A and Sentinel-3B satellites.
Validation work involved:
 Collection of the ground-based AOD data worldwide; matching satellite product with ground-based measurements
 Validation of the
 AOD
 AOD uncertinties
 Angstrom exponent
 Fine mode AOD
 Inter-comparison with similar satellite products: MODIS, MISR
 Recommendations to product developers
Sy-2-AOD product from both Sentinel-3 satellites, S3A and S3B, was validated against AERONET, SURFRAD, SKYNET and MAN ground-based networks. We evaluated validation results as a function of sun/satelite geometry, retrieved surface reflectance, cloudiness, revealed regional and seasonal differences in algorithm performance.
Validation results, in general, are often similar for S3A and S3B or slightly better for S3B. Validation results for the Southern Hemisphere (SH) are slightly better. However, the number of validation points in the SH is considerably lower than in the NH. Thus, global validation statistics are strongly biased to the validation statistics in the NH. AOD offset is increasiung with increasing solar zenith angle. For more details, see LAW validation web portal https://law.acri-st.fr/validation-results

ACKNOWLEDGEMENTS
This work is supported by the ESA/Copernicus LAW project.

Validation of satellite-derived products, in particular water reflectance, using in situ data is essential to ensure the quality of derived parameters useful for water quality monitoring, like turbidity and chlorophyll-a concentration. The use of autonomous systems, like AERONET-OC, has shown to be effective for increasing the number of validation match-ups compared to oceanographic cruises. However, the multispectral nature of the CIMEL instrument used in the AERONET-OC network prevents its use to validate hyperspectral missions, such as PRISMA (Hyperspectral Precursor of the Application Mission) and upcoming EnMAP (Environmental Mapping and Analysis Program) and CHIME (Copernicus Hyperspectral Imaging Mission for the Environment), as well as missions with different spectral band configurations, like Sentinel-2.
Within the H2020/HYPERNETS project, a new generation hyperspectral radiometer, the HYPSTAR® instrument, with a pointing system and auxiliary sensors has been designed and is being tested in different test sites to provide water reflectance at fine spectral resolution (3nm FWHM) in the 350-1100nm region with high quality at lower cost. This system has been installed in one of the test sites in Argentina, located in the turbid waters of Río de la Plata upper estuary, 60 km south of the city of Buenos Aires, an ideal scenario to test atmospheric correction algorithms performance.
The quality of the match-ups at a test site will depend on how representative is the point radiometric measurement at the site compared to the satellite pixel of a given size which will be influenced by the spatial homogeneity of the water body and also by the proximity of the site to land (adjacency effect).
In the present study, we describe and assess the quality of the HYPERNETS water site located in the Río de la Plata through the analysis of satellite-derived time-series of turbidity and its spatial variability (homogeneity) and the influence of land using the time series of high spatial resolution L8/OLI and S2/MSI data extracted at the site location. Furthermore, the first HYPSTAR® data and match-ups of different satellite systems are shown and analyzed showing its potential to validate satellite products in a multi-mission perspective.

This work is an alternative to the classical uncertainty assessment of the ocean colour radiometry (OCR) based on the direct comparison of satellite derived marine reflectance with concurrent in situ radiometric measurements collected at sea level. When dealing with the standard atmospheric correction (AC) algorithm employed on an operational basis by the Space agencies, the uncertainty of the marine reflectance in the visible (VIS) part of the solar spectrum originates from two sources: (i) the uncertainty of the level-1 (L1) calibration of signal measured at top of the atmosphere (TOA), which also includes the uncertainty of the vicarious calibration, and (ii) the uncertainty of remote-sensing aerosol algorithm and the AC. The level-2 (L2) aerosol product, usually comprising the aerosol optical thickness (AOT) at a given near-infrared (NIR) wavelength and its spectral dependence (Angstroem exponent), cannot be validated only with in situ radiometric measurements of the solar extinction. This validation also requires a good determination of the atmospheric scattering functions (aerosol path radiance and transmittance) needed to perform the AC of remotely sensed data. The aim of this work was to assess the uncertainty of these atmospheric scattering functions through the use of the AErosol RObotic NETwork (AERONET). The latter provides a worldwide database with inherent optical properties (IOPs) of aerosols, extracted from inversions of ground-based radiometric measurements (i.e. solar extinctions and sky diffuse radiances). Using these IOPs as inputs to a radiative transfer tool allows the estimation of the atmospheric scattering functions at the time of sensor’s overpass. This study was conducted for the MEdium Resolution Imaging Spectrometer (MERIS on-board the ESA/ENVISAT platform), with the L2 data derived from the third reprocessing over three AERONET sites selected in the European seas (Aqua Alta Oceanographic Tower, AAOT, Gustav Dalen Tower, GDT, and Lampedusa) and one in Pacific ocean as a reference (Lanai). An uncertainty budget based on the statistical analysis of the relative errors between MERIS and AERONET-derived atmospheric scattering functions was developed following the Guide to the expression of Uncertainty in Measurement (GUM). At 443 nm, for both the four stations, the combined standard uncertainty of the atmospheric path reflectance is estimated around 2 to 3%, while that of the total atmospheric transmittance is observed around 1 to 2%. However, the impact of these uncertainties on the OCR radiometry is largely varying between sites, due to the distinctive contribution of the marine signal to the total radiometry acquired at TOA. Indeed, the median values of the relative uncertainty of the marine reflectance at 443 nm range from 17.1% at Lanai to 88.3% at GDT (with a significant intra-site variability), with 24.1% at Lampedusa and 31.2% at AAOT. Clearly, these results demonstrate quantitatively the importance of the AC, for which the current performance does not allow to reach the OC mission requirements. Moreover, they stress the importance of deriving a pixel-by-pixel uncertainty budget, instead of providing overall estimates as generally done in the OC community. Finally, this study confirms that oligotrophic waters under clear atmospheres, as observed at the Lanai’s station, are strictly required for performing a system vicarious calibration (SVC).

The Copernicus Hyperspectral Imaging Mission for the Environment (CHIME) is one of the High-Priority Candidate Missions (HPCM) endorsed by European Space Agency for the expansion of the Copernicus Sentinel missions. CHIME will provide routine hyperspectral sampling of Earth surface reflectance over the solar spectral range (400-2500 nm) at a 30 m spatial resolution with a revisit of 22(11) days with one (two) satellite(s). CHIME observations will support EU- and related policies for the management of natural resources and assets providing a major contribution in the domains of raw materials and sustainable agricultural management with a focus on soil properties, sustainable raw materials development and agricultural services, including food security and biodiversity.
Currently under phase B2 the development of the CHIME mission is performed by a consortium led by Thalès Alenia Space in France (as prime contractor) and OHB System AG in Germany (for the instrument). The Observation Performance Simulator (OPSI), is a software tool being developed by ACRI-ST under the management of the above partners as ATBD providers, to support the development and verification of the space segment and the development of the ground segment.
OPSI is devoted to simulate the instrument acquisition and its different acquisition modes (along with the platform behaviour), to prototype the corroborating ground segment processors which calibrate the payload measurements to TOA radiance (at L1b) and orthorectified TOA reflectance (at L1c) and to assess the instrument performance by comparing true and estimated parameters generated at different stages. In order to accomplish the above objectives OPSI is composed of an Instrument Performance Simulator (IPS), a Ground Processor Prototype (GPP) module and a Performance Assessment Module (PAM).
This presentation is dedicated to the status of the OPSI Software System at its Preliminary Design Review, for which it offers to the leading consortium and to ESA a software tool able to simulate the radiometric aspects as well as to provide first mission performance figures to be expected from the instrument design. In particular we focus on the simulation of measurement artefacts as well as on the tracing of data quality within the ground processing.

Sen2Cor is a Level-2A (L2A) processor whose main purpose is to correct mono-temporal Copernicus Sentinel-2 (S2) mission Level-1C (L1C) products from the effects of the atmosphere in order to deliver radiometrically corrected Bottom-of-Atmosphere (BOA) data. Byproducts are Aerosol Optical Thickness (AOT), Water Vapor (WV) and Scene Classification (SCL) maps. Sen2Cor is used for systematic processing of Sentinel-2 L1C data thus generating the S2 L2A products distributed to users by the Copernicus SciHub. In parallel, a Sen2Cor Toolbox can be downloaded from the ESA website for autonomous processing of S2 L1C data by the users.
Several important evolutions had been realized from Sen2Cor version 2.8 to 2.10. Version 2.9 runs with Copernicus DEM if correctly formatted. Significant improvements were realized on scene classification module from Sen2Cor 2.8 to 2.10. Whereas atmospheric correction core modules and the AOT estimation based on dense dark vegetation (DDV)-pixels remain unchanged, a new AOT estimation fallback solution was implemented in the recent version. This new fallback solution takes AOT from Copernicus Atmospheric Monitoring Service (CAMS) data in snowy and arid landscapes in case the Sentinel-2 granule does not contain enough DDV-pixel required for AOT estimation. Sen2Cor 2.10 is designed to work with next generation product format which includes this external AOT information in the metadata.
Sentinel-2 products processed with Sen2Cor are almost compliant with the Analysis Ready Data (ARD) specifications. Knowledge of uncertainty of products is one major key to foster interoperability both through time and with other datasets. This presentation will provide a status update on the Sentinel-2 product uncertainties.
Both AOT and WV retrievals are validated by comparing with reference data provided by AERONET sun photometers at 80 locations distributed over the globe, all continents and climate zones. Spatial average of retrieval from Sentinel-2 over 9x9 km2 region around AERONET station is compared to ±15 min time average of AERONET data around satellite overpass time. Quality of SR retrieval is assessed by comparison with pseudo reference data. These are generated by running Sen2Cor with fixed aerosol optical thickness as input which is set equal to the value provided by the AERONET. The presentation will compare uncertainty of AOT, WV and SR per band for different geographical regions and climate zones.
The presentation will also analyze the sensitivity of Sen2Cor processing to parameters which can be configured by the user with running Sen2Cor Toolbox. The difference between processing with rural or maritime aerosol type and between summer or winter atmospheric profile will be discussed.

To allow for a robust exploitation of the information provided by Copernicus Sentinel-2 satellites or by the future decametric resolution TRISHNA (CNES) and LSTM (Copernicus), users must have access to high quality time series of surface reflectance, which includes an accurate and robust cloud detection, and a high quality atmospheric correction. This is the role of the Level 2A product (L2A). The French Land Data center, Theia, produces and distributes L2A products thanks to the MAJA processor, developed by the French and German space agencies, CNES and DLR, and by the Centre d'Etudes Spatiales de la BIOsphère (CESBIO), and implemented by CS-GROUP. Compared to classical processors based on multi-spectral relations to detect clouds and estimate aerosol content, Maja also involves multi-temporal criteria, assuming a usually slower variation of surface reflectance as compared to the atmospheric contribution, including clouds and aerosols.

A large scale validation effort has been done to measure the accuracy of the L2A products generated with MAJA, including the development of a method to build reference cloud masks interactively and efficiently, the use of atmospheric aerosol measurements from the Aeronet Network, or the verification of surface reflectance using automated ground stations in three sites operated by CNES, one of which, Lamasquere, is situated on an agricultural farm, while the two others are on more bright and uniform landscapes. The validation results against ground truth data are complemented with statistical tools to measure the noise on time series. All these tools indicate that the noise on time series is lower than 0.01 in reflectance units.

But even with accurate surface reflectance and cloud masks, Sentinel-2 data are still complicated to use for a lot of users. The presence of clouds causes spatial and temporal data gaps in the time series that have to be handled by automatic processing, and the swath of Sentinel-2 satellites is limited. It is therefore impossible to observe a large region in one overpass, and users have to rely on data acquired on different dates, which can cause apparition of artifacts at the limits of swaths or around the data gaps due to clouds. Overcoming this difficulty is the role of the Level 3A product (L3A), which provides a monthly synthesis of cloud free surface reflectance. In 2017, Theia has started distributing L3A products to users based on the Weighted Average Synthesis Processor (WASP), which simply makes a weighted average of surface reflectances after a directional correction to normalize data acquired from different swaths. The WASP processor has been designed to minimize the artifacts, and does it quite well as it may be seen on the joined image. The low level of visible artifacts mainly results from the quality of the L2A products generated by MAJA.

After an introduction showing all the services that have adopted the MAJA and WASP products or open source softwares, our presentation will describe the methods behind MAJA and WASP processors, detail the improvements due to our recent work on adjacency effects, explain how these codes will be adapted to the TRISHNA and LSTM missions which have a wider field of view, and provide validation results concerning the quality of cloud detection and atmospheric correction, as well as the quality of L3A syntheses.

The European Space Agency (ESA) strategy for Land products Calibration and Validation (Cal/Val) was presented in a recent scientific paper [1]. The main pillars underpinning this strategy are a reinforced focus on metrological practices to ensure traceability along the full processing chain and the need of providing uncertainty per-measurand, ideally at pixel-level. This approach strictly follows the guidelines and best practices firstly defined as part of the Quality Assurance for Earth Observation (QA4EO) framework. The application of this generic approach to ESA operational land products at different processing stages is here discussed and the readiness level for each element of the generic Cal/Val framework is assessed. The objective is to identify knowledge and data gaps and prepare the necessary actions to address these gaps. The aim of this paper is to provide a detailed review of the readiness level of Cal/Val ground-based infrastructure and relevant methods and tools for estimating uncertainty of optical land products at Top-Of-Atmosphere (TOA) and Bottom-Of-Atmosphere (BOA) level, as well as for derived land bio-geophysical products. As a result of this critical review we can state that, while for the assessment of TOA radiometry there is a long-lasting experience and a very good level of readiness in terms of infrastructure and practices, important gaps appear at the BOA stage. Among the most urgent gaps, there is an obvious need for setting up and maintaining a global network of surface reflectance ground-based measurements. This is the main priority for ESA, since it is currently the weakest point in terms of readiness level along the whole processing chain, while it is an essential element to facilitate the interoperability of current and future land imaging sensors. In terms of ground-based facilities, there is a large number of potential national and continental networks, although they significantly differ with respect to the used practices and there is little attention in providing uncertainty in Cal/Val data. The way forward for the coming years will be to contribute to harmonising best practices across existing networks to enhance their interoperability moving, in the longer term, to a sustained operational global network of Fiducial Reference Measurement (FRM) sites.

[1] Niro, F.; Goryl, P.; Dransfeld, S.; Boccia, V.; Gascon, F.; Adams, J.; Themann, B.; Scifoni, S.; Doxani, G. European Space Agency (ESA) Calibration/Validation Strategy for Optical Land-Imaging Satellites and Pathway towards Interoperability. Remote Sens. 2021, 13, 3003.

In the past 2 years, the “ESA/Copernicus Space Component Validation for Land Surface Temperature, Aerosol Optical Depth and Water Vapour Sentinel-3 Products” (referenced in the following by LAW) aimed to perform a more extensive and systematic validation against ground-based measurements of three Sentinel 3 core products: the Integrated Water Vapour included in OL_2_LFR products, Aerosol Optical Thickness included in SY_2_AOD products and Land Surface Temperature provided by SL_2_LST products.
Concerning LST validation, the LAW project was the opportunity to perform a gap analysis of the current LST networks, in particular regarding the biome distribution. Following this analysis, five appropriate locations have been defined and selected to deployed new LST stations (Svartberget – Sweden; Hyytiälä – Finland; KIT Forest – Germany; Puéchabon – France; Robson Creek - Australia). Data acquired by these stations are then compared to the SLSTR L2 LST dataset. Originally planned in Summer 2020, the deployment of these stations has been achieved in November 2021 due to the COVID-19 situation, enabling the start of the LST validation phase.
Concerning AOD and IWV validation, a selection of appropriate ground-based measurements networks has been performed and all relevant data gathered in a dedicated database. To ease the validation work, this database has been combined with specific extraction and collocation tools: each time, Sentinel 3 satellites overpassed one selected station, a macro-pixel was extracted from the Sentinel 3 product and recorded in the LAW database. Matchups, gathering combination of ground-based measurements and satellite macro-pixel collocated in time and space, were then analyzed by the LAW validation team.
Thanks to these collocated satellite/ground-based measurements, the AOD and IWV validation phase has been ended in December 2021 with consolidated performance analysis of these products, definition of their limitations and specification of relevant evolutions.
In addition to this validation phase, a dedicated web portal has been created to provide description of the project and the latest Validation results. All matchups created for AOD and IWV validation phase are also accessible (upon subscription) through https://law.acri-st.fr/home website.
A global overview of this project will be provided, followed by a description of the main achievements obtained since January 2020.

The snow cover area, defined as the spatial extent of the snow cover on the land surface, is a key variable in many hydrology, climatology and ecology studies. Earth observation satellites have been used to routinely map the snow cover area at continental scale since the late 1960s. However, most available products have a spatial resolution of 500 m and above and therefore do not meet a range of user needs for both science and operational applications. On behalf of the European Commission, the European Environment Agency has commissioned the development and real-time production of the Copernicus High Resolution Snow & Ice products (HRSI), including a snow cover component to address these needs. In particular, this service provides a canopy-adjusted fractional snow cover (FSC) at 20 m resolution along with a cloud and cloud shadow mask and quality flags. The products are derived from Sentinel-2 observations, resulting in a revisit time less than or equal to 5 days (except during low illumination periods in winter at high latitude). The products are distributed with a maximal latency of 3 h after the availability of the level 1C product in the Sentinel-2 mission ground segment, which means that they are generally available on the same day as the sensing time.

The products are generated using the MAJA-LIS pipeline. MAJA is a level 2A processor which provides slope corrected surface reflectance images including a coarse resolution cloud and cloud shadow mask, while the Let-It-Snow (LIS) algorithm provides the fractional snow cover (FSC) of every pixel identified as containing snow and a refined cloud mask. To detect snow and refine the cloud mask from MAJA, the LIS algorithm relies on a digital elevation model and four threshold parameters for the differentiation between no-snow, snow, and cloud pixels. A sigmoid-shaped empirical function for estimating the FSC from the Normalized Difference Snow Index of the level 2A pixels was calibrated using Pléiades very high resolution images. In forest regions, the FSC is adjusted to subcanopy “on-ground” FSC using the Copernicus Sentinel-2-derived tree cover density.

The snow cover detection was evaluated over Europe using in situ snow depth observations at 1764 stations from 36 countries, covering a wider range of climate and topographic conditions. We found a good agreement between both datasets with an accuracy (proportion of correct classifications) of 94 % and kappa of 0.81. More accurate (+6 % kappa) retrievals were obtained by excluding low-quality pixels at the cost of a reduced coverage (−13 % data). Performances decreased at sites with higher tree cover density due to the obstruction of the ground surface by the tree canopy. The FSC in open terrain was evaluated using SPOT 6/7 satellite images, time lapse camera photographs, terrestrial lidar scans and crowd-sourced in situ measurements yielding a root mean square error (RMSE) of about 25%. A study over an instrumented forested site in Sierra Nevada (USA) showed that the operational HRSI algorithm yielded similar performances (25-30% RMSE) as a computationally intensive spectral unmixing approach while retrieving the subcanopy ground FSC.

Monitoring the surface albedo of snow is crucial for assessing the Earth energy budget in polar and high-mountain areas. Multiple algorithms are thus proposed to compute surface albedo from satellite observations. Still, satellite-retrieved information is limited by scarce calibration/validation sites in remote areas. Ground-based observations guarantee the continuity of measurements under cloudy weather and serve as a cal/val tool, but deployment of autonomous instruments is limited in remote cold areas. In this context, we introduce the Multiband albedometer: a new instrument that performs autonomous albedo measurements. It is battery-operated with low energy consumption, hence suited for long-term monitoring in polar regions. Measured snow spectral albedo and Specific Surface Area (SSA) time series are here used to validate the Pre-operational Sentinel-3 Snow and Ice products (SICE), derived from Copernicus Sentinel-3 Ocean and Land Color Instrument (OLCI) and yet not fully validated on snow. The Multiband albedometer includes two radiometric cosine detectors facing the sky and snow surface respectively. Each channel measures the incoming radiation in 7 selected bands in the visible to short-wave infrared spectrum. Processing includes corrections for the sensor dark current, the imperfect cosine response and cross-calibration of the two channels, the effect of surface slope and sensor tilt. Two multiband albedometers were installed close to the French station of Dumont d’Urville, in Adélie Land in East Antarctica, near the snow equilibrium line and in the upper accumulation area. Albedo and SSA time series were extracted from data collected during austral summers between January 2017 and January 2021. A new cloud mask developed from Sentinel-3 SLSTR products was used to validate the day-to-day weather and validate SICE at the same location. The good agreement between Sentinel-3 derived data and in-situ measured SSA highlights the potential of the Multiband albedometer for calibration and validation of satellite-retrieved snow albedo and SSA in remote regions.

The HYPERNETS project aims to establish a new network of test sites measuring surface reflectance to validate satellite products. All sites are equipped with the newly-developed next-generation hyperspectral radiometer HYPSTAR®. The land part of HYPERNETS instuments various land cover types including homogenous surfaces like desert and snow and expanding to more complex vegetated surfaces such as agricultural fields and forests. The measurement results are routinely transmitted to a central server and automatically processed by the HYPERNETS processor. The HYPERNETS products include downwelling irradiance, directional upwelling radiance and surface reflectance, all accompanied by uncertainty and covariance information.

Validation of satellite derived products is an essential step to ensure their reliability and quality. The hyperspectral aspect of the new network is important for validation of hyperspectral satellite sensors such as PRISMA and upcoming EnMAP and CHIME (Copernicus Hyperspectral Imaging Mission for the Environment). By using a hyperspectral instrument, the results can also be convolved with the spectral response function of any multispectral instrument (e.g. Sentinel-2) within the wavelength range of the HYPSTAR® (380-1700 nm). The HYPERNETS land network has multiple sites, with automated measurement every 30 mins. These frequent measurements maximise the number of satellite overpasses available for validation. The detailed HYPERNETS uncertainty budget allows to combine these measurements and check to which extent the satellite is consistent with the traceable HYPERNETS measurements. In addition, the new instrument and its pointing system offer significant benefits of hyperspectral and multi-angular measurements which can match any satellite of interest viewing geometry or lead to the site-specific bidirectional reflectance distribution function (BRDF) derivation. Together, these characteristics make HYPERNETS an ideal network for the validation of satellite land products.

Naturally, there are some challenges to be addressed, as the new system requires more power than a multispectral instrument alternative such as the Cimel sun photometers, and the measurements take longer to capture the data for all spectral bands. Thus, a carefully considered and optimised measurement protocol is critical for successful running of the land network, which is yet to be fully validated. Only a small selection of the viewing geometries is executed at the current stage and includes six azimuth and six zeniths angles to enable frequently repeated measurement sequence. This short protocol is unlikely to be sufficient to derive the full site BRDF. Thus, in the coming months several modifications to that protocol will be applied and the results from them compared to either confirm the current selection of the angles is optimal or a new protocol will be rolled out across the network.

The land network progress will be presented including first results from the deployed sites, their performance during the winter 2020/2021 season and lessons learnt from the measurement protocol optimisation.

Geophysical validation of Integrated Water Vapour (IWV) product from Sentinel-3 Ocean and Land Colour Instrument (OLCI) was performed as a part of “ESA/Copernicus Space Component Validation for Land Surface Temperature, Aerosol Optical Depth and Water Vapour Sentinel-3 Products” (LAW) project.

IWV product was compared with reference observations from three networks: GNSS (Global Navigation Satellite System) derived precipitable water vapour from SUOMINET network, integrated tropospheric total columns from radio-soundings from IGRA (Integrated Radiosonde Archive) database and column averaged water vapour abundances from TCCON (Total Carbon Column Observing Network) Fourier transform spectrometer observations.

Dedicated matchup database was created for the validation work. For each overpass of ground-station, a macro-pixel (31 x 31 OLCI pixels) was extracted from the Sentinel-3 product and recorded in the LAW database. Matchups were screened according to preset criteria (maximum separation in time and space, cloud conditions, quality flags etc.). All matchups created for the IWV validation phase are also accessible (upon subscription) through the LAW website https://law.acri-st.fr/home.

IWV matchups were analyzed by the LAW validation team, in contact with the OLCI IWV algorithm team. In addition to general statistical validation, special attention was given to validation of cloud-flagging and error estimates provided by the IWV algorithm, as well as investigation of anomaly observed in IWV observations from center camera of OLCI/Sentinel-3B.

The Copernicus Hyperspectral Imaging Mission for the Environment (CHIME) is in preparation to carry a unique visible to shortwave infrared spectrometer. CHIME will globally provide routine hyperspectral observations to support new and enhanced services for, among others, sustainable agricultural and biodiversity management. The mission shall provide Level 1B, 1C and 2A products, as well as a set of downstream products related to the different environmental applications, such as the quantification of vegetation traits, soil and mineral properties. In this context, this work presents the latest status of developed retrieval models for the operational delivery of vegetation traits products.
As part of the CHIME mission preparation, ESA had initiated the development of an end-to-end (E2E) mission performance simulator, which is able to simulate realistically and very accurately the complete chain starting from data recording, sensor calibration and data pre-processing to sensor products up to final surface parameter maps. The E2E simulator generates realistic and synthetic CHIME data sets at raw data level (L0) in forward mode. It subsequently creates L1B, L1C and L2A products, together with the presented vegetation traits (L2BV). One of the main advantages of the E2E simulator is that any of the generated products can be validated against reference input data. In this way, the accuracy of the retrieval models can be assessed given multiple scenarios.
Within CHIME's E2E project, the L2BV module is in charge of the processing of L2A atmospherically corrected reflectance data into vegetation variables (traits), such as leaf and canopy water, chlorophyll and nitrogen content among others. Apart from the processing software (i.e., the L2BV module), the core work involved the development of retrieval models covering the priority vegetation variables. Based on experience within the context of developing retrieval models for the FLEX Earth Explorer mission, we present a workflow of building so-called hybrid models. Hybrid models combine elements of machine learning statistics with physically based methods. The same strategy is pursued to all the variables, and boils down to the following steps:
To enable training generic hybrid models, generation of a broad spectral database using simulations coming from the Soil Canopy Observation, Photochemistry and Energy fluxes (SCOPE) radiative transfer model (RTM). For the variables not present in SCOPE, alternative leaf RTMs were explored, (1) PROSPECT-DYN and PROSPECT-PRO. These RTMs were subsequently coupled with the 4SAIL model for generating top-of-canopy reflectance simulations.
Once having the training data sets prepared, Gaussian noise was applied to render the simulated spectra more similar to real (noisy) data. Then, the spectral data was transformed into 20 components using principal component analysis (PCA).
For each variable, a Gaussian process regression (GPR) algorithm was trained, validated and refined during the course of the E2E project. For validation, the models were first validated against in situ data collected during different field campaigns. In addition, models were tuned using active learning (AL) methods providing representative training data sets. Further, to enable processing complete heterogeneous images, and non-vegetated pixels were included in the training data sets in the form of non-vegetated spectra (i.e., bare soil, man-made surfaces, water, etc.).
Following some internal rounds of model development, testing and improving, we present here the v.1.8 vegetation models. This version covers the development of 13 trait retrieval models with band settings according to CHIME E2E L2A spectral configuration.
The prototype retrieval models were applied to both hyperspectral airborne (HyPlant) and spaceborne (PRISMA) imagery that were resampled to CHIME band settings and processed through the E2E chain. Among the provided vegetation products, it led to a first space-based canopy nitrogen content map over a heterogeneous landscape (Verrelst et al., 2021). In addition the retrieval models provide uncertainty information along with the estimates for each trait, which could serve, for instance, to mask out the most uncertain areas. We conclude that the obtained CHIME-like L2BV traits maps demonstrate the feasibility to routinely deliver a collection of next-generation vegetation products across the globe.


Verrelst, J., Rivera-Caicedo, J. P., Reyes-Muñoz, P., Morata, M., Amin, E., Tagliabue, G., ... & Berger, K. (2021). Mapping landscape canopy nitrogen content from space using PRISMA data. ISPRS Journal of Photogrammetry and Remote Sensing, 178, 382-395.

The Database for Imaging Multi-Spectral Instruments and Tools for Radiometric Intercomparison (DIMITRI) is a database of Level-1 products (TOA radiance/reflectance) of optical sensors; coupled with a software offering users the capability of radiometric performance assessment of optical imagers. DIMITRI was initially prototyped at ESA/ESTEC. It is currently maintained by ESA, ARGANS and MAGELLIUM (https://dimitri.argans.co.uk/).
DIMITRI offers a suite of tools for comparison of more than 15 optical sensors over more than 20 radiometrically homogenous and stable sites at TOA level, within the spectral range of 0.4 –4 μm wavelength. The database covers the period 2002 to present. DIMITRI’s interface enables radiometric intercomparisons between sensors or against simulated signals over Pseudo Invariant Calibration Sites (PICS) such as ocean, snow/ice and desert sites.
The Level-1 products ingestion into the DIMITRI-database consists on the extraction for each site of the TOA-reflectance, the sensor and solar geometries and auxiliary/meteorology information where available (wind-speed and direction, surface pressure, humidity/Water vapour and ozone concentration). Each observation is automatically assessed for cloud cover using automated algorithms making use of each sensor’s spectral coverage; manual cloud screening can also be visually performed using product quick-looks.
The monitoring of sensor’s performance using the observed TOA reflectances can be done by either:
• Sensor to sensor inter-comparison (e.g. Angular Matchup method);
• Sensor's TOA-reflectance to simulations (e.g. Rayleigh Scattering, Sun glitter, PICS and DCC methods)

This presentation provides an overview of the recent evolutions of DIMITRI toolbox such as:
- Improvement of the ingestion process
- Implementation of an installation test-unit
- Improvement of Rayleigh and Sunglint methods
- Improvement of PICS method
- Implementation of DCC method

DIMITRI is used in both Copernicus Sentinel-2 and Sentinel-3 mission performance centres, and will be used in the future Copernicus OPT-mission performance centre to ensure the accurate calibration and validation of the user products.

Keywords: Optical Sensors, Radiometric Calibration and Validation

This work describes a dataset of burned area reference perimeters derived from Sentinel-2 (S2) images for the year 2019 over Sub-Saharan Africa. The dataset has been produced in the framework of the ESA FireCCI project (https://climate.esa.int/en/projects/fire/) for validation of burned area products. This dataset covers a set of suitable S2 tiles defined as sample units and selected with a stratified random sampling approach among all tiles covering the continent. The work proposes a methodology for both sampling and for the extraction of fire perimeters from time series of S2 images over the sample units. The information strata used for sampling are the Olson world ecoregions (Olson et al., 2001) and the fire activity as derived from the FireCCI51 Burned Area product (Lizundia-Loiola et al., 2020) to identify within, each biome, regions of high/low fire activity. A total of 50 tiles, according to sampling effort limitation, were identified and time series of consecutive S2 images were exploited to derive fire perimeters with a supervised Random Forest algorithm implemented in Google Earth Engine (GEE). Time series over each sample unit were built based on data availability and to satisfy a set of requirements in terms of cloud cover of each single S2 image (< 15%), cumulated cloud cover over the time series period (< 30%), length of the time series (>100 days) and step between consecutive images (≤ 16 days). Cloud cover value of S2 acquisitions is extracted from image metadata: total scene cloud cover percentage is given by the “High probability clouds percentage”, “Medium probability clouds percentage” and “Thin cirrus” values. Random Forest algorithm was applied to each pair of consecutive S2 images to identify areas burned between two dates (short unit). Synthetic maps are derived by summing up all burned polygons and by preserving the first date of detection of burned surfaces for each polygon; clouds are also cumulated over the time series so that regions that have been covered by cloud at least once are discarded as “not observed”. The final maps represent the total burn map for the specific considered time series (Long Unit).
Output BA reference files are delivered as Open Geospatial Consortium (OGC) Geopackage format containing the following categories: burned and unburned, unobserved areas or not valid pixels (i.e. regions where the surface conditions could not be observed and assigned to neither of the two burned and unburned categories). A validation exercise to assess quality of the generated data set have been performed. Fire perimeters were compared to burned area polygons derived from very-high resolution (VHR) Planetscope images over five additional S2 tiles selected by convenience. Estimated error metrics (DC>85%) confirm that the presented dataset is suitable to be used as a reference for the validation of operational continental scale Burned Area products. Moreover, the proposed methodology for sampling S2 tiles suitable for validation could be applied to other regions and/or years and, in particular, it could be used for regions where S2 tile overlapping has a greater influence thus undermining a robust and efficient selection of tiles to process.


Lizundia-Loiola, J., Otón, G., Ramo, R. and Chuvieco, E. (2020). A spatio-temporal active-fire clustering approach for global burned area mapping at 250 m from MODIS data. Remote Sensing of Environment, 236, 111493, https://doi.org/10.1016/j.rse.2019.111493.

Olson, D.M., Dinerstein, E., Wikramanayake, E.D., Burgess, N.D., Powell, G.V.N., Underwood, E.C., D'Amico, J.A., Itoua, I., Strand, H.E., Morrison, J.C., Loucks, C.J., Allnutt, T.F., Ricketts, T.H., Kura, Y., Lamoreux, J.F., Wettengel, W.W., Hedao, P., & Kassem, K.R. (2001). Terrestrial Ecoregions of the World: A New Map of Life on Earth. BioScience, 51, 933-938.

Landsat 9 was launched on September 27, 2021, replacing Landsat 7 on-orbit, and continuing the long heritage of the Landsat program. The satellite carries the Operational Land Imager-2 (OLI-2) and the Thermal Infrared Sensor-2 (TIRS-2), which are effectively clones of the instruments onboard Landsat 8. Little has changed in the way of instrument design between Landsat 8 and Landsat 9, though the TIRS-2 thermal instrument has had baffling added to prevent the stray light issue that impacted TIRS data quality. TIRS-2 was also upgraded to a Class-B mission, which means additional redundancy is built into the instrument to ensure a longer mission life.

The Landsat 9 spacecraft will transmit all 14 bits of OLI-2 data, a change from Landsat 8, which transmits only the upper 12 bits of Earth image data. This is expected to improve the signal-to-noise ratio (SNR), particularly for dark targets.

During a three-month on-orbit commissioning phase, the radiometric performance and initial absolute calibration were assessed. The OLI-2 onboard calibrators include a shutter to measure dark signal levels, two full-aperture diffusers for calibration using the sun, and multiple lamp pairs for monitoring short term stability of the instrument’s response. The SNR is calculated using a model based on the diffuser, lamp, and dark data. One absolute calibration source unique to the commissioning phase is the simultaneous imaging with Landsat 8 OLI for three days in November 2021. This paper will cover the early assessment of radiometric stability, SNR, and absolute calibration as determined from the first months of Landsat 9 operations.

This presentation provides a status of the Sentinel-2/MSI and Sentinel-3/OLCI radiometric validation activities, performed by the Copernicus Sentinel Optical Imaging Mission Performance Cluster (OPT-MPC).
Nominal calibrations are based on the exploitation of the on-board sun diffuser images. The advantage of the on-board sun diffuser is to provide the instrument with a very uniform and well-known signal, allowing very accurate absolute and relative radiometric calibrations.
The validation activities assess all radiometric performances related to image quality requirements: absolute radiometric uncertainties, multi-temporal and inter-bands relative radiometric uncertainties, instrument response non-uniformity, signal-to-noise ratio (SNR) etc (see Foungie et al. 2016; Gascon et al. 2017). Vicarious methods (e.g. Rayleigh Scattering, Sun glitter, Desert- Pseudo Invariant Calibration Sites (PICS), Lunar Irradiance Model ESA (LIME) and Deep Convective Clouds (DCC)) are used for the absolute radiometry validation and the multi-temporal and inter-bands relative radiometry validation (Lamquin et al. 2019; Neneman et al. 2020). Absolute radiometric performance is assessed as well using ground-based instrumented sites (e.g. RadCalNet) (Alhammoud et al. 2019). In addition, cross-missions inter-comparisons are systematically performed.
The results show the good radiometric performances of the mission products, thanks to a robust in-flight calibration strategy. The radiometry in both missions is accurate (better than 5% absolute uncertainty) and stable (no detectable trend).
Nonetheless, the radiometric validation suggests that MSI-A measured reflectances are slightly brighter than MSI-B ones by about 1% while OLCI-A measured reflectances are brighter than OLCI-B ones by about 2% over the VNIR bands. In order to ensure better consistency of the TOA measurements time-series of both missions, the alignment between the two units of each mission and its impact on the downstream products are still under investigation.

Keywords: Optical sensors, Radiometric Vicarious Calibration and Validation, Cross-mission inter-calibration and inter-comparison.


References:
- B. Fougnie, C. Desjardins, B. Besson, V. Bruniquel, N. Meskini, J. Nieke, and M. Bouvet "Results from the radiometric validation of Sentinel-3 optical sensors using natural targets", Proc. SPIE 9972, Earth Observing Systems XXI, 99720O (19 September 2016); https://doi.org/10.1117/12.2237575
- F. Gascon, C. Bouzinac, O. Thépaut, M. Jung, B. Francesconi, J. Louis, V. Lonjou, B. Lafrance, S. Massera, A. Gaudel-Vacaresse, F. Languille, B. Alhammoud, F. Viallefont, B. Pflug, J. Bieniarz, S. Clerc, L. Pessiot, T. Trémas, E. Cadau, R. De Bonis, C. Isola, P. Martimort, and V. Fernandez, “Copernicus Sentinel-2A Calibration and Products Validation Status,” Remote Sensing, vol. 9, no. 6, p. 584, Jun. 2017 [Online]. Available: http://dx.doi.org/10.3390/rs9060584
- M. Neneman, S. Wagner, L. Bourg, Blanot, M. Bouvet, S. Adriaensen, J. Nieke, “Use of Moon Observations for Characterization of Sentinel-3B Ocean and Land Color Instrument”. Remote Sens. 2020, 12, 2543. https://doi.org/10.3390/rs12162543
- N. Lamquin, E. Woolliams, V. Bruniquel, F. Gascon, J. Gorroño, Y. Govaerts, V. Leroy, V. Lonjou, B. Alhammoud, J. A. Barsi, J. S. Czapla-Myers, J. McCorkel, D. Helder, B. Lafrance, S. Clerc, B. N. Holben, “An inter-comparison exercise of Sentinel-2 radiometric validations assessed by independent expert groups”, Remote Sensing of Environment, Vol. 233, 2019, https://doi.org/10.1016/j.rse.2019.111369
- B. Alhammoud, J. Jackson, S. Clerc, M. Arias, C. Bouzinac, F. Gascon, E. G. Cadau, R. Q. Iannone, and V. Boccia, "Sentinel-2 Level-1 Radiometry Assessment Using Vicarious Methods From DIMITRI Toolbox and Field Measurements From RadCalNet Database," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 12, no. 9, pp. 3470-3479, Sept. 2019, doi: 10.1109/JSTAR

In the last decade, some missions started to offer operational Level-1 (L1) uncertainty estimates for top-of-atmosphere (TOA) radiance/reflectance factor. Among them, the Sentinel-2 (S2) mission delivers uncertainty products associated to the L1C products using the Radiometric Uncertainty Tool (RUT). The delivery of uncertainty products represents an important milestone that requires consecutive efforts so that further processing levels can also offer these uncertainty estimates. This work presents the advances towards the upcoming release of a prototype version of the S2 L2A products (surface reflectance) with attached uncertainty estimates. The prototype version is under implementation in Sen2Cor atmospheric correction processor. Sen2Cor is the official processor that generates L2A surface reflectance products from the TOA reflectance L1C products. This processor is both used by the Payload Data Ground Segment (PDGS) to produce operational products and by the users as an off-line processor. The method that is proposed for the calculation of S2 L2A uncertainty is a hybrid method that combines a MonteCarlo method for the off-line calculation of look-up tables (LUT) and a GUM analytical approach for the internal Sen2Cor implementation. The LUTs contain an uncertainty estimate for each of the atmospheric functions in Sen2Cor and the correlation between them. These values are the result of an offline Monte Carlo calculation of the atmospheric functions perturbed by different uncertainty contributions. These contributions include the parameterisation input space (angular information, altitude, visibility, water vapour, aerosol type, atmosphere type and ozone), the surface lambertian assumption or molecular mixing ratio. Sen2Cor internally combines the different atmospheric functions (e.g. path radiance) to retrieve a surface reflectance estimate. The method combines the uncertainty of these atmospheric functions together with the correlation and jacobian matrices to generate a surface reflectance uncertainty estimate. The GUM analytical approach has been selected for the internal combination of the uncertainty in Sen2Cor since it provides an efficient uncertainty propagation with limited impact on the memory and processing requirements. Moreover, the presentation will show the first prototype version of the RUT over selected scenes and the validation of the first results. Some examples about the potential of these products will be discussed that include a better quality metrics for end applications such as agricultural monitoring, a better definition of prior conditions in retrieval processes or the support of Earth Observation (EO) products conformance tests.

Retrievals of geophysical products from the Sea and Land Surface Temperature Radiometer (SLSTR) Level 1 data are based on the combination of data from multiple spectral bands from the same position on Earth and rely on accurate geolocation and precise co-registration of the instrument pixels observed by different channels. Small misalignments of detector arrays in the focal plane assembly (FPA) combined with processing effects can result in positional offsets between images observed by different channels, resulting in errors in the retrieved data. The absolute geometric calibration of SLSTR VIS/SWIR products has been determined via analysis of a reference visible channel (S3) over ground control points. Inter-band co-registration is assumed and depends on the pre-launch alignment data being valid for flight conditions.

This study investigates the co-registration between images of all SLSTR channels, in particular the thermal infrared channels with respect to the channel S5. Key challenges when co-registering the thermal infrared channels with the VIS/SWIR channels are that the channels have different spatial resolutions (TIR= 1 km, VIS/SWIR= 0.5 km at nadir view), and importantly are not measuring the same signals. We have derived an algorithm that allows us to quantify the differences in the co-registration between each spectral band. The method is based on maximising the correlation between two images. It is not dependent on the presence of ground control points and can detect band to band position changes at the sub-pixel level of precision. However, because of the different spatial resolutions, it is necessary to remap the channels to a common image grid.

We present results of the inter-band co-registration analysis, showing long term trends and global distribution of the measured positional offsets. The measurements allow us to identify variations in detector misalignments possibly due to thermoelastic effects or other events in the instrument that arise during the mission.

An increasing number of Earth Observation images, and in parallel increasing number of Land Monitoring Services (LMS), defines new requirements for quality assessment in the operational productions. The Quality Assurance Framework for Earth Observation (QA4EO) developed by CEOS (Committee on Earth Observation Satellites) and GEO (Group on Earth Observations) defines the fundamental principles of the quality assurance. Such principles shall become an integral part of the LMS productions. However, there is not yet operational technology implementing the principles to be applied in the productions.
The Quality Control Metadata Management System (QCMMS), ESA GSTP project, aimed at filling the technological gap. The goal was to design and develop a prototype software system for automated quality control in the process of future LMS to allow early identification of potential quality problems through the production stages, to assure better traceability of quality-related issues, and to report QA results in the LMS process. The main component of the system is so-called Quality Control Manager (QC Manager).
The QC Manager performs set of analysis along the production pipeline from images download, trough pre-precessing steps (including multi-sensor harmonization) up to map production. Quality and quality-related indicators (~100) are collected and evaluated at each of the steps at image level, pixel level or set of images to evaluate fitness for the purpose. The indicators can be grouped into number of topics: completeness, consistency, accuracy, uncertainty and additionally lineage is stored. The QC Manager system was developed and demonstrated on three use cases: soil sealing mapping, forest change monitoring a seasonality monitoring. The source code for the QC Manager is available at https://github.com/mapradix/qcmanager as Open Source.
The system was evaluated in a set of operational scenarios with external stakeholders. The results of the use cases demonstrated the veracity of the image products, typically hidden in the LMS production, and proved the need of the quality assurance. On the other hand, the demonstration use cases revealed high computational demand of the quality indicators collection (especially the geometric and radiometric quality assessment steps), lack of common reference data sets and problem of potentially rejecting large portions of the images where there are not enough acquisitions.

Sentinel-1 and Sentinel-2 (S-1, S-2) data have jointly been analyzed to assess an incredibly impacting wildfire event that occurred in the Sardinia Island (Montiferru region), Italy, during summer 2021. The fire burned for about one week starting from July 23, 2021. In the first phase of its propagation, the wildfire affected an area mainly covered by herbaceous vegetation and wooded pastures. Then, on July 24, the fire got quickly out of hand due to the complex topography and extreme weather conditions, with temperatures close to 40°C, low relative humidity and strong winds. The most significant wildfire rates of spread were observed from the afternoon of July 24 to the evening of July 25, 2021. In this work, S-1 and S-2 data were exploited for assessing i) the area affected by fires (i.e., burned area, BA) and ii) the damage that occurred to vegetation community (i.e., burn severity). Six clear-sky Sentinel-2 multispectral instruments (MSI) images over tile 32MTK were downloaded as Level 2A with the Sen2r toolbox. Two S-2 images were selected as pre- (July 05) and post-fire (July 30) dates to extract fire perimeters with a supervised classification algorithm based on fuzzy set theory. To assess burn severity, we used NBR difference between pre- and post-fire images (dNBR=NBRpre-fire - NBRpost-fire). Moreover, seven Sentinel-1 SAR images (Path 88, VH polarization) were collected in Interferometric Wide (IW) mode (Swath 2, ascending orbits), with single-look-complex (SLC) format from July 6 to August 11, 2021. To analyze the main fire event, we focused primarily on acquisition dates of 18 July, 24 July and 30 July 2021, representing the conditions before, during, and after the fire event. SAR images were radiometrically calibrated to sigma naught maps, post-processed by applying a de-speckling noise-filtering algorithm, and co-registered; sigma naught values were finally converted from linear in decibel. Interferometric SAR (InSAR) coherence and coherence difference before and after the fire were also analyzed to investigate the relationship with fire occurrence and impact; in particular, we considered two indices: normalized coherence difference and coherence ratio indexes.
The fire perimeter extracted from s-2 image highlighted that about 13000 ha were burned during the Montiferru wildfire affecting mainly natural vegetation (55%) (Broadleaved forest, Coniferous forest and Sclerophyllus) and pastures (25%). Preliminary analysis of S-1 SAR revealed a significant drop in the VH sigma naught (~2dB) over areas most severely affected by the fire (high fire severity).
The complementary use of S-1 and S-2 observations offered further information on the dynamic of the event in time and on the impact of fire event (burn severity). The results are valuable contributions to evaluating the potential of the integrated optical/microwave remote sensing systems to assess fire damage in the Mediterranean basin.

In support of a range of operational applications, the Copernicus Global Land Service (CGLS) systematically produces biophysical products from Earth observation (EO) data. Recognising the need for quality assurance of these products, in 2017, the European Commission’s Joint Research Centre (JRC) initiated Phase 1 of the Ground Based Observations for Validation (GBOV) service, with the aim of developing and distributing robust in situ datasets for the purposes of product validation. A second four-year phase has now been commissioned to provide continuity to the service.

In Component 1 of GBOV, raw observations from a range of existing networks are collected and processed to provide datasets suitable for validating global albedo, land surface temperature, soil moisture, top-of-canopy reflectance, and vegetation products. In this contribution, we focus on the vegetation variables considered with GBOV, namely leaf area index (LAI), the fraction of absorbed photosynthetically active radiation (FAPAR), and the fraction of vegetation cover (FCOVER).

During Phase 1, over 6,500 in situ reference measurements (RMs) were derived from raw digital hemispherical photography (DHP) acquired by the National Ecological Observatory Network (NEON) between 2013 and 2020, covering 24 sites representing cropland, deciduous forest, evergreen forest, mixed forest, grassland, shrubland, and woody wetlands. Derived RMs were provided with quality indicators and estimates of uncertainty.

Using high spatial resolution imagery from the Landsat 8 Operational Land Imager (OLI) and Sentinel-2 Multispectral Instrument (MSI), quality controlled RMs were then upscaled to provide land products (LPs) suitable for validating moderate spatial resolution EO data. LPs took the form of 3 km x 3 km reference maps, including quality indicators and estimates of uncertainty, both provided on a per-pixel basis.

In this contribution, we describe the main achievements and lessons learnt in Phase 1 of the GBOV service in terms of vegetation product validation, including methodological developments and validation results for LAI, FAPAR, and FCOVER products derived from Sentinel-2 and -3. We then provide perspectives for Phase 2, including the expansion to further environmental monitoring networks and potential synergies with other initiatives such as the Fiducial Reference Measurements for Vegetation (FRM4VEG) programme.

Several airborne measurement campaigns have been carried out in recent years, with the support and funding of ESA, using the laboratory and aircraft operated by the Freie Universität Berlin. This research infrastructure (LARSIA: Laboratory for Airborne Remote Sensing Instruments and Applications) includes a laboratory with optical calibration facilities, an aircraft Cessna T207A used mainly for remote sensing campaigns and a motor glider ASK16 for in-situ measurements and training purposes. The infrastructure is part of EUFAR (European Fleet of Airborne Research) on a European level. Measurement campaigns have been conducted as part of preliminary studies of planned future satellite systems, systems and instruments already launched during the commissioning phase, for validation of existing systems (AROMAT, AROMAPEX, NITROCAM, QA4EO, SVANTE) and science/research driven activities. The focus of the campaigns carried out in recent years has mostly been the detection and spatial distribution of various trace gases (mainly NO2, SO2, NH3, but also CO2 and CH4). The respective project partners were assembled according to tasks and expertise on a European level and additionally supplemented with further required ground-based instruments and experts. A description of each campaign was done jointly within the framework of the preparation of a campaign implementation plan. The instrumentation used for remote sensing and/or in-situ measurement campaigns was assembled according to the requirements of the projects with the help of the project partners, including ground-based measurements from mobile platforms to validate the airborne measurements. The data evaluation and quicklooks were summarized with the preparation of a data acquisitions report.
The optical instruments for remote sensing purposes can be mounted at up to three bottom openings and one side window with a clear view downwards or sideways. In addition, racks can accommodate additional equipment. With the help of the cooperation partner GFZ (German Research Center for Geosciences), a number of existing hyperspectral sensors can be used to simulate the remote sensing instruments on satellite systems. With these, large airborne surveys with a spectral coverage from the UV to the thermal infrared are possible. In addition, specially developed systems or engineering models of the satellite systems can be used and tested before a deployment on larger Aircraft or Satellite. Additional instruments for describing and characterizing the state of the atmosphere can also be used to correct the acquired data. Navigation and aircraft attitude data are collected by appropriate sensors and additionally the aircraft attitude and its change during the flight can also be compensated by attitude stabilization platforms. Some of these systems were provided by the cooperation partner DLR-OS (German Aerospace Center-Optical Sensor systems). Here we introduce the infrastructure and give some examples of campaigns conducted and present some results.

Novel hyperspectral data were acquired with HyperScout 2 during the ESA FSSCat and PhiSat-1 mission and will be distributed within the Copernicus framework through the Deimos data portal. These hyperspectral data provide possibilities to advance existing capabilities both in open science practices as well as the exploitation of commercial services. Furthermore, preparatory activities in view of the upcoming ESA CHIME hyperspectral mission can benefit from these data.

The hyperspectral data injected into Copernicus target a diversity of applications as flooding, fire prediction, detection and monitoring, Urban Heat Islands, vegetation and crop status, water quality, change detection, soil moisture. During this talk an overview is provided of the acquired data made available, the in-flight calibration activities, and the achieved radiometric and geometric quality of the Level-1C hyperspectral data product.

HyperScout is a miniaturized Earth Observation instrument, developed by a European consortium led by cosine, and it has been launched for its first in-orbit demonstration on the 2nd of February 2018 as part of the GOMX-4B mission. It is a push-broom hyperspectral imager based on a 12 megapixel 2D image sensor with 50 spectral bands over a spectral range from 450 till 950 nm, a spectral resolution of 16 nm, and a spatial resolution of 70 m at an orbit height of 500 km. It is designed to be operated upon nano-, micro- and larger satellites. The extremely compact reflective telescope ensures high optical quality in the VNIR range as well as in other spectral bands. HyperScout 2 is an enhanced version of the instrument, equipped with an additional thermal channel and a Vision Processing Unit for advanced Artificial Intelligence applications.

The calibration and data processing for such a small instrument are challenging due to the lack of on-board calibration equipment. An in-flight calibration based on vicarious calibration and cross-calibration with institutional satellites as Sentinel-2 is presented. Geometric processing and geo-referencing is based on a processing chain employing machine vision techniques. The achieved radiometric and geometric quality of the data is discussed.

In addition, the authors report on the use of Artificial Intelligence and Machine Learning to cross-calibrate small satellite instruments and institutional satellites, connecting New Space to Copernicus. MATCH (MAchine learning Tool for Calibration of Hyperspectral sensors) leverages machine learning analytical power to enable the scientific use of commercial data blended with institutional data, thus, making use of the best features of the two worlds: the unprecedented high temporal resolution of data produced by large commercial constellations, and the unchallenged high quality standard of the data produced by Copernicus satellites.

Like ScanSAR images from other sensors, wide-swath images from the two Sentinel-1 satellites can exhibit noticeable intensity discontinuities at the subswath seams as well as – in some cases – periodic scalloping patterns in azimuth. These effects are particularly pronounced in images of ocean scenes, where the mean Doppler offset due to wave motions and surface currents, as well as the low backscattered power at large incidence angles and at cross polarization, can lead to suboptimal SAR processing results. The calibration and instrument noise level data provided with each image are not sufficient for correcting these artifacts.

We have demonstrated in a presentation at Living Planet 2010 [1] and in a journal paper from 2013 [2] how scalloping patterns in the different subswaths of a ScanSAR image can be extracted and divided out in an iterative approach based on spectral analysis. Similarly, one can identify and correct intensity jumps at subswath seams by analyzing mean intensity profiles across an image. While the results of such procedures may look perfect to the naked eye, adjustments of local mean intensities alone may be insufficient where the image interpretation requires analyses of variations and patterns on a pixel-to-pixel scale. A closer look at the statistics of wide-swath SAR images reveals that properties such as the effective number of looks and the line-to-line and column-to-column correlations of speckle noise patterns can be different in different subswaths and vary with the range and incidence angle within each subswath. This corresponds to a more or less blurry appearance of patterns in different parts of the image when zooming in.

In this presentation at Living Planet 2022, we will show how the statistical properties can be homogenized by further post processing of the images beyond the correction of scalloping patterns and subswath seams in the local mean intensity. Our objective in the development of the proposed technique has been to achieve the best possible statistical homogenization with a minimal degradation of spatial resolution, maximal preservation of meaningful geophysical signatures, and an acceptable processing time. Our solution is another iterative optimization scheme, which blends the original image with various isotropically and non-isotropically smoothed versions of itself and adjusts spatially varying weighting factors until a satisfactory product is obtained. We will explain the technique, show examples of homogenized images, and discuss their properties before and after the homogenization.

[1] Romeiser, R., J. Horstmann, and H. Graber, A new algorithm for descalloping ScanSAR images by post-processing, Proc. Living Planet 2010, ESA Special Publication SP-686, 5 pp., European Space Agency, Noordwijk, Netherlands, 2010.

[2] Romeiser, R., J. Horstmann, M.J. Caruso, and H.C. Graber, A descalloping post-processor for ScanSAR images of ocean scenes, IEEE Trans. Geosci. Remote Sens., 51, 3259-3272, 2013.

One of the main reasons for building a space-borne radar constellation instead of operating a single instrument only is the higher revisit rate for acquisitions over the area of interest. For synthetic aperture radar (SAR), a number of such radar constellations are currently in operation, e.g., SAR-Lupe (5 satellites), COSMO-SkyMed (4 satellites), the RADARSAT constellation mission (3 satellites), or the TerraSAR-X/TanDEM-X formation in combination with the PAZ satellite. Furthermore, within virtual constellations instruments from different platforms are compiled. For this purpose, multiple SAR sensors with different characteristics are mixed or even combined with optical instruments in order to derive multi-temporal products.
In the frame of the COPERNICUS program, the European Space Agency (ESA) has been implementing a fleet of Sentinel-1 satellites. The first one, Sentinel-1A (S-1A), was launched as a single SAR satellite in April 2014. The Sentinel-1 SAR constellation was established approximately two years later with the start of the operational phase of Sentinel-1B (S-1B). The launch of the upcoming Sentinel-1C (S-1C) is planned for end of 2022 / early 2023. The routine operation of S 1C shall start after a six months commissioning phase; then, a three-satellite Sentinel-1 constellation can be operated for a certain period of time. All Sentinel-1 satellites carry nearly identical SAR instruments which operate in C-band at a center frequency of 5.405 GHz and a maximum bandwidth of 100 MHz. The 280 transmit and receive modules (TRMs) of the phased array antenna are able to transmit a nominal peak power of about 5 kW.
On behalf of ESA, the German Aerospace Center (DLR) performed independent calibration campaigns for S-1A in 2014 and for S-1B in 2016 during their respective commissioning phases. Since then, both SAR instruments have been permanently monitored during routine operations with respect to the radiometric accuracy and stability using remotely controlled corner reflectors and C band transponders deployed within the DLR calibration field.
Radiometric calibration adjusts the measured pixel intensity in the SAR image to a physical property, the radar cross section (RCS). After calibration, this relationship is defined over the entire backscatter range, from low image power (near noise) up to high reflections (below saturation). After a proper radiometric calibration with well-characterized reference targets, the measured radar backscatter for the Sentinel-1 SAR satellite constellation is then verified over a wide backscatter range using different target types. For this purpose, the RCS derived from point targets and radar brightness from distributed targets are compared between S-1A and S 1B acquisitions over the same observation area for regions where a stable target backscatter is expected for a certain period of time.
Low differences (in the order of 0.3 dB) are found between S-1A and S-1B for high and medium backscatter derived from point targets or rainforest regions, but higher differences for low backscattering regions like ice areas and lakes. For comparing radar brightness containing low backscatter targets, an accurately derived noise level has to be taken into account. In addition to the measured lower noise equivalent beta zero (NEBZ) level, higher transmit power was detected for S-1B compared to S-1A. The slightly higher antenna gain of S-1B leads finally to a higher sensitivity for low backscattering areas of S-1B compared to S-1A and explains the found differences.

Acknowledgement: The results presented here are outcome of the ESA contract Sentinel-1 / SAR Mission Performance Cluster Service 4000135998/21/I BG. Copernicus Sentinel-1 mission is funded by the EU and ESA. The views expressed herein can in no way be taken to reflect the official opinion of the European Space Agency or the European Union.

The Copernicus POD (Precise Orbit Determination) Service is part of the Copernicus Processing Data Ground Segment (PDGS) of the Copernicus Sentinel-1, -2, -3, and -6 missions. A GMV-led consortium is operating the Copernicus POD (CPOD) Service being in charge of generating precise orbital products and auxiliary data files for their use as part of the processing chains of the respective Sentinel PDGS.
The two Copernicus satellites Sentinel-1A and Sentinel-1B are Synthetic Aperture Radar (SAR) satellites, launched in April 2014 and 2016, respectively. The POD of the satellites is done based on the dual frequency high precision GPS data from the on-board receivers. Three different orbital products are currently provided for both satellites: PREORB, RESORB and POEORB with different timeliness, accuracy and coverage. Since beginning of 2020 the new PREORB product is generated to be used instead of the on-board navigation solution; this new product has a latency of maximum 30 minutes, and an accuracy requirement better than the on-board solution (1 m in 2D). It provides a propagation of four orbital revolutions to the future, from the last ascending node. The near real-time (NRT) RESORB product has a latency of maximum three hours and an accuracy requirement of 10 cm in 2D. The non-time critical (NTC) POEORB product has a latency requirement of less than 20 days and a very high accuracy requirement of 5 cm in 3D.
The orbit accuracy validation is mainly done by cross-comparing the CPOD orbits with independent orbit solutions provided by the CPOD Quality Working Group (QWG). This is essential to monitor and improve the orbit accuracy, because for Sentinel-1 this is the only possibility to externally assess the quality of the orbits. Typical differences with respect to external solutions are well below the requirements.
In April 2021 a reprocessing of the orbits from the entire mission times of both satellites has been published on the ESA Copernicus Open Access Hub. The aim of the reprocessing is to provide a consistent set of orbits to the user community for the full mission triggered by several major updates in the operational orbital processing during the last years.
This paper presents the Copernicus POD Service in terms of operations and orbital accuracy achieved for all orbital products of Sentinel-1A and -1B. Focus is given to the new orbit prediction product (PREORB) and the reprocessed orbital product.

Remote sensing offers various approaches to extract height information from satellite images. This paper compares two quite different, yet very common methods, which are applicable to Sentinel-1 data, namely interferometry using the phase information, and radargrammetry using the amplitude information of the SAR signal (see e.g. Crosetto and Aragues, 1999).
The Sentinel-1 (S1) satellite constellation is a key element of the European Earth observation programme Copernicus. The Copernicus programme, coordinated and managed by the European Commission, provides global, timely and easily accessible information and services in application domains such as land, marine, atmosphere, emergency management, climate change and security. The S1 satellites are equipped with a C-band synthetic aperture radar (SAR). Main objectives of the S1 mission include land and marine monitoring, geohazard monitoring, support to emergency response, monitoring of ice sheets, glaciers and snow cover, and climate change applications. The operational S1 SAR instrument imaging mode over land surfaces is the Interferometric Wide swath (IW) mode which enables regular wide-spread repeat-pass interferometric SAR (InSAR) surveys. This is a powerful tool for monitoring surface motion and deformation over stable (coherent) targets for which the InSAR phase is preserved over the repeat-pass time span (6 days with two S1 satellites in orbit).
Due to these main objectives, the orbital tube of the S1 satellites constellations allows only short baselines, which limit the achievable interferometric DEM accuracy. However, very soon several scholars demonstrated the possibility to extract height information applying an interferometric DEM generation workflow (see e.g. Gallaun et al., 2015, Nikolakopoulos and Kyriou, 2015 or Mohammadi et al., 2018). In this work, the standard interferometric DEM generation workflow was extended and applied to a test site in Austria. The first extension facilitates the phase unwrapping step by reducing the number of fringes to minimum. This is accomplished by removing not only the flat terrain phase from the original interferogram but also a topographic phase part that is derived from coarse prior DEM information. After successful phase unwrapping this topographic phase is re-added to the unwrapped phase. The second extension automates the determination of the constant phase which is required to eliminate the rank deficiency of the phase unwrapping.
Again, a coarse prior DEM is used to set-up a 3D point grid over the area covered by the unwrapped interferogram. This 3D information then is transformed into the SAR slant range geometry using a map-to-image transformation which is based on the well-known range Doppler equations (Lukes and Raggam, 1986). Each pair of 3D point information and 2D SAR image information form a so-called ground control point (GCP). Of course, inaccuracies of the input DEM and of the satellite meta information (e.g. orbit information) have strong impact on the derived GCPs. However, a large number of GCPs well distributed over the SAR scene can be simulated. Since we are only interested in a constant phase shift, this results in a large overdetermination and the solution is robust to large outliers.
As alternative a radargrammtric approach as described in Gutjahr et al., 2014 was applied to Sentinel-1 data acquired from neighbouring orbits. To the best knowledge of the author, such an experiment was never published before. The accuracy of radargrammetric DEMs depends on the matching quality and the stereometric baseline. Geometric effects in SAR images like foreshortening or layover and the immanent speckle noise reduce the matching quality because the requirement of high image similarity is not met. Geometric effects are reduced as much as possible using the approaches of Gutjahr et al., 2014 and Perko et al., 2017. A simple but effective speckle reduction was achieved by multi-timporal averaging over a time series of SAR images. The stereometric baseline is fixed by the S1 orbits and varies with geographic latitude.
Fig. 1 summarizes first results obtained for province of Vorarlberg, Austria. Sentinel-1 scenes from ascending orbit 015 and 117 were used for the stereometric processing, whereas only interferometric results from ascending orbit 015 are shown. A first accuracy assessment is given Fig. 2.
References
Crosetto, M. & Pérez Aragüés, F.
Radargrammetry and SAR Interferometry for DEM Generation: Validation and Data Fusion
In SAR Workshop: CEOS Committee on Earth Observation Satellites, 1999, 450, 367-372}
Gallaun, H.; Zeug, G.; Dahmer, J.; Fontanot, F.; Uttenthaler, A.; Gutjahr, K. & Raggam, H.
EU-DEM Upgrade
European Environment Agency, European Environment Agency, 2015
Nikolakopoulos, K. & Kyriou, A.
Preliminary results of using Sentinel-1 SAR data for DEM generation
European Journal of Geography, 2015, 6, 52-68
Mohammadi, A.; Ahmad, B. B. & Shahabi, H.
Extracting Digital Elevation Model (DEM) from SENTINEL-1 Satellite Imagery: Case Study a Part of Cameron Highlands, Pahang, Malaysia
The IIER International Conference, 2018, 31-36
Lukes G. and Raggam H.
Implementation of Map-to-Image Correspondence for SAR Image Analysis.
Proc. of the 52nd Annual Convention of the ASPRS, Vol. 5, pp. 202-210, Washington D.C., March 16-21 1986.
Gutjahr, K.; Perko, R.; Raggam, H. & Schardt, M.
The Epipolarity Constraint in Stereo-Radargrammetric DEM Generation
IEEE Transactions on Geoscience and Remote Sensing, 2014, 52, 5014-5022
Perko, R.; Gutjahr, K.; Kruger, M.; Raggam, H. & Schardt, M.
DEM-based Epipolar Rectification for Optimized Radargrammetry
2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), IEEE, 2017

Abstract

During the last decades, the unprecedented development of EO-satellites have dramatically changed remote sensing in general, and SAR-based data acquiring in particular. A large number of satellites orbit around Earth and continuously scan the surface with sensors of different spatial and temporal resolutions and various wavelengths, providing reliable input data for monitoring the stability of the urban environment.

However, the greatest proportion of these raw data are originated from free medium resolution (C-Band) sensors (such as Sentinel-1 [S1]), and as such, it bears limitations in connection to the spatial resolution and wavelength of the sensor. These parameters affect both the detectability of individual buildings and the accuracy of the measurements. Still, a medium resolution sensor can provide general information about the deformations of larger areas, as well as highlight risk zones, and hotspots but the exact identification of the spatial and temporal pattern of displacements is usually highly encumbered. Despite the extensive amount of papers, which evaluate both high-resolution X-, (such as TerraSAR-X [TSX] - and medium resolution C-band sensors simultaneously [1], [2], our knowledge is still scarce regarding how to combine these measurements for effective urban deformation monitoring. Therefore, our overall goal was to define the best practice in the case of evaluating S1 and TSX InSAR measurements simultaneously.

The historical city centre of the Hungarian capital (Budapest) was selected as a test site, from where various urban distortions with different spatial extent were reported during the last decades. TSX and S1 data acquired between February 2017 and April 2019 over the site have been processed using the Persistent Scatterers (PS) and Small BAseline Sub Sets (SBAS) algorithms of Envi SARscape (5.6) software to create the basis for the stability map of the test site. For determining the displacement hotspots identifiable on S1 and TSX InSAR data, spatial and temporal analysis tools of R and Python software have been evaluated. The spatial and temporal presence and behaviour of displacement zones identified by the two sensors were statistically described and compared. As a secondary task, different environmental phase contributors (temperature-, soil moisture changes etc.) causing seasonal trends and fluctuations were identified and removed from both S1 and TSX time-series data to receive clear deformation behaviour of distortion areas. Climate data from Copernicus Climate Change Service and statistical approaches were evaluated to accomplish this task.

In essence, we intended to identify and categorize any detectable movements according to visibility, specifically determining what sort of movements are visible only to TerraSAR-X sensor, and to Sentinel-1, and to both. Then we categorized these movements according to the types (building deformation, mass movement, movements of infrastructure etc.).

Within our study area, in Budapest, significant movements are only limited to a few known and well-researched areas [3],[4]. Apart from that, there are plenty of detectable small scale displacements (railway deformations, or super-slow mass movements on the flanks of the Gellert-hill), but only traces of these movements, primarily at the most intensive parts are distinguishable from S1 images. During our research, it became evident that the S1 sensor was only able to detect more significant movements, relative to its wavelength, as opposed to TerraSAR-X, which is much more precise due to its shorter wavelength. Important to note, that the movements that were detected from the S1 data were not too exaggerated for the shorter wavelength of the X-band, so those remained clearly visible with TSX as well, but on the contrary, it is usually is not the case.

References:

1. Tosi, L., Teatini, P., Strozzi, T. (2013) Natural versus anthropogenic subsidence of Venice. Science Reports 3, 2710. https://doi.org/10.1038/srep02710
2. M. Polcari, A. Montuori, C. Bignami, M. Moro, S. Stramondo, C. Tolomei (2017) Using multi-band InSAR data for detecting local deformation phenomena induced by the 2016–2017 Central Italy seismic sequence. Remote Sensing of Environment Vol. 201, 234-242, https://doi.org/10.1016/j.rse.2017.09.009.
3. Grenerczy Gy.,Virág,G., Frey, S., Oberle, Z. (2008) Budapest műholdas mozgástérképe: a PSInSAR/ASMI technika hazai bevezetése és ellenőrzése, Geodézia és Kartográfia vol. 11., 3-9
4. Pasquali P. – Cantone A. – Riccardi P. – Defilippi M. – Ogushi F. – Gagliano S. – Tamura M. (2014) Mapping of ground deformations with interferometric stacking techniques. In: Holecz F. - Pasquali P. - Milisavljevic N. (eds.), Land applications of radar remote sensing. (ISBN: 978-953-51-4233-1) 234–259. Intech. doi: 10.5772/55833.

The principle of InSAR is based on the extraction of ground surface data to generate topography or deformation maps through phase differences between two acquisitions. While in particular early research has been extensively exploring possibilities regarding the application of the InSAR technique, they also revealed multiple error sources connected to imaging geometry, propagation path, and InSAR processing workflows. These error sources make the interpretation of the interferograms difficult, and thus might limit users’ understanding of the resulting topography without in-depth knowledge of the InSAR technique.
Based on previous research, we here separate main sources of error into six groups: errors related to the (1) imaging system, such as receiver thermal noise; (2) imaging geometry, such as orbital errors, geometric decorrelation, Doppler decorrelation, geometric distortion; (3) propagation path, including tropospheric delay and ionospheric delay; (4) InSAR processing procedures, such as interpolation errors, coregistration errors, phase unwrapping errors; (5) random noises through, e.g., a direct result of temporal decorrelation and vegetation canopy; and (6) surface deformation, in the case of investigation of phase variation targeted at topography phase.
Due to the complex interaction of various effects and challenges to distinguish the source of various phase signals, it becomes crucial to develop an understanding of the potential error source as well as their impacts, in order to be able to perform a better assessment. As it was noticed that the same InSAR workflow over the same area of interest can produce the results with completely different qualities in a previous study, it becomes important to know what factors cause undesirable outcomes, and to what extent each factor has undermined the accuracy. To this end, there are two main aims in this research. First, we construct a formal description that attempts to explain the relationship between SAR-measured elevation heights and the influencing factors; second, we carry out an error regression analysis to observe if the derived equation is realistic enough to meet empirical observations. With these two aims, we attempt to find out the gap between theoretical assumption and real signal, which might help to improve models that describe the impacts of error sources in potential subsequent studies. The ultimate goal is to better understand the scale and the influence of each error source on the final results, by which users could know what causes the problem, what corrupts the results, and where to start seeking solutions to correct them.
In this study, eighteen pairs of Sentinel-1A images were selected with the largest perpendicular baselines ranging from 70 to 140 m. The pairs were processed with SNAP software. To simplify the experiment, this research only discussed the errors sourced from tropospheric delay and surface deformation and try to mitigate random errors by choosing a proper test site and applying filtering during InSAR processing. Vertical integral refractivity data were acquired from the Weather Research and Forecasting model developed based on work published by Liu, and this data was used to stand for the tropospheric delay. PSInSAR results were provided by the Center for Space and Remote Sensing Research (CSRSR) using 121 Sentinel-1 images acquired from 2016 to 2020, and the result was used to stand for surface deformation.
In our result, a stronger correlation was shown between the error of InSAR-generated DEM and deformation. This outcome supposedly resulted from the fact that deformation was derived from the PSInSAR technique. It implies that the output of the PSInSAR technique possibly contains the same magnitude of error from the same source that exists in InSAR-generated DEMs. In addition, among 18 image pairs, only 6 pairs have shown an R-squared value better than 0.6, which indicates that the proposed mathematical description for the relationship between the error of InSAR-generated DEM and two selective parameters is not realistic enough to meet empirical observations. However, the accuracy of the collected data of two parameters also plays a key role in this study. In our study, the error range of integrated refractivity is about 5-10 mm/m, and the deformation result from the PSInSAR technique is about 10 mm. With this level of accuracy, the outcome of error regression analysis could be easily affected and misled.

With the recently published Sentinel-1 Global Backscatter Model (S1GBM) Version 1.0 [ref 1, 2], we provide a new perspective on Earth’s land surface through normalised microwave backscatter maps from Sentinel-1’s Synthetic Aperture Radar (SAR) observations. The S1GBM V1.0 describes Earth’s land surface for the period 2016–17 by the mean C-band radar cross section in VV- and VH-polarisation at a 10 m sampling, giving a high-quality impression on surface- structures and -patterns. Supporting not only the design and verification of upcoming radar sensors (including Sentinel-1C), the obtained S1GBM data also serve land cover classification and determination of vegetation and soil states, as well as water body mapping.

In the course of its generation, we processed 0.5 million Sentinel-1 scenes totalling 1.1 PB, and performed semi-automatic quality curation and backscatter harmonisation related to orbit geometry effects. S1GBM V1.0 comprises land observations from Sentinel-1’s Interferometric Wide (IW) swath mode in VV- and VH-polarisation, collecting Ground-Range-Detected data at High-resolution (GRDH) from the full 2016-2017 period. The overall mosaic quality excels (the few) existing datasets, with minimised imprinting from orbit discontinuities and successful angle normalisation in large parts of the world. Hosted at our institutional data repository [ref 3] following the FAIR principles, we provide under open license the VV and VH mosaic as tiles that are organised in a folder structure for six continents (omitting non-covered Antarctica). With this, twelve zipped dataset-collections per continent are available for download [ref 4].

In this conference contribution, we present the first extension of the S1GBM, V1.1, providing an additional set of normalised mosaics covering the northern and southern polar zones and sea ice regions. V1.1 ingests Medium-resolution data (GRDM) from Sentinel 1’s Extra Wide (EW) swath mode in HV- and HH-polarisation, at a pixel sampling of 40m. To reflect cold and warm conditions in the high latitudes, and in particular to capture the varying snow pack extents along Greenland’s coastline, data collections are set to the months January and July of the period 2016-17, respectively. Processing, normalisation- and mosaicking methods, and publication terms follow with minor adaptions the existing V1.0 dataset publication.

This extension of the S1GBM allows the characterisation of the Earth’s land surface in 2016-17 by the mean C-band radar cross section, now also including the Arctic and the Antarctic coastline and the polar sea ice regions, which are of particular interest of when mapping climate change impacts.

References
[ref 1] https://www.nature.com/articles/s41597-021-01059-7
[ref 2] https://sentinels.copernicus.eu/web/sentinel/missions/sentinel-1/sentinel-1-global-backscatter-model
[ref 3] https://researchdata.tuwien.ac.at
[ref 4] https://doi.org/10.48436/n2d1v-gqb91

Many applications of synthetic aperture radar (SAR) backscatter data require analysis-ready data (ARD) as input, which can be derived from Level-1 SAR images by applying e.g. geocoding, radiometric calibration, multi-looking, and imaging noise adjustments. Several publications [1, 5] present guidelines on how to weave these operations together using well-known SAR toolboxes like SNAP (https://step.esa.int/main/snap-8-0-released/), GAMMA (https://www.gamma-rs.ch/software), or ISCE2 (https://github.com/isce-framework/isce2) to create a preprocessing workflow with calibrated and georeferenced backscatter datacubes as output. In addition to backscatter data, most of these software packages also allow to produce certain auxiliary layers like projected local incidence angle (PLIA) data or shadow/layover masks, which are indispensable for enhancing the quality and interpretability of ARD data. Producing all these layers for every individual scene is very time consuming and may become a bottleneck for services which demand efficient, worldwide preprocessing in near-real-time.

Geocoding, which creates a link between orbit and ground geometry, claims most of the preprocessing time. The ground geometry—usually represented by a digital terrain model—is assumed static, whereas orbit data is dynamically updated with the latest vectors for every scene. Assessing the stability of orbit trajectories could thus help to significantly improve the overall preprocessing performance.

With orbital tube diameters of around 500m (1σ), Sentinel-1’s predecessor C-band SAR missions ENVISAT and ERS-1/2 had already proven their value for manifold interferometric SAR (InSAR) applications [4]. In this respect, the Sentinel-1 constellation goes one step further. Both platforms, currently Sentinel-1A and Sentinel-1B, share the same orbital plane, and revolve the earth in an orbital tube with a diameter of only 100m (RMS) [2]. Hitherto, studies have only analysed the role of Sentinel-1’s orbital tube in terms of certain InSAR parameters [3], but did not contemplate on the potential of the satellite’s orbit stability over time when generating SAR backscatter data.

In this conference contribution, we address the question how Sentinel-1’s orbit fluctuations propagate into different SAR layers and how Sentinel-1’s preprocessing can be made more efficient—without reducing the quality of the output. On the basis of four years of Sentinel-1 orbit data, we aim to quantify the impact of the orbital tube to state whether a layer can be declared as static or not. First results indicate a steady behaviour of PLIA over time, thus paving the way to establish it as a static layer per relative orbit. Such knowledge can be incorporated into the design of performant workflows, allowing to save costly resources, in particular when conducting Sentinel-1 preprocessing on a global scale.


[1] Federico Filipponi. “Sentinel-1 GRD preprocessing workflow”. In: Multidisciplinary Digital Publishing Institute Proceedings. Vol. 18. 1. 2019, p. 11.

[2] Dirk Geudtner et al. “Sentinel-1 system capabilities and applications”. In: 2014 IEEE Geoscience and Remote Sensing Symposium. IEEE. 2014, pp. 1457–1460.

[3] Pau Prats-Iraola et al. “Role of the orbital tube in interferometric spaceborne SAR missions”. In: IEEE Geoscience and Remote Sensing Letters 12.7 (2015), pp. 1486–1490.

[4] Fabio Rocca. “Diameters of the orbital tubes in long-term interferometric SAR surveys”. In: IEEE Geoscience and Remote Sensing Letters 1.3 (2004), pp. 224–227.

[5] John Truckenbrodt et al. “Towards Sentinel-1 SAR analysis-ready data: A best practices assessment on preparing backscatter data for the cube”. In: Data 4.3 (2019), p. 93.

Towards NRT Sentinel-1 ARD products: model based atmospheric corrections for improved geolocation

Ongoing activities in the context of the definition of Analysis Ready Data (ARD) products for Sentinel-1 showed that one of the critical points for the compliance to CEOS Analysis Ready Data for Land (CARD4L) specification [1] for the Normalized Radar Backscatter (NRB) product is the geo-location accuracy.
Version 5.5 of the specification [2] has a target requirement for geo-location accuracy that is 0.1 pixels rRMSE (radial Root Mean Square Error). The S1-NRB product is planned to have a pixel spacing of 10m resulting in a target requirement for geolocation accuracy of 1m rRMSE.
According to S1-MPC annual reports the Sentinel-1 geolocation accuracy is very good once one compensates the known approximations and propagation effects. The “out of the box” geo-location performance (i.e. the accuracy of the product without the compensation of known effects) suffers of a relatively small loss of precision but a significantly higher bias (3-4m) that makes it not suitable, as it is, for the generation of S1-NRB products with the target geo-location accuracy.

Recently ESA developed a new auxiliary product, the Sentinel-1 Extended Timing Annotation Dataset (S1-ETAD) [3], that consists of a set of layers for S1 SLC timing corrections that, used in combination with the corresponding S1 SLC, can ensure a geo-location accuracy better than 20cm.
S1-ETAD would be the obvious solution to the problem, unfortunately the generation of ETAD products has a latency of 21 days that is too long for some of the application scenarios connected to the use of S1-NRB products, like e.g. landside monitoring applications.
To overcome the issue, one option is to use a S1-ETAD product with better timeliness but degraded accuracy. This solution would fit needs of NRT S1-NRB products provided that the performance of the “degraded” S1-ETAD products are compatible with the geo-location accuracy requirements of the NRB product itself.

The main sources of geo-location error that are addressed by the S1-ETAD product are:
- orbits: for 5cm to 1m (1σ) accuracy depending on the orbit product type
- atmosphere: 2-4 m (slant range direction), including ionosphere and wet and dry troposphere components
- system effects and processing approximations: < 3 m (both range and azimuth directions)
- geodetic effects: < 20 cm (both range and azimuth directions)

Among the above listed error sources, the ones mostly impacting the timeliness of the S1-ETAD product are:
- POEORB orbit products, with a latency of 20 days,
- ionosphere: final auxiliary products (TEC maps) necessary to compute ionosphere corrections, with a latency of 11 days,
- troposphere: Numerical Weather Prediction products necessary to compute corrections, with a latency of 1 day

Corrections for system effects and geodetic effects have basically no constraints in terms of timeliness and can be computed as soon as the S1 SLC product is available.
In principle, exploiting RESORB orbits, it could be possible to generate a “degraded” or “fast” S1-ETAD product with a timeliness between 3 and 6 hours, by replacing the very accurate models and auxiliary product used to compute atmospheric corrections with something less accurate but faster to compute.
The overall accuracy of the new models for atmospheric corrections should be compatible with the 1m rRMSE accuracy required for the S1-NRB (10m sampling) product to be compliant with the target requirement of CARD4L specifications.
The slant range component of the ETAD correction would be the one more impacted by the change, the azimuth component, 10cm (1σ) nominal accuracy in standard S1-ETAD products, would have a minimal impact related to the slightly worse accuracy of RESORB orbit products (10cm) with respect to POEORB ones (5cm).

In order to achieve the reduced latency, possible options to be investigated are:
- using a different set of auxiliary products for ionosphere (CODE) and troposphere (NWP), e.g., derived from forecasting, to improve the timeliness.
- using very simple analytical models to represent the propagation delay in the atmosphere.

For the first option, and in particular the ionospheric corrections, the final TEC maps can be replaced by “rapid” or even “predicted” ones at the cost of accuracy. While rapid TEC maps still have a latency of about 1 day, predicted TEC maps are available few days in advance and do not affect the timeliness of the S1-ETAD product.
Similar considerations should be possible for troposphere products considering forecasting solutions.

Regarding the second option, analytic models considered in this study are very simple and have important limitations, in particular they do not consider the spatial and temporal variations of the propagation delay of the electromagnetic signal in the atmosphere linked to local conditions of the ionosphere and troposphere at different levels. Moreover, some of the models considered do not even take into account the dependency on the viewing angle or the dependency on the topography, which is a very important element for the tropospheric delay computation.

Mixed solutions including different approaches for the ionospheric and tropospheric components will be also considered.

The main expected improvement deriving from the use of simple analytic models for atmospheric propagation error correction is to reduce as much as possible the mean error bias currently observed in the “out of the box” geolocation measurement on 1 SLC products.

All the models considered require a proper calibration in order to be used in a production context.

The key point of this study is to assess the performance of the different models considered to understand if they are compatible with the “very fast” generation of S1-NRB products compliant with the target CARD4L specifications in terms of geo-location accuracy.

The methodology for calibration and validation is outlined and main results are presented and discussed.


References
[1] CEOS Analysis Ready Data (https://ceos.org/ard)
[2] CEOS (2021) Analysis Ready Data For Land: Normalized Radar Backscatter. Version 5.5 (https://ceos.org/ard/files/PFS/NRB/v5.5/CARD4L-PFS_NRB_v5.5.pdf)
[3] Sentinel-1 Extended Timing Annotation Dataset (ETAD) (https://sentinels.copernicus.eu/web/sentinel/missions/sentinel-1/data-products/etad-dataset).

Digital elevation models (DEMs) hold a key role as the main input data for a large variety of applications as they provide an accurate representation of the Earth’s surface and its corresponding topographic parameters. Scientific applications require current, reliable and precise elevation information to be implemented in research to generate correct and valuable results. The performance and accuracy of DEMs must therefore be subject to quality assessment.
Synthetic Aperture Radar (SAR) principles and methodology allow the generation of DEMs independent from day light and cloud coverage which act as a massively limiting interference factor in optical imagery.
In 2014, a C-band radar mission Sentinel-1 launched within the Copernicus Programme of the European Space Agency (ESA), providing open access to radar data of regular global coverage (Braun 2021). The capabilities of DEM derivation from Sentinel-1 imagery are limited by parameters like slope and vegetation coverage as well as by the mission design itself as the large temporal baseline of two or more Sentinel-1 acquisitions hinders the stability of the interferometric phase and degrades the coherence. One crucial step in the analysis of SAR data and therefore optimizing its potential for DEM generation is the preprocessing of suitable interferometric data pairs, a complex workflow which includes coregistration, interferogram formation, phase unwrapping and terrain correction (Braun 2021).
Within the scope of the TanDEM-X mission by the German Aerospace Center (DLR), consisting of two satellites (Terra-SAR-X and Tandem-X), a global DEM was generated from bistatic X-Band interferometric SAR products acquired between December 2010 and January 2015 (Rizzoli 2017). The TanDEM-X mission is the original source of the radar data from which the Copernicus DEM (COP-DEM) by ESA was derived. The COP-DEM is available in varying resolutions with the 30m resolution COP-DEM being openly accessible.
The Ice, Cloud and land Elevation Satellite-2 (ICESat-2) mission by National Aeronautics and Space Administration (NASA) provides openly accessible data since October 2018. ICESat-2 data is a continuous, equally distributed, high resolution reference dataset and allows to independently evaluate the performance of SAR-derived DEMs. An Advanced Topographic Laser Altimeter System (ATLAS) on board of ICESat-2 fires shots of photons in 532nm wavelength separated into six beams arranged in three pairs. Each pair consists of a strong and a weak beam with an energy ratio of 4:1 (Neuenschwander und Pitts 2019). Terrain and canopy heights retrievals are stored in the ATL08 data product of ICESat-2 and computed for fixed 12x100m land segments in which a valid height value is given if a threshold of 50 signal photons is met in order to reliably represent the land surface within the segment. ICEsat-2 provides a large variety of parameters and flags to filter the data adapted to the needs of the research purpose.
In this presentation, we will introduce several DEMs derived from Sentinel-1 data and perform a quality and accuracy evaluation. As reference data, we will use highly accurate ICESat-2 data points and two reference DEMs, TanDEM-X in 10m resolution and Copernicus DEM in 30m resolution. The main focus of this study is the correct processing and application of various parameter filter techniques of ICESat-2 data to provide accurate height reference data for the accuracy assessment of SAR-derived DEMs and the evaluation on DEM quality in consideration of different land cover types and topographic conditions regarding study sites from different continents.

References
Braun, A. (2021): Retrieval of digital elevation models from Sentinel-1 radar data - Open applications, techniques, and limitations. Open Geosciences, 13, 532–569.
Neuenschwander, A.; Pitts, K. (2019): The ATL08 land and vegetation product for the ICESat-2 Mission. Remote Sensing of Environment, 221, 247-259.
Rizzoli, P., Martone, M., Gonzalez, C., Wecklich, C., Borla Tridon, D., Bräutigam, B., Bachmann, M., Schulze, D., Fritz, T., Huber, M., Wessel, B., Krieger, G., Zink, M., and Moreira, A. (2017): Generation and performance assessment of the global TanDEM-X digital elevation model, ISPRS J. Photogram. Remote Sens., 132, 119–139.

Satellite based radar systems (Synthetic Aperture Radar - SAR) are well known for all-day and all-weather capabilities. In contrast to optical sensors, which are strongly influenced by cloud coverage, radar signals are able to penetrate clouds, haze and fog etc., thus delivering information reliably independent from weather conditions. However, in order to reach the earth’s surface, radar signals have to travel through the atmosphere twice. This causes multiple effects such as e.g. range delays and interferometric phase delays (Fig.1), which have to be considered when interpreting results based on radar data. Therefore, atmospheric correction is of crucial importance when processing radar signals..
Increasing research activities in recent years demonstrate that certain interrelationships between radar signas and the atmosphere are still not fully understood, yet. It has already been shown that using data from numerical weather prediction (NWP) models is beneficial for atmospheric correction modelling. So far mainly global model or re-analysis data sets (e.g. ERA-5 from ECMWF) have been used, i.e. rather coarse gridded data sets in which local weather phenomena are not resolved sufficiently. SAME-AT therefore improves the modelling of atmospheric correction and the use of error budgets on the atmospheric input parameters. This information is derived from forecast uncertainties from a convection permitting limited area ensemble prediction system C-LAEF (Wastl et al. 2021). A novel atmospheric correction approach is developed and specifically tested in the complex topography of the Alps in Austria. In order to achieve this main goal, the SAME-AT consortium aims to improve the quality of the reference data in Austria by further developing and exploiting corner reflector networks.

Fig.1: Comparison of ZPD GPS station Graz Lustbühel – ERA-5 for year 2017 to 2019 – summer season
Corner reflectors (CRs; Fig.2, left) are artificial passive reflectors of different shape, size and material etc. (e.g. Qin et al. 2013 and Jauvin et al., 2019) which have been used in many (In)SAR related studies. Within SAME-AT, we use a recently established corner reflector network consisting of at least 2 vertical profiles in Eastern and Central Austria.



Fig.2. Left: Design of double head CR used in the study. Right: Absolute localization error and the effect of the individual SAR corrections for a test CR set-up (according to Czikhardt, R., 2021)
In a first test phase, one double head CR was located in Seibersdorf / Austria in April 2021. We used the GECORIS software (Czikhardt, R., 2021) to analyze the effects of the individual SAR corrections on the absolute localization error (Fig. 2, right). Within SAME-AT, we are focusing on the tropospheric effects, which cause a shift in range of about 2.5 m. In our first analysis only standard atmospheric parameters were evaluated but we will continue with input from ECMWF (ERA-5) and C-LAEF.
Numerical weather models provide valuable information for SAR/InSAR correction approaches. Vice versa, observed SAR/InSAR delays and their error statistics can serve as data sources for the determination of the initial state (data assimilation) of the NWP systems. SAR/InSAR delays allow conclusions to be drawn about the tropospheric moisture content, which act as extremely valuable information for weather models. An important part of SAME-AT is therefore the investigation of the possible benefit of SAR/InSAR delays on the quality of NWP systems. SAME-AT therefore allows an improvement in both disciplines (Satellite based radar systems and development of numerical weather models) in designing the concept as a feedback loop of both novel approaches. The high spatial and temporal (e.g. hourly) resolution of current weather models offers great added value for radar applications. The developed methods for the correction of atmospheric delays is integrated into ongoing projects in order to assess the feasibility for various applications such as deformation monitoring. In recent months the Covid-19 pandemic situation and the related reduction of aircraft based observations dramatically revealed the importance of the availability of data sets entering NWP models showing a high level of diversity. Thus again both disciplines will benefit from developments of SAME-AT.
In the high mountain context, services for the assessment of mass movements, the monitoring of permafrost areas or hydrology are a crucial aspect in terms of ongoing climate change. The added value of improved monitoring services of high mountain areas are given by the increased threat on the infrastructure such as settlements, infrastructure for transport, energy supply, skiing resorts or tourism in general. SAME-AT aims to increase the attractiveness of d-InSAR based monitoring services to stakeholders and decision makers from civil protection agencies, regional geological surveying authorities, natural disaster manager, engineering offices, hydrologists, cadastre manager and infrastructure provider. SAME-AT aims to serve as door opener to new additional or complementary monitoring strategies for such stakeholders.
The free availability of the Sentinel-1 radar data still opens new perspectives for a multitude of different applications and also expands the spectrum of potential user groups. Users of SAR-based interferometry can improve the accuracy of their applications with the corrections of SAME-AT or incorporate the methods of SAME-AT in their workflows as all software developments will be made open source. This significantly increases the attractiveness of SAR-based technologies.

The construction sector has been booming recently, construction sites seem to arise everywhere around the urban or industrial agglomerations. This results from the excellent economic situation as well as the low interest rate policy of the recent years. The good order situation is expected to continue in the near future. Not even the pandemic situation in Europe could stop the construction boom. Furthermore, the amendment of the building regulations, with accelerated approval procedures, will presumably boost residential construction, according to the relevant ministry. This results in an increasing number of building permits and due to a missing automated connection between the building activities and the surveying of buildings by the Bavarian Agencies for Digitisation, High-Speed Internet and Surveying (ÄDBV), there is a high increase in the workload of the employees. In order to keep an overview and to create a relief of the employee’s new methods for the investigation of construction activities are explored, e.g., automated services that recognize new buildings from remote sensing data. Hence, time-consuming methods such as the visual comparison of the True Digital Orthophoto with the digital cadastral map could be replaced or at least much simplified when possible construction hotspots are already marked by the detection service. In this context, the following question arises: Do time series of the freely available Sentinel-1 and -2 images, processed to analysis ready data (ARD) in the extended Kennaugh framework, allow for the detection of newly built residential houses?
We present the prototype of an automated service that continuously evaluates the acquisitions of Sentinel-1 and Sentinel-2 using a temporal Wavelet analysis in the extended Kennaugh framework of the MultiSAR System for the detection of newly built structures. It reports on an optimized post-classification data fusion approach to join the weekly Sentinel-1 image and (due to cloud coverage) less than weekly Sentinel-2 image over Europe. Although the direct SAR-OPT image fusion was feasible in the extended Kennaugh framework to perform a pre-classification data fusion, we decided in favour of a weekly Sentinel-1 evaluation whose results are further stabilized by the evaluation of a Sentinel-2 acquisition whenever available. The whole process is designed to be implemented as a steady service that highlights construction activities over large areas. Thanks to the most sophisticated pre-processing to ARD, e.g. total backscattering intensity of Sentinel-1 or the newly developed LEAI for Sentinel-2, this approach needs neither training data nor computationally intensive machine learning algorithms. The validation of the individual time series of Sentinel-1 and -2 with ground truth data collected during a survey showed accuracies of up to 90% whereas the missed hits were residential houses smaller than the spatial resolution of the satellite images.
In our study, Sentinel-1 and -2 satellite imagery of the European Copernicus program from begin of October 2019 to end of November 2020 are processed. The study area is located in the northeastern part of the Bavarian town Dingolfing where numerous residential buildings are currently under construction. The need for more living space arises from the near car factory of the ’Bayerische Motoren Werke AG’ (BMW). Thanks to the weather-independent imaging capability of the Synthetic Aperture Radar system of Sentinel-1, a time series of 67 images is used and for the multi-spectral system of Sentinel-2, due to the strong influence of the atmosphere and clouds, 46 images are considered. Before creating the time series, the optical data are transformed into Kennaugh-like elements based on the method of an orthogonal transform on hyper-complex bases and are decomposed into the individual Kennaugh elements in the same way as the polarimetric images processed by the MultiSAR System processor of DLR. Furthermore, a novel multi-spectral index, the LEAIndex (LEAI), is developed from the signatures of the Kennaugh-like elements. The peak in the temporal signature caused by the construction of a new building is extremely amplified using the LEAI in comparison to using the normalized Kennaugh-like elements or using the colour channels directly, which would be the standard way. Therefore, the LEAI is derived from the Kennaugh-like elements for the whole data cube. The LEAI time series of Sentinel-2 and the ones of the total intensity of Sentinel-1 are convolved with different scales of the Haar wavelet kernel. Applying the Haar wavelet, local changes in the time series are enhanced in a robust and reliable way so that the construction of a new building can be detected as a minimum in the temporal signature. Comparing different scales of the Haar wavelet enables the determination of the minimum number of time steps needed for a reliable detection of a new building.
Afterwards, a simple threshold classification is applied to these Wavelet amplified data. The threshold is determined by a sophisticated pre-classification validation approach examining the completeness and correctness at the same time. The reference data can be seen as real ground truth data because it was provided directly by the house owners and collected in a local survey. An accuracy of up to 90 % is achieved with only the Sentinel-1 data set, combined with the Haar wavelet kernel on a scale that includes ten acquisitions. Using only Sentinel-2 data set shows an accuracy of up to 87 %, but due to the irregular intervals within the time series, it is necessary to extend the Haar wavelet to a scale covering twenty acquisitions. Especially by delimiting the data to the areas relevant to the construction of buildings, the developed method proves to be an excellent way to capture newly built residential houses from remote sensing data of the Copernicus mission. In summary, the LEAI for Sentinel-2 and the total intensity for Sentinel-1 combine all necessary information to detect the construction of a new building with an expected accuracy of more than 90%. In combination, the regular time series of Sentinel-1 providing structural information is stabilized by the time series of Sentinel-2 providing spectral information. A certain time gap concerning the detection by Sentinel-1 and Sentinel-2 was observed and could be explained by the collected ground truth data. Sentinel-1 reports on the erection of the first wall whereas Sentinel-2 highlights the foundation of the base plate. This is an important fact that justifies the decision for a post-classification data fusion in retrospect. Based on these two multi-temporal data stacks and the known time gap of the respective peak in the temporal signature an automated process can be implemented.
The original idea was to transfer the approach to the whole of Bavaria in order to support the ÄDBV. In general, even the extension to complete Germany or neighbouring countries would be feasible. From the methodological perspective, the time series analysis tested and validated on the ARD cube should be transferred to the original Level 1 data for sensitivity reasons. In this way, even the finest reasonable ground sampling of Sentinel-1 Interferometric Wide Swath Mode with about 5 by 20 metres could be investigated in order to refine the detection of smaller houses. Furthermore, the smoothing effect of the Haar wavelet is increased. From a user’s point of view, the output format should be reworked, i.e., one must evaluate whether vector or raster format is preferred or which semantic resolution is wished, e.g. only binary (yes/no) or given as probability. In the optimal case, the detected building is written to a database for further handling in the competent authority. For now, we can resume that the automated detection of construction activities on residential buildings from satellites of the Copernicus mission is feasible with a really high accuracy.

Satellite remote sensing can provide an important tool for effective and low-cost monitoring of forested areas. The availability, frequency and coverage of satellite remote sensing data have increased considerably during the last years, in particular due to the Copernicus program with Sentinel-1 (S-1). Optical images, like from Sentinel-2 (S-2), typically have the advantage of being largely designed for forest and vegetation monitoring because they are easy to interpret. However, they are severely affected by cloud cover and meteorological conditions, and they are dependent on solar illumination. The data can be supplemented with data from synthetic aperture radar (SAR) sensors, like S-1, which in general are unaffected by day-night, clouds and almost any weather condition. There is a need for mapping of forest areas with young stands under regeneration, as a basis for conducting tending, or precommercial thinning (PCT), whenever necessary. Forest clear-cut detection would be valuable for forest management, if it could be done routinely in near real-time with a spaceborne synthetic aperture radar (SAR) system, which provides data all year and all-weather.

The first objective of this presentation is to show the potential of multitemporal Sentinel-1 (S-1) data for characterization and detection of forest stands under regeneration in two study sites from southeastern Norway. We identify the most powerful radar features for discrimination of forest stands under regeneration versus other forest stands. A number of radar features derived from multitemporal S-1 data were used for the class separability and cross-correlation analysis. The analysis was performed on forest resource maps consisting of the forest development classes and age in two study sites from south-eastern Norway. Important features were used to train the classical random forest (RF) classification algorithm. The study shows that forest stands under regeneration in the height interval for PCT can be detected with a detection rate of 88% and F-1 score of 64%, while tree density and broadleaf fraction could be estimated with coefficient of determination (R2) of about 0.71 and 0.76, respectively.

The main aim was to test the SAR features to discriminate forest stands under regeneration from other forest stands. We provided a systematic investigation of S-1 parameters for the specific purpose of discrimination between forest stands under regeneration versus the other forest age classes using the class separability, cross-correlation analysis, and the features’ importance ranked by the RF approach. We found that the most powerful S-1 features are firstly the repeat-pass coherence and secondly the radar backscatter intensity of the VH channel. We used the most important features chosen in the previous step to train the classical random forest (RF) classification algorithm. Regarding the use of S-1 variables, the classification results of this study indicated that the 6-day interferometric coherence was shown to perform better than the backscatter intensity for forest age classification, although this analysis showed that it is beneficial to include both backscatter intensity and coherence.

The second objective of this presentation is to show the capability of S-1 time series for clear-cut detection. In S-1 time-series data, a forest clearing will lead to reduced backscatter intensity and increased interferometric SAR (InSAR) coherence magnitude. A time-series of 108 interferomtric wide (IW) Single look complex (SLC) S-1 images collected in 2016, 2017, and 2018 are used to study the potential for mapping clear-cut areas in eastern Ireland. We combined multitemporal InSAR coherence and backscatter intensity for the detection. This is an extension of previous studies that used either backscatter intensity or InSAR coherence magnitude, while we show the added value of both together. Coherence magnitude was the strongest predictor of the two. A novel and fully automatic procedure were developed for clear-cut detection that utilizes a long time series of S-1 SAR images acquired before, during, and after logging. We processed one prelogging time-series being the 2016 acquisitions and one time-series during and after clear-cuts, i.e., the 2017 and 2018 acquisitions. In particular, the proposed method has two innovative aspects, one is to exploit a long time-series of SAR data acquired before the clear-cut event to characterize the area in terms of SAR backscatter and coherence variations, while the other is to automatically train a machine-learning algorithm to estimate the changes in the SAR images acquired before, during, and after the clear-cut. The clear-cut detection operation consisted of five steps, i.e., the preprocessing, the SAR feature extraction, the forest/nonforest (FNF) masking, an unsupervised machine learning algorithm based on the EM approach, and assigning labels to the two classes “clear-cut” and “not clear-cut”.


References:


[1] V. Akbari and S. Solberg and S. Puliti, “Mutitemporal Sentinel-1 and Sentinel-2 images for characterization and detection of young forest stands in Norway,” IEEE Journal of Selected Topics in Geoscience and Remote Sensing, vol. 14, pp. 5049-5063, Apr. 2021.

[2] V. Akbari and S. Solberg, “Clear-cut detection and mapping using Sentinel-1 backscatter coefficient and short-term interferometric coherence time series,” IEEE Geoscience and Remote Sensing Letters, Early Access, Dec. 2020.

Nitrogen dioxide (NO2) is one of the major pollutants effecting air quality across Europe. Though a decreasing trend in NO2 pollution can be observed, most countries in Europe report concentrations above the defined EU and WHO annual limit values (European Environment Agency, 2020). Combustion sources such as emissions from industries and power plants, heavy traffic density in urban regions, biomass burning and lightning events are mainly responsible for releasing NO2 into the atmosphere (Zhang et al., 2003).Exposure to NO2 for longer duration can be harmful to human health (WHO, 2017). Furthermore, higher concentrations of NO2 in the atmosphere can also lead to the formation of tropospheric ozone and aerosols, thus negatively impacting the environment.

It is important to monitor and measure the surface concentrations of NO2 to help in strategizing decisions related to controlling NO2 pollution. In-situ measurements from ground monitoring stations provide an account of surface NO2 concentrations at hourly intervals, but they represent only a small area around their location. On the other hand, satellite measurements provide a good global spatial coverage at resolutions such as 5.5km x 3.5km. However, the satellite measured NO2 represent the concentration of the entire tropospheric column and not that of the surface. In this study, we estimated surface NO2 concentrations over Europe by mainly combining NO2 tropospheric vertical column density from Sentinel-5 Precursor (S5P) satellite observation and surface NO2 concentrations from in-situ measurements into Machine Learning models such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost). Along with S5P observations, the model was trained using following additional datasets as input parameters: night radiance from day-night band of Visible Infrared Imaging Radiometer Suite (VIIRS) instrument, solar zenith angle, digital surface model (DSM), CORINE Land Use Land Cover, Planetary Boundary Layer Height, indicators for weekdays and ERA-5 meteorological parameters such as temperature and wind velocity for the periods August 2018 - August 2021. To test the accuracy of estimated NO2 concentrations, the models were trained using 70% of the dataset and tested with the remaining 30% of the datasets, wherein RF model reported a mean absolute error (MAE) of 6.98 μg/m3 and XGBoost model reported a MAE of 6.48 μg/m3. To further validate the spatial sensitivity of the trained models, station locations were divided into two subsets such that the training and testing stations do not overlap and are spatially well distributed . The models were capable of estimating surface NO2 at previously unknown locations with MAE of 7.45 μg/m3 and 7.06 μg/m3 for RF and XGBoost, respectively. The estimated surface NO2 concentrations also showed good agreement with station measurements with correlation coefficients ranging between 0.72 and 0.79. Furthermore, on evaluating the models’ capacity to estimate seasonal patterns of NO2, relatively larger errors were observed during summer months as opposed to winter season when there is less variability in NO2 concentrations. However, these results still depict a good correlation coefficient of 0.9 with monthly averaged station observations. The good agreement of machine learning model estimated surface NO2 concentrations with in-situ data indicate the potential of using these models with satellite measurements in developing surface NO2 maps over larger regions for decision making and policy planning.

References
European Environment Agency. (2020). Air quality in Europe - 2020 report. In EEA Report (Issue No 09/2020). https://doi.org/10.2800/786656
WHO. (2017). Evolution of WHO air quality guidelines: past, present and future. http://www.euro.who.int/pubrequest
Wu, S., Huang, B., Wang, J., He, L., Wang, Z., Yan, Z., Lao, X., Zhang, F., Liu, R., & Du, Z. (2021). Spatiotemporal mapping and assessment of daily ground NO2 concentrations in China using high-resolution TROPOMI retrievals. Environmental Pollution, 273, 116456. https://doi.org/10.1016/j.envpol.2021.116456
Zhang, R., Tie, X., & Bond, D. W. (2003). Impacts of anthropogenic and natural NOx sources over the U.S. on tropospheric chemistry. Proceedings of the National Academy of Sciences, 100(4), 1505–1509. https://doi.org/10.1073/PNAS.252763799

Cropping pattern information, such as intercropping, is important for understanding agricultural trends, influencing agricultural policies, planning appropriate management practices, and supporting sustainable intensification practices. Globally, the reporting of intercropping is rarely done in agricultural statistics, and such information is still scarce, especially for small-scale farms in developing countries where there is the most need to increase food production. This study aims to understand the differences in temporal-spectral behaviour of single cropped and intercropped maize fields at different growth stages using time-series Sentinel-2 data. Information on the location, boundaries and crop types for 126 single cropped and 69 intercropped maize fields in Kenya were obtained from Plant Village and Radiant Earth Foundation during the 2019-2020 growing season. Maize was commonly intercropped with cassava, groundnut and soybean. 28 Sentinel-2 images corresponding with crop growing seasons were downloaded and processed. We examined whether there was a significant difference in the temporal mean reflectance of single cropped and intercropped maize fields at different growing stages. For this, we assessed the temporal-spectral behaviour of single crop and intercropped maize using ten Sentinel-2 spectral bands and several vegetation indices (VIs). Our results showed that the response of spectral VIs varied depending on the phenological stages of maize in both cropping patterns. Among studied VIs, NDVI showed better separability between single and intercropped maize fields during various crop growth stages. At the early growth stages of maize majority of the spectral bands showed different spectral reflectance for single and intercropped maize fields. While at the maturity stage, the differences in single and intercropped maize fields were better detected using spectral bands from red-edge, NIR, and SWIR regions. Our study shows that temporal-spectral information from Sentinel-2 data can be used to discriminate single cropped from intercropped maize fields and, hence, can improve the detection and classification of intercropping practices.

Multi- and hyper-spectral, multi-angular, or multi-sensor top-of-canopy reflectance data call for an efficient generic retrieval system which can improve the consistent retrieval of standard canopy parameters as albedo, Leaf Area Index (LAI), Fraction of Absorbed photosynthetically Active Radiation (fAPAR) and their uncertainties, and exploit the information to retrieve additional parameters (e.g. leaf pigments). Furthermore consistency between the retrieved parameters and quantification of uncertainties are required for many applications.
We present a retrieval system for canopy and sub-canopy parameters (OptiSAIL), which is based on a model comprising SAIL (canopy reflectance), PROSPECT-D (leaf properties), TARTES (snow properties), a soil model (soil reflectance anisotropy, moisture effect),
and a cloud contamination model. This allows for the retrieval under snowy conditions (snow under canopy), and the removal of minor cloud contamination.
The inversion is gradient based and uses codes created by Automatic Differentiation. Thus, also the full per pixel covariance-matrix of the retrieved parameters is computed efficiently, based on second derivatives of the code. For this demonstration, single observation data from the Sentinel-3 SY_2_SYN (synergy) product is used, which contains 16 OLCI bands and 2x5 SLSTR bands with both views (nadir and oblique). The results are compared with the MODIS 4-day LAI/fAPAR product and PhenoCam site photography.
OptiSAIL produces generally consistent and credible results, at least matching the quality of the technically quite different MODIS product. For most of the sites, the PhenoCam images support the OptiSAIL retrievals. The system is computationally efficient with a rate of 150 pixel per second (7 ms per pixel) for a single thread on a current desktop CPU using observations on 26 bands. Not all of the model parameters are well determined in all situations. Significant correlations between the parameters are found, which can change sign and magnitude over time.
OptiSAIL appears to meet the design goals and puts real-time processing with this kind of system into reach. It seamlessly extends to multi-mission and hyper-spectral data, which should lead to further improvements in terms of uncertainty reduction of the retrieved parameters. In addition, OptiSAIL can serve as framework for sensitivity studies. An extension to include SIF, thermal, and microwave sensors is possible.

Air-sea gas flux, ocean phytoplankton and Sea Surface Temperature are fundamental oceanographic parameters necessary to monitor shifts in carbonate chemistry, pH and oceanic CO2. Combining multi-sensor data from Copernicus satellites Sentinel-1 and Sentinel-3 that measure surface roughness, ocean colour and SST, represent a means of calculation of global ocean-atmosphere CO2 fluxes and ocean acidification parameters in coastal and oceanic environments, at unprecedented spatial and temporal scales.
There is a growing need to provide Fiducial Reference Measurements (FRMs) to validate both the input satellite data and the final CO2 flux and ocean acidification products. Few sites and platforms can support these measurements in the deep blue, case 1 waters. The Atlantic Meridional Transect (AMT) undertakes research from the UK and the Southern Ocean, along transects that covers several ocean “provinces” and remote regions.
The AMT4OceanSatFlux project deployed a suite of semi-autonomous underway optical and infra-red radiometers and Synthetic Aperture Radar (C-band) measurement systems on AMT to validate Sentinel-1 and Sentinel-3 products. Eddy covariance and carbonate chemistry measurements were also taken to validate the CO2 flux and ocean acidification satellite products developed by the project.
This talk will illustrate the range of FRMs collected, and the validation results for both the input parameters and final air-sea CO2 flux and ocean acidification products.

The advancement of the imaging and data handling technologies offers new and multiple perspectives to look at land. Nowadays, the bio-physical cover on the Earth’s surface could be observed through complementary methods, each revealing different characteristics and spatial-temporal aspects. For example, data from aerial imagery with very high spatial resolution could provide accurate information on the spatial aspects, while high resolution EO data from Copernicus Sentinel with their superior temporal resolution could bring evidence on the temporal behaviour and life cycle. The improved standards and tools on data interoperability allow consistent integration of these data to extract the relevant information for an application domain.

A notable example is the Checks by Monitoring (CbM), introduced in 2018 as an alternative control method for area aid declarations under the European Union’s common agricultural policy (CAP). CbM relies on a combination of different capturing and processing methods to observe the nature and behaviour of the land units (agricultural parcels). The perimeter of the agricultural parcel and the relevant topographic features within, are derived from the Geo-Spatial Aid Application (GSAA), using aerial othoimagery and the large-scale thematic data from Land Parcel Identification System (LPIS). The different bio-physical stages associated with specified land management and their timing of occurrence, are mined from the signal from Copernicus Sentinels, extracted for each agricultural parcel (statistics/clusters from set of pixels). In such a way, CbM could resolve the complexity of the observed land phenomena by breaking down the problem into smaller, logically coherent, parts that are dealt separately (tiered approach) and by eliminating noise and unnecessary processing (reductive approach).

CbM acknowledges that, while relatively coarser in terms of spatial detail, the strength of Copernicus Sentinel data is in its very high temporal resolution. The sampling rate of the Sentinel data unlocks the systematic detection of short stages and tell-tale events, associated with farming activities (ploughing, mowing). The vast amount of imagery, despite the progress of artificial intelligence and cloud computing, presents a major challenge. By working with a Sentinel signal extracted for a pre-defined spatial object, rather than the individual pixel, CbM reduce this “data deluge” to something that is easily, reliably, and accountably managed.

For such purpose, CbM introduced the concept of the Feature of Interest (FOI); as being the spatial entity for signal extractions from Sentinel data and acting as “business unit” for the major decisions made for the relevant land management practice expected. The FOI is the surface of the earth where the specified practice will be performed. In farmer’s terms, this surface would correspond to a particular field, meadow, or orchard. In Sentinel’s terms, it’s the surface area where reflectance and backscattering is measured. For many agricultural practices, a FOI’s lifecycle lasts a single growing season, for some others, it can cover multiple consecutive years; thus, its physical properties (incl. spatial boundaries) are neither given nor stable. As consequence, the digital representation of the FOI “the polygon” recorded in the system - could vary over time. Spatial congruency between the two FOI representations – one from the initial perimeter defined in the GSAA and the other obtained from automatic mapping on Sentinel imagery – is a key “boundary” condition for the correct performance of a CbM process, as it ensures that:
• the area component - officially known hectares of agricultural land cover - provided by the Integrated Administration and Control System (IACS), is correct and needs no further processing.
• data derived from Sentinel is exclusively associated with the properties of the monitored land unit (agricultural parcel), which guarantees meaningful analysis and reliable CbM decisions.

This importance has triggered a series of developments and methods to assess spatial cardinality and spatial variability within the FOI representations. The work involved three aspects:
1. Select number of Sentinel signals, design relevant methods and prototypes
2. Test the performance of the developed prototypes;
3. Compile the most successful methods in a Jupiter notebook (to be further deployed in JRC DIAS GitHub).

Four methods were initially developed and tested:
• Analysis of Sentinel 1 backscattering’s speckle noise: Assumes that in homogeneous fields the SAR backscattering speckle is following a Gamma distribution.
• Threshold on Sentinel 2 signal-to-noise ratio: It uses the ratio between the observed NDVI average and observed NDVI standard deviation.
• Unsupervised clustering through S2 image segmentation: It looks for clusters of pixels (grouped in segments) with distinct behaviour in time.
• Multi-temporal S2 supervised classification: It assesses the land cover types found within the FOI representation derived from pixel-based supervised machine learning.

The comparison of the results from the methods with the collected ground truth provided a complex picture, with no standalone method performing sufficiently well on combined commission and omission errors. Methods seem to be complementary rather than competing, as each addresses and reveals different aspects. The research, however, provided important insights into the nature of the FOI conditions as well as valuable input for the setting the parameters of the methods. It also led to the elaboration of a new statistical method for FOI assessment, based on the based on “interquartile analysis” of the pixel value distribution of the NIR and RED Sentinel – 2 bands within the FOI representation. Another result is a generic “best practice” notebook for the detection of clusters within FOI representation, provided by any thematic raster (available through https://github.com/ec-jrc/cbm).

The full JRC progress report on Assessing the Validity of the Feature of Interest, in the context of the CAP Checks by Monitoring is available on: https://marswiki.jrc.ec.europa.eu/wikicap/images/7/75/JRC123711_foi_assessment_final22.pdf

In the E-SHAPE "EuroGEO Showcase: Applications Powered by Europe" project, the Pilot 2 of “GEOSS for Disasters in Urban Environment” show case has as main goal the use of Copernicus data and services through the APIS Copernicus Open Access Hub. These data can be assimilated into the WRF meteorological model coupled with reflectivity radar data and lightning observations to improve the forecast of hight impact weather events in urban environment. Starting from the ESA STEAM project results, this pilot presents the added value of homogeneous Earth Observation assimilation into the high spatio-temporal resolution meteorological simulations for the extreme event happened over North-western Italy on 21-22 October 2019 characterized by very intense and persisting rainfall with consequent wide flooded areas.
During the the flood one person died and 130 people were evacuated after flooding in Alessandria Province, Piedmont. Piedmont Regional Agency for Environmental Protection (ARPA) reported 253mm of rain felt in 12 hours in Casaleggio Boiro on 22 October. The Bormida river in Alessandria jumped from 1.2mon 20 October to around 7.5m on 22 October, above the danger mark of 7m. The Orba river in Casal Cermelli has also exceeded the danger mark 4.5m, reaching 5.78m on 22 October.
The synoptic analysis shows that between 19-22 October, north-western Italy was affected by intense humid southeastern currents associated with an Atlantic trough centred over Spain, which brought adverse weather with very intense rainfall. The precipitation between Liguria and Piemonte was mainly convective. In particular, in the early afternoon of Monday a self-regenerating thunderstorm structure was formed on the Ligurian Sea where it remained stationary for about 12 hours, determining a series of rain showers of exceptional intensity. The so-called V-shaped Mesoscale Convective System (MCS) was stuck because of the wind convergence in the low atmosphere and the region’s steep orography played a fundamental role. This kind of events are very hard to be predicted, and during the forecasting phase in this case almost no LAM (Local Area Model) has been able to predict the correct position and intensity of such a dangerous and stationary structure.
Thus, this use case is chosen to test the complete hydro-meteorological chain of the E-SHAPE project starting from the Sentinel-1 OCN (wind over ocean) product processing and assimilation into the WRF meteorological model, coupled to reflectivity and lightning assimilation, up to the hydrological peak discharge prediction with the Continuum model. The main aim is to analyse the value added of such innovative observations assimilation during a forecasting-nowcasting phase starting from the meteorological output up to the peak discharge forecast that become very important in such regions characterized by complex orography.

Space-geodetic techniques such as Global Navigation Satellite Systems (GNSS) and Syntenic Aperture Radar interferometry (InSAR) are powerful tools to measure and monitor ground surface motion. InSAR has widely been used for the detection and quantification of slow mass movements over the past three decades mainly at the local and regional scales. The high performance and millimeter-level measurement accuracy of radar satellite to provide a dense deformation map at different spatial and temporal resolutions are the key factors to think of using SAR data and InSAR technique as an efficient tool for geohazards motoring system at the nationwide scale.
Sweden has recently joined to the countries having InSAR Ground Motion Service (GMS) at a nationwide scale. The InSAR service of Sweden, which will soon be freely available for users, provides the displacement time-series of measurement points for the entire country. The Swedish GMS project was started last year and is an ongoing collaboration between the Geological Survey of Norway (NGU) and several Swedish organizations (led by the Swedish National Space Agency (SNSA)). The InSAR-based GMS of Sweden has been generated by NGU using Sentinel-1 data (2015–2020) and the Persistent Scatterer Interferometry (PSI) technique. The web-based GMS of Sweden consists of ~1,5 billion time-series measurement points obtained from both descending and ascending satellite orbital modes.
Currently, the Swedish GMS is under evaluation and validation phase and the given plan has been designed to assess the quality or validate the GMS products. We plan to conduct the data validation through two main phases: 1) a cross-comparison between InSAR measurement points and ancillary data such as GNSS, Corner Reflectors (CR), Electronic Corner Reflectors (ECR) and leveling data, and 2) assessment of tropospheric and ionospheric effects on InSAR measurement points. Specifically, we will evaluate different approaches and data for the InSAR tropospheric corrections, such as Very-Long-Baseline Interferometry (VLBI), Water Vapour Radiometry (WVR), and GNSS data at the Onsala Space Observatory (OSO).

In the first phase of validation, leveling data collected in Gothenburg and Stockholm cities, mainly over the residential areas and public transport infrastructures compared to the corresponding InSAR measurements points (vertically converted) for a five-year period. The initial results present a high correlation between two sets of the vertical displacements. The same procedure will be performed for the Kiruna city where the mining activities resulted in a drastic urban land subsidence. Since the CRs and ECRs have recently been installed in different parts of Sweden, we do not have them as PS points in the current version of the GMS. Therefore, those CR-based measurement points will be used in future accuracy assessments.

As the InSAR-based GMS can be used to monitor and identify the potential risk of geo-related hazards in Sweden, the society will directly benefit from the outcomes of this project. This open access product will help the stakeholders with decision support for prioritization of risk-reducing measures, and identification of the need for further investigations for areas in danger. The service could also assist municipalities and county administrative boards to have an update information regarding urban areas which are more prone to land subsidence and disruption urban infrastructure.

The Copernicus program is an Earth monitoring initiative led by the European Union (EU) and carried out in partnership with the EU Member States and the European Space Agency (ESA) to access accurate and timely information services to better manage the environment, understand and mitigate the effects of climate change and ensure civil security.

Anthropogenic CO2 Monitoring (CO2M) Mission is part of the Copernicus Space Component Expansion Programme and shall observe the human-made atmospheric CO2 and CH4 content by providing the XCO2 and XCH4 total column concentration. CO2M will perform greenhouse gas monitoring in an operational context with high spatial and temporal resolution and short time updates. Though also able to quantify natural sources and sinks for CO2 and CH4, the main focus of CO2M will be the quantification and monitoring of anthropogenic CO2 emissions with high precision.

The CO2M mission, lead by a consortium of which OHB System is the industry prime, will consist of at least two spacecraft in an orbit with a semi-major axis of 7113 km and a repeat cycle of 11 days. The mission has kicked-off its implementation phase in July 2020 and is expected to complete phase B2 early 2022. The launch is planned before end of 2025 to meet the commitments made by the Paris Agreement to limit the global average temperature increase compared to pre-industrial levels to below 2.0°C and targeted to be below 1.5°C.
Each satellite carries an imaging spectrometer (CO2I) dedicated to measuring atmospheric CO2 content as the main instrument. Designed for high spatial and high spectral resolution, this instrument will image the spectral radiance in the NIR, SWIR-1 and SWIR-2 bands as input to the retrieval algorithms for determining column-averaged mixing ratio of CO2 and CH4 in the atmosphere. An additional imaging spectrometer dedicated to measuring the atmospheric NO2 column is integrated into the CO2I providing data at a spatial resolution of 4km2 and fully co-registered measurements. By combining CO2 and NO2 measurements, the greenhouse gases can be clearly traced to human-made emissions of high temperature combustion processes.
In addition to CO2I, the CO2M spacecraft is equipped with a multi-angle polarimeter (MAP) and a cloud imager (CLIM). The CO2 instrument relies on the quality of the reflectance measurements in the Near-Infrared and Short-Wave Infrared spectral regions. Aerosols in the atmospheric column viewed by the CO2I instrument perturb the accuracy of the CO2 retrieval, by scattering light into/out of the CO2I light path and hence accurate characterisation of aerosols is required. The strength of this scattering as a function of both wavelength, angle, and degree of polarization allows for sufficient correction in the retrieval. Low water clouds and thin cirrus also affect the CO2 retrieval process. Measurements from a dedicated cloud imager (CLIM) will allow for detecting and filtering the data for cloud-contaminated cases. The combined measurements performed by these instruments will be used to achieve a XCO2 precision of less than 0.7ppm and a systematic error below 0.5 ppm per spatial sample of 4km2 sampling size.

The CO2M satellite utilises OHB’s standard Earth Observation Platform, Eos. The Eos platform is founded on common and well-established state-of-the-art technologies, thereby providing the opportunity of a fast track and low risk platform adaptation. The platform will apply the latest PUS-C standard and implements the File Based Operation concept, allowing to send also unbounded files during ground contact, ensuring fast availability of the measured data to the end-user.

In this presentation OHB System will provide an overview of the mission, technology and outlook of the CO2M Satellite to reach the commitments made at the Paris Agreement.

Since the first launch in 2014, the Sentinel-1 (S1) satellites have acquired several tens of thousands of valuable Extra Wide (EW) swath data over the polar regions, and still counting. The use of interferometric SAR techniques (InSAR) on EW data is theoretically possible given that the technical design of EW acquisitions is similar to the Interferometric Wide (IW) swath mode. However, it requires the ability to focus raw level-0 data to Single Look Complex (SLC) data as most data in the EW archive is only available as level-0 or as Ground Range Detected (GRD). Over the last decade, and with support of ESA, scientists at NORCE Norwegian Research Centre have developed this unique capability, providing new opportunities for novel research on various applications within the cryosphere.

In this presentation we will show first results of an ESA pilot study (EW-Explore) to investigate the potential of the level-0 EW data archive for cryospheric applications. Here, the focus is primarily on the use of EW data in glaciological applications in East Antarctica. As most EW data is acquired over coastal areas, we naturally look at the boundary where land ice overflows into the ocean as ‘ice shelves’. These floating bodies of ice serve as gatekeepers for the massive amounts of ice stored inland while at the same time, they act as indicators for climatic change as they tend to quickly respond to alterations in oceanic- and atmospheric conditions. Understanding the processes that govern the stability of these ice shelves is a very active field of research, in which SAR data play a key role.

In our project we use InSAR methods (like double difference interferometry (DDInSAR), classical 2-pair InSAR and (in-)coherent offset tracking) on a subset of the EW archive to capture essential climate variables (ECVs) of specific ice shelves in Queen Maud Land, East Antarctica. These ECV parameters include the estimation of ice velocity, grounding line- and calving front locations. Additionally, we look to monitor the development of ice shelf cracks, estimate calving fluxes, and map surface features (channels, pinning points etc.). To understand selected geophysical processes such as the migration of grounding lines, we plan to utilize the almost continuous 6-days repeat pass for time series analyses. Lastly, we will show how EW InSAR compares to classical IW InSAR and how we can use IW and EW data in a synergistic approach.

Using InSAR techniques on the extensive S1 EW archive is likely to be a game changer when it comes to monitoring glaciological processes in coastal Antarctica, especially when processing is scaled up to use the entire archive in both space and time. Likewise, the archive covering the Arctic has a fair potential for glaciology too, but also within other cryospheric applications like ground surface deformation in permafrost terrain (thaw subsidence and frost heave time series).

Remote sensing observations of the composition of Earth’s atmosphere are performed with instruments operating on space-borne and airborne platforms and from ground-based stations. In this context, vertical profiles of atmospheric variables are often obtained with an inversion procedure (retrieval) from the observed radiances.
When one or more instruments observe the same portion of the atmosphere, the information obtained from the different measurements can be combined in order to get a unique vertical profile of improved quality with respect to that of the profiles retrieved from the single observations. The most comprehensive way to combine different measurements of the same quantity is considered to be synergistic retrieval, which jointly inverts all the observations and produces a single output profile. However, recently a new method referred to as Complete Data Fusion (CDF) was proposed that, in linear approximation conditions, provides products equivalent to that of the synergistic retrieval with simpler implementation requirements.
Using a series of examples based on real data, we will present the different contexts in which the CDF can be used and the key benefits that can be achieved. This is particularly interesting considering the forthcoming operation of atmospheric Sentinels. In fact, the examples deal with precursor instruments to those operating on the new platforms or others that could provide complementary information to them. Another essential aspect to be underlined is that the CDF can modify the characteristics of a product (spatial and vertical resolution, a priori information), making it more compatible with other tasks such as assimilation, source points detection and time-series calculation.
In particular, here, we present some results of the CDF application to measurements of vertical profiles of Ozone, Temperature, Water Vapour and eventually other trace gases, performed by different instruments (GOME2 and IASI at least, but eventually also TROPOMI, MIPAS and others). The method’s inputs are the profiles retrieved directly from the single measurements, characterized by their a priori information, covariance matrices and averaging kernel matrices. The output consists of a single profile also characterized by an a priori information, a covariance matrix and an averaging kernel matrix, which collects the information content of the input profiles.
The fused product is compared with the input ones in terms of errors and number of degrees of freedom (DOFs). We will see that, in general, the fused product has lower errors and higher DOFs if compared with the L2 ones, so we will analyse the mechanism that provokes this quality improvement, also considering the shape of the individual averaging kernels rows. We will discuss the strategies to allow the fusion of non-coincident measurements and the relative implications in terms of information content. We will also focus on the actual open problems and the desirable future developments and applications.

As global demand for liquefied natural gas (LNG) increases, so does the need for continuous monitoring of these important production locations. Gas flaring is a standard procedure for safety and process-related flow balances during plant utilization changes but is also automatically triggered when key equipment fails. The most relevant compounds emitted from flaring activities are mainly carbon dioxide (CO2), and short-lived climate pollutants, such as black carbon, known for their warming potential. In addition, gas-driven compressors add emissions at continuous production-related levels.

As a result of the lack of reported detailed information from such industrial facilities worldwide, satellite remote sensing provides one of the few avenues to close substantial information gaps and provide timely, high-quality LNG supply chain data. Satellite data has been utilized over the past decade to estimate the amount of gas flared, and subsequently emissions. The rate of CO2 emissions can be approximated from fire radiative power (FRP) calculations, and the combined use of multiple satellite instruments has significant potential for improving the estimation of gas flaring volumes and emissions. Here we concentrate on the synergistic use of Sentinel-2 and Sentinel-3 data, in addition to comparisons with the more frequently used VIIRS and MODIS sensors. We focus on data from several LNG sites, i.e., Corpus Christi, Sabine Pass, and Cove Point (USA), Curtis Island (3 individual plants; Australia), Ras Laffan (Qatar), and Oman LNG (Oman), to cover a range of liquefaction capacity, flaring frequency and type (ground flares and flare stacks), site complexity (buildings, gas turbine-trains), as well as meteorological and surface conditions.

Sentinel-2 provides higher temporal and spatial resolution compared to Landsat-8 (20 m versus 30 m, A+B ca. 5 days versus 16 days). Similar to Landsat-8, top-of-the-atmosphere (TOA) radiances in the shortwave infrared region (SWIR) (1.6 μm and 2.2 μm) bands of the Sentinel-2 Multi Spectral Instrument (MSI) can become saturated for hot flames. Therefore, the calculation of FRP is constrained to very small flares. The high spatial resolution of Sentinel-2, however, allows for the detection of very small gas flares as well as non-gas flare hotspots at LNG sites. In addition, EUMETSAT/ESA provide FRP products from the Sea and Land Surface Temperature Radiometer (SLSTR) instrument onboard of Sentinel-3, which features SWIR, mid-infrared (MIR) and thermal infra-red (TIR) channels with spatial resolution between 500 m and 1 km. which both help avoid detector saturation. One of the disadvantages of the Sentinel-3 FRP product is that it may show detections that cannot uniquely be attributed to the flare sites.

Here, we present the methodology and first results of the synergistic exploitation of Sentinel-2 and Sentinel-3 for improved flare and general hotspot monitoring at LNG plants. Our study also shows the advantages of the combined use of the Sentinels to determine omission and commission errors, thus generating a more reliable flare-FRP and subsequently CO2 emission estimates from gas flaring at LNG plants worldwide. In addition, we have successfully implemented plant monitoring operationally.

Despite still facing various challenges of accuracy, sampling, and continuity, remote sensing is considered to be most promising for supporting sustainable development, especially in regions where dense in-situ measurement networks are missing. In this case-study we apply remote sensing to support sustainable water management in the Muringato sub-catchment in Nyeri County, Kenya, where accurate and up-to-date information regarding the land’s surface is needed for both modelling and planning purposes. For water resources, possibilities in remotely sensed land use and land cover do not only cover the detection of waterbodies, but also urban or cultivated areas as indicators of local demands. Most of the region’s specific problems have also been identified to be directly related to poor land use practises or unfavourable change of land cover such as water abstraction, soil erosion, and degradation. As the situation is highly dynamic, reinforced by climate change, and regularly in-situ data acquisition is not practical nor economical, automated and repeatable methods are needed. The study’s results aim to support the local stakeholders, decision makers, and resource users in monitoring changes of the ~225 km² large Muringato sub-catchment and to automate tasks that previously had to be done in-situ. The combination of ESA’s Sentinel-1 and Sentinel-2 satellites offers the possibility of spatially explicit data in high temporal and spatial resolution. It enables monitoring even in regions with very high and frequent cloud cover like Kenya with two rainy seasons which add up to almost half a year. The study utilizes a synergetic approach using data of both sensor types for annual monitoring while exploring additional data products like the Evapotranspiration Stress Index derived from the Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) and the Global Canopy Height dataset derived from combined Landsat optical data and Global Ecosystem Dynamics Investigation (GEDI) Lidar data. ECOSTRESS and GEDI are both situated on the International Space Station (ISS). Sentinel-2 data is pre-processed in the public available Google Earth Engine to cloud-free median images and used along matching Sentinel-1 data as basis for an object-based-approach. Resulting objects are also evaluated regarding annual Sentinel-1 characteristics, leading to a classification that also considers seasonal changes, for instance in cultivated areas. We could show that an object-based synergetic approach combining Sentinel-1 and Sentinel-2 data is suitable for monitoring land use and land cover changes in areas with high and frequent cloud cover such as Muringato. Additionally, we found that including ISS-borne Earth observation data is beneficial for the monitoring of the region’s water resources. Especially ECOSTRESS also hints at the added value for the operational use of data from upcoming missions like The Copernicus Land Surface Temperature Monitoring (LSTM) within a constant monitoring.

Intertidal flats provide essential services including protection against storm surges and coastal flooding. They are characterized by a continuous redistribution of sediment and changes in topography. Nowadays, there is an increasing need for measuring the topography of intertidal environments in a regular manner. The continuous monitoring of their topography is fundamental for hydrodynamic and morphodynamic modeling of coastal systems [1]. Measuring the topography of such environments using conventional ground and airborne-based methods is logistically very challenging. Spaceborne-based techniques are a viable alternative that can provide recurring large-scale monitoring [2]. Spaceborne monitoring of intertidal environments is essentially performed using the waterline method. The waterline method consists in extracting waterlines (water/flats interface) from remote sensing images, assigning heights to the extracted waterlines using sea level information, assembling of waterlines, and interpolation [3], [4]. Used traditionally, this method is dependent on a significant amount of manual effort and manual processing. This limits its use for long term and large-scale monitoring [5]. The main difficulties are the following: (i) Insufficient number of satellite images acquired in a relatively short acquisition period, (ii) Manual elimination of noisy SAR images, (iii) The subjective satellite-image thresholding required in the edge detection process, and (iv) The manual post-processing of the extracted waterlines prior to interpolation. In this study we propose a new automatic approach for using the waterline method to derive intertidal Digital Elevation Models (DEMs) from Sentinel-1 and Sentinel-2 images. The changes include a faster, more efficient, and quasi-automatic detection and post-processing of waterlines. Using Sentinel-1 and Sentinel-2 images, DEMs were generated by this new waterline method approach for the Arcachon Bay and the Bay of Veys located both on the French Coast. The comparison of the generated DEMs with lidar observations showed an error of about 19–29 cm. Finally, the derived DEMs were used to track topographic changes and the intertidal water storage variation for the studied intertidal areas from 2016 to 2021. Furthermore, we investigated the performance of the future wide-swath altimetry mission SWOT to use the waterline method in order to generate intertidal DEMs. In contrast to conventional altimetry missions intended to ocean surfaces only, SWOT mission will provide high precision measurements of water surface elevation over oceans, coastal, and inland water surfaces. Over coastal and land areas, SWOT will provide sea surface heights measured in 2 dimensions with a water mask [6] . This type of data is ideal for the waterline method as the water mask can be used to extract the waterline directly, and the height of this waterline is measured by SWOT which will make the waterline method completely independent of in situ measurements. To investigate the performance of SWOT tests were made by generating intertidal DEMs from SWOT-type observables simulated using the SWOT Large scale simulator [7]. By comparing SWOT-derived DEMs to the validation DEMs (Sentinel-derived DEMs), SWOT showed a great potential for monitoring intertidal topographies. Mean Absolute Errors (MAEs) and Root Mean Square Errors (RMSEs) reached respectively 5.2 cm and 8.4 cm for the Arcachon Bay and 10.2 cm and 17.3 cm for the Bay of Veys [8]. The ability of SWOT to track topographic changes and measure the intertidal storage volume were also assessed and showed very promising results.

[1] J. Benveniste et al., “Requirements for a Coastal Hazards Observing System,” Front. Mar. Sci., vol. 6, p. 348, Jul. 2019, doi: 10.3389/fmars.2019.00348.

[2] E. Salameh et al., “Monitoring Beach Topography and Nearshore Bathymetry Using Spaceborne Remote Sensing: A Review,” Remote Sens., vol. 11, no. 19, p. 2212, Sep. 2019, doi: 10.3390/rs11192212.

[3] D. C. Mason, I. J. Davenport, G. J. Robinson, R. A. Flather, and B. S. Mccartney, “Construction of an inter-tidal digital elevation model by the ‘water-line’ method,” Geophys. Res. Lett., vol. 22, no. 23, pp. 3187–3190, 1995.

[4] G. Heygster, J. Dannenberg, and J. Notholt, “Topographic Mapping of the German Tidal Flats Analyzing SAR Images With the Waterline Method,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 3, pp. 1019–1030, 2010.

[5] E. Salameh, F. Frappart, I. Turki, and B. Laignel, “Intertidal topography mapping using the waterline method from Sentinel-1 & -2 images: The examples of Arcachon and Veys Bays in France,” ISPRS J. Photogramm. Remote Sens., vol. 163, pp. 98–120, May 2020, doi: 10.1016/j.isprsjprs.2020.03.003.

[6] S. Biancamaria, D. P. Lettenmaier, and T. M. Pavelsky, “The SWOT Mission and Its Capabilities for Land Hydrology,” Springer, Cham, 2016, pp. 117–147.

[7] CNES, “SWOT Hydrology Toolbox.” Toulouse, 2020, [Online]. Available: https://github.com/CNES/swot-hydrology-toolbox.

[8] E. Salameh, F. Frappart, D. Desroches, I. Turki, D. Carbonne, B. Laignel, 2021. Monitoring intertidal topography using the future SWOT (Surface Water and Ocean Topography) mission. Remote Sens. Appl. Soc. Environ. 23. https://doi.org/10.1016/j.rsase.2021.100578

Operational monitoring of land management practices is crucial for the effective design, implementation and evaluation of thematic policies. This is particularly important in the case of the European Union common agricultural policy (CAP) and the European Green Deal. Given the complexity of agricultural management activities and environmental protection actions, the complementary evidence provided by optical and radar sensors has the potential to contribute to a more reliable detection of land management activities in operational contexts.

Several algorithms to identify agricultural and environmental protection practices based on Copernicus Sentinel-1 and Sentinel-2 satellites have already been published in the literature. However, the operational application of these methods for near real-time monitoring of land management practices over large areas requires innovative approaches that can efficiently analyse the large volume and streams of Sentinel data. These developments must go together with recent cloud-based advancements in data storage and management.

In this context, the GTCAP team at the Food Security Unit, JRC - EC, has developed a processing framework to explore and analyse Sentinel-1 and Sentinel-2 data. This framework was built using a set of integrated Python modules and Object Oriented Programming (OOP) to facilitate its potential extension with new modules and algorithms designed to detect specific land practices. This tool enables researchers to explicitly explore and model mechanistic relationships between land management practices and the variations in satellite signals over time.

The system is designed to reduce the data volume from Sentinel satellites and transform it into manageable information. Following this logic, the analysis is based on information extracted from satellite bands for each homogeneous land management unit, moving from pixel to object analysis. The information used is band statistics (mean, standard deviation, min, max, median, quantiles) and the histogram derived from the Scene Classification Layer (SCL) for Sentinel-2, and statistics of backscattering and 6-day coherence for Sentinel-1. The signal processing modules take these statistics as input for the detection of activities occurring at the level of the spatial unit. The statistics are generated by a backend in a Copernicus Data and Information Access Service (DIAS) environment with access to the Sentinel archive and stored in a database in tabular format. The backend is described in the abstract “Integration of Copernicus Sentinel-1 time series in ML analytics for agricultural practice monitoring” presented in session “Advanced Solutions for SAR processing and analytics” (C01-08).

The main focus of the data analysis tool presented here builds on these statistics and can potentially be applied to a variety of monitoring domains, for example agricultural management and carbon monitoring. As an example to illustrate the logic, we present the results from a mowing detection pilot study in the context of the CAP, using agricultural parcels declared by farmers as reference spatial objects.

The processing architecture is built on four main modules: 1) import of band statistics, 2) data preprocessing, 3) identification of relevant changes in the signal time series (marker), 4) aggregation of markers from different sensors and association with land management practice, through the relevant bio-physical stages. Marker is a key concept in the analysis, and is defined as an observation of a relevant bio-physical change of state, made on the Sentinel signal, on a given date.

The first module loads the object-based statistics generated by the backend and formats them according to the analytical requirements. Different sources can be set as data source, including direct database access, flat files (csv for statistics and shapefile/geojson for spatial units) and RestFul Application Programming Interfaces (APIs).

The second module offers a set of preprocessing tools to deal with incomplete and irregular time series (i.e., resampling, interpolation, smoothing, signal combination, cloudy observation removal , noise reduction). These signal processing blocks can be chained to allow the implementation of complex processing schemes before actual marker detection. Several processing chains can be allocated in parallel to simultaneously process Sentinel-1 and Sentinel-2 data. In addition, the code architecture of this framework offers the possibility to improve the overall computational performance of each module by exploiting multi-threading to process observations from several spatial objects in parallel.

The third module analyses the temporal profiles of the relevant signals to detect markers resulting from land management practices. For example, aboveground biomass reduction associated with a mowing activity is expected to result in a drop in NDVI. Similarly, an increase in coherence is also expected. In this respect, several algorithms (detectors) able to identify, in a given time series, a specific feature corresponding to a marker have been implemented based on minima and maxima values of band statistics. The flexibility of the framework developed allows incorporating Machine Learning (ML) approaches for marker detection. For each marker, parameters such as duration, amplitude and period of occurrence are determined. These parameters can be used to exclude false positives and improve detection performance and fine tune the identification of relevant variations according to the specific environmental and climatic context and the practices monitored. Several graphical outputs help users to explore the results and build the most appropriate model.

In the last module, the results from the marker detection run on all the specific bands and derived indices are combined to identify the relevant stages of the land management practices under assessment and obtain a more reliable detection. This approach is demonstrated for mowing detection using NDVI from Sentinel-2 and coherence from Sentinel-1. If an NDVI drop and a coherence increase are found on close time intervals, they can be associated leading to a more reliable activity detection. In a similar way, coherence can be used to compensate for data gaps in the NDVI time series.

The framework proposed demonstrates how to fully exploit the complementary nature of Sentinel-1 and Sentinel-2 observations. It is constantly under development and new features are frequently added. The code is open source and available with its documentation in https://github.com/ec-jrc/cbm. Collaboration and contributions from the scientific and application communities are very welcome.

During the Tropospheric Ozone Assessment Report (TOAR) [1] we built an air quality database that contains time series of measured ozone, ozone precursors, and meteorological data from surface observation stations. One aspect of the TOAR database that substantially contributed to its adoption by the research community, is the augmentation of provider metadata for these stations with globally consistent information derived from multiple Earth Observation data products. This adds additional context to the description of measurement locations and thereby enriches the analysis possibilities. For this we developed a workflow called Geolocation Service that we want to present here.

Our Geolocation Service exposes REST APIs to the user where they can specify an area of interest in the form of latitude, longitude, and possibly radius as well as a specific time where we have data with a time resolution. With the radius parameter it is possible to extract points (no radius) or areas. Different REST API endpoints provide different services, like for example topographic information and nighttime lights. The advantage of REST APIs is that they not only make human interaction possible but also machine to machine communication.

After retrieving the requested data from a performant geodata service (in our case a Rasdaman service), the service can run different analyses which the user can specify and will return any results in a standardized way, namely Geo-JSON. The user can choose between a range of aggregation methods (mean, min, max, etc.) or can choose to return the closest value to the given coordinates. The aggregation method can be specified directly in the REST APIs which makes it very flexible for the user.

Since this workflow consists of modular components, it is easy to exchange or expand some of its parts. We are interested in expanding the available datasets to include the Copernicus Sentinels (i. e. retrieving data from the Copernicus Open Access Hub instead of from our Rasdaman service) to run the existing analyses on those datasets while still providing the user with the same interface and responses as they already know.

To go even further, it would also be possible to include landcover detection, flood mapping, and other spatial analyses via the same geodata workflow.


[1] https://igacproject.org/activities/TOAR

Typically, two thirds of Earth’s surface are obstructed by different types of clouds, by different types of aerosols, and mixtures thereof. The significance of obstruction for the observation of Earth’s surface depends on the amount and optical properties of materials along the optical path but also on the application itself. Ocean applications do have a different sensitivity to obstruction than land and atmosphere applications, and different levels in a processing sequence do have different needs. Some applications require a very clear-conservative cloud masking, whereas others are more tolerant with respect to cloud contamination or even require a cloud-conservative masking. Though the ground segments for OLCI and SLSTR include cloud masking, each of which is based on its native sensor, the present cloud masks are often criticised because they do not fulfil the requirements of some specific applications or users (besides typical weaknesses in problematic cases like semi-transparent clouds, clouds over ice and snow, or coast lines).

EUMETSAT initiated the research study “Sentinel-3 Synergy Cloud Mask Development” to explore novel synergistic cloud masking approaches for OLCI and SLSTR, which demonstrably tackle the problems described above. The study included a thorough review of the cloud masking achievements of European and international scientists working with medium spectral resolution imagery, an evaluation of the operational OLCI and SLSTR cloud masks with respect to the needs of Sentinel-3 users for Level-2 land, marine and atmosphere applications, and an internal EUMETSAT Expert Workshop to critically review and debate the proposed cloud masking concept and its demonstration results.

The consolidated algorithm starts from a very clear-conservative and a very cloud-conservative mask. Subsequent optimal inversion of a simplified radiative transfer equation incorporating spectral wavebands from OLCI and SLSTR to quantify the perturbation of assumed clear-sky reflectance and brightness temperature and to help discriminating between clouds and aerosols plays a key role. The eventual atmospheric masks combine all results and associated uncertainties obtained from the inversion into a comprehensive set of cloud and aerosol categories and associated confidence indicators. The generic concept is applicable to spectral imagers with a set of wavebands reduced to the essentials, too, and proved suitable to develop an initial cloud screening scheme for the future European CO2 Mission.

In this contribution, we explain the concept of the algorithm synergistically using OLCI and SLSTR, and present results from the validation of the atmospheric cloud and aerosol masks. The algorithm will run at EUMETSAT’s in-house offline environment for Sentinel 3, but the concept is easily transferable to other sensors, too.

As the world starts including ecosystem services into the economic reality there is a need for tools that allow us to manage transactions and trust between people and managers of ecosystems. Ecosystem services are extremely complex and hard to measure. And a transparent and easily understandable method for measurement and communication between people and managers of ecosystems is essential.
Here we present a case study of a prototype implementation that allows anyone to buy carbon directly from farmers who have managed to store carbon in their soils. The implementation process relies on a combination of decades of soil research and on engaging users to start a process of interacting with the technology, taking mini steps at a time.
Soils represent one of the largest carbon storage elements on the planet. The combined research efforts of the scientific community have reached a point that now allows us to model changes in soil organic carbon at 30m resolution and understand changes that happen at 5 year timescales. Combining the entire wealth of scientific knowledge with our decades of experience in soil research now culminates in a product with reliable uncertainty indications. Reliable uncertainties that are easy to communicate are essential when engaging farmers and buyers of carbon. For generating predictions and uncertainties of soil organic carbon we use our python eumap package and implement a fully optimized and parallelized machine learning and mapping approach.
Direct contact with people and farmers allows us to work iteratively with user feedback, building trust, so we can develop a prototype of the system that makes sense to everyone involved. During this presentation we discuss the lessons learned and the pitfalls and best-practices that come with implementing a prototype like this.
The final goal of this talk is to show one way we can use the open market and directly connect managers of ecosystem services with buyers. We discuss our experiences in this process so that people in similar positions can learn from the mistakes we have made. Hopefully we inspire other highly technical institutions to try similar things and bring their science to the field and start including the ecosystem into our economic reality.

Fluvial islands are a common geomorphological forms present in almost all major rivers. These are a fundamental component of the riparian corridor, serving as an important ecological habitat for aquatic and terrestrial species. Also, characterized by morphological, economic and even political values, these landforms may provide information about recent geomorphic processes of a river system. Being part of the fluvial system, climate changes and anthropic activities from the last centuries have had implications on the morphodynamics of these landforms. For example, in Europe, nearly all rivers have been altered through channelisation, dam constructions, pollution or overexploitation. As a hydro-morphological consequence, rivers have been flow-regulated to some degree, they have been disconnected from their floodplains, the sediment transport capacity, the flood peaks have been reduced and there have been changes in the number and area of the islands. A modern tool in monitoring fluvial islands is the use of remote-sensing technology. It is known about the amazing potential of the satellite images in the identifying and the monitoring the environmental changes at different scales. Multitemporal and multispectral satellite observation from the Landsat and Sentinel programs are a significant resource, which can allow continuous monitoring landforms and ecosystem evolution. This paper aims to present an approach to monitor fluvial islands by using satellite images, combining with the data from the historical map and data from ”in situ”. The area of interest for this study focuses on the Danube River, in the common Romanian-Bulgarian sector between Giurgiu – Oltenița (Romania) and Russe – Tutrakan (Bulgaria) towns. With a 70 km-long, the studied reach presents 25 vegetated islands, which are covered a total emerged area of 23.96 square kilometres for the mean annual flow. The results show that, for the selected section of the Danube River it is possible to detect vegetated islands and sandbars in the river channel base of the satellite’s data, especially Sentinel-2 imagery. They also make it possible to determine the shape, size and rate of migration of river islands over a period of time. Using QGIS Open Source Software some morphometric parameters were calculated, such as: length, width, area and elongation ratio. Knowledge of the information on the number and the shape of those landforms is relevant for river restoration strategies, nature conservations of islands and maintenance of the protected habitats.

The Black Sea is located in the northeastern part of the Mediterranean Sea. It is a semi-closed basin that communicates with the Planetary Ocean through the Bosphorus and Dardanelles Straits. The water balance is highly imposed by the freshwater inputs from some of the biggest rivers in Europe in terms of solid and liquid discharge: (e.g. Danube). As an endorheic system, the main characteristics that make the Black Sea a special study place are the input of significant freshwater, the lack of strong vertical currents, and the limited water exchange with the Mediterranean Sea.
Earth Observation services for Black Sea Protection (EO4BSP) overlap the entire area of the Black Sea and propose a holistic approach that covers different elements with potential environmental impact. The project is implementing six services that are being delivered to a number of 13 stakeholders from the Black Sea riparian countries and one International organization - The Black Sea Commission. The services will comprise a suite of cloud-based applications which will detect, monitor, analyze and characterize the sources of environmental problems using available Earth Observation imagery in conjunction with in-situ inputs and other reference data. Service applications are developed for deployment in the cloud and shall employ advanced dispersion modeling techniques in conjunction with EO Data to deliver meaningful (actionable) maps, statistics, and other data.

S1 - Land Use Land Cover coastal changes
Economic development is associated with land-use changes, transforming the natural green zones into exclusive anthropogenic areas. Analysis and modeling of land-use change trends and urbanization allow us to evaluate the spatial development patterns providing a key for effective planning practices in the context of Marine Strategy Framework Directive (MSFD) and Maritime Spatial Planning (MSP) implementation.
Service 2 - Eutrophication
Eutrophication represents one of the most severe and widespread environmental problems for coastal zone managers (IOCCG Report Number 3, 2000). In the “Black Sea region briefing - The European environment — state and outlook 2015” published by European Environmental Agency, eutrophication is considered one of the main four key transboundary challenges of the Black Sea.
S3 - Marine Front Identification and mesoscale circulation
This service will include data fusion, satellite observations, numerical modeling, and data assimilation, as well as skill assessment and metrics with a focus on sea state, temperature, turbidity, and SPM, identification of ocean fronts.
EO4BSP will provide services, based on numerical simulation and data assimilation, of currents, salinity and temperature, and distribution, height and period of wind waves, ocean color, sediment transport dynamics, and biogeochemical components as well as the forecast of these parameters.
S4 - Oil Tankers path identification
This service will make use of historical AIS data. Provided by EMODnet, the present data can be used in many ways, not only for oil tankers' path identification but also for illegal trafficking in the Black Sea. S.4 will be used as a decision tool for stakeholders. This application will be intimately linked with the Oil spill monitoring service.
S5 - Oil spills identification and monitoring
Maritime surveillance activities are traditionally carried out by patrol ships or aircraft. However, in recent years the use of synthetic aperture radar (SAR) and optical satellite imagery has proved highly effective in ship traffic and oil spill monitoring. The capability of observing wide areas in almost all-weather and light conditions makes SAR sensors the most suitable tool for maritime surveillance purposes. The service is able to apply the necessary pre-processing steps and an algorithm to Sentinel 1 Ground Range Detected Level 1 data to detect and characterize possible oil spills with tens of meters long.
S6 - High-resolution water quality monitoring in anchorage areas
Monitoring water quality parameters through remote sensing techniques may offer a comprehensive overview of water bodies due to the spatial and spectral capabilities of the sensors. The spatial and temporal distribution of these indicators will reveal the improvement or alteration of the surface water health status. This may be a consequence of nutrients or organic pollution or contamination of waters with hazardous substances. The service will focus on chlorophyll a (chl_a), turbidity, and total suspended matter (TSM). Monitoring the evolution of this parameter at several moments would reveal the anchorage areas aquatic ecosystem's health status.

The main technical objectives of the project are related to the development and demonstration of EO platform capabilities for regional-scale, which can respond to specific needs for:
provision of cloud computing-based solutions and resources
data processing of large amounts of EO data
fusion and integration of multiple sources of information and products, from different satellite products to in-situ and ancillary information
delivery to users (advanced visualization tools, products dissemination workflows, integration of results into users systems)
The concept of the EO4BSP platform implementation architecture focuses on the standard-compliant interfaces that facilitate interoperable data access. These interfaces are part of the components of the platform implementation architecture, which may be assembled themselves of parts specific to platform services as well as software and data that is external to the project.

The EO4BSP platform architecture comprises the following component types: services, processing, data integration, data access, and applications.

Changes in seasonal patterns of natural phenomena occurring on terrestrial naturally vegetated ecosystems are influenced by fluctuations of biotic and abiotic factors taking place on a seasonal as well as annual basis. Changes affecting expected phenology patterns/ cycles for a species and an area, i.e. germination, growing, florescence, herbage, maturation, leaf browning and withering, constitute significant indicators towards the identification of the status (structure and function) of terrestrial naturally vegetated ecosystems. Land surface phenology (LSP) may be used as a well aligned proxy to the observed phenology on the ground.
Time series of spaceborne imageries and their products (e.g. Vegetation Indices - VI) are used to estimate land surface phenology. Higher temporal resolution along with enhanced radiometric sensor features allow for an improved result. This capacity is used within the ENI CBC Joint Operational Programme Black Sea Basin 2014 – 2020 «Copernicus assisted environmental monitoring across the Black Sea Basin - PONTOS» project (under Grand Agreement BSB 889) to elaborate on the assessment of forest cover changes in pilot areas located in Armenia and Georgia. Final goal is to determine whether such changes impact the surrounding environment and ultimately the water quality of the closely located water bodies and supported ecosystems.
LSP metrics and changes are calculated using PhenologyMetrics and PhenologyChanges modules, which were developed by the Centre for Research and Technology Hellas (CERTH) within the completed H2020 ECOPOTENTIAL project. The PhenologyMetrics module enables the generation of phenology related layers of information relying on VI time series covering a vegetation growth period, which commonly corresponds to a one-year period. LSP metrics are estimated per pixel with the exploitation of R phenex package. The PhenologyChanges module facilitates the registering of abrupt changes along the vegetation phenology cycles of sequential years through numerous annual VI series based on the R BFAST (Breaks For Additive Seasonal and Trend) package. This way the dynamics of forest cover loss and gain as a result of human-induced pressures (e.g., fires and deforestation) across the last 2 decades can be registered. Estimated features comprise of the time of canopy emergence, peak and senescence through a vegetation cycle, as well as the identification of time instances, when the phenology cycle seems to be abruptly disrupted (date, number of instances, date of maximum change per pixel).
This study will present a cross-scale comparison of the benefits brought along for the user, when the spatial resolution of the spaceborne input products is increased. In this context, LSP features are produced sequentially by utilizing MODIS Normalized Difference Vegetation Index (NDVI) time series in a spatial resolution of 250m, NDVI time series calculated with Landsat 8 OLI TIRS and Sentinel-2 MSI images. Ground data support the analysis. Results show the benefits of the increment in resolution, while they demonstrate lesser difference between the Landsat and Sentinel-2 ones in dense forested areas. The benefits for each region of interest are highlighted based on an assessment by local actors.

Earth Observation (EO) data acquired by spaceborne sensors are increasingly coupled with cutting-edge technologies to acquire valuable information about changes in the atmo-, hydro-, bio- and geo-sphere in order to assess associated drivers and impacts. Policy implementation and decision-making shall be therewith supported. A vast number of remote sensing multi-source and multi-modal data with various temporal, spatial and spectral resolutions became available and reachable through the open distribution policy of the European Space Agency and Copernicus program, along with the National Aeronautics and Space Administration of the United States of America.
Traditional approaches for acquisition, distribution, storage, management and analysis of EO data are subject to limitations in this new Big Data era. This is related to the volume (e.g., tera to petabytes), variety and heterogeneity (e.g., radar or optical) and velocity (e.g., new data are becoming available on a daily basis). Entrepreneurs, governmental agencies, and supporting infrastructures face the challenge of exploiting the thesaurus of data. Main bottleneck is, beyond the processing and storage capacity, the existence and employment of experts to perform analysis and interpretation. Moving away from conventional data handling, the need for analysis ready data emerged. This way, focus is shed away from data preparation and analysis towards their interpretation and usage. Multi-dimensional data cubes and open source software, allow for communal fast development in the ecosystem of knowledge and knowledge sharing.
Numerous initiatives to create Earth Observation data cubes at national or regional scales have been successful in capturing the momentum of amalgamating remote sensing techniques with information technologies solutions to deliver thematically important and semantically standardized exchangeable information. Open Data Cube (ODC) software paves the way forward as a result of researchers and developers joint operation with the Committee on Earth Observation Satellites (CEOS). ODCs are emerging in increasingly numbers across the world as the result of diverse initiatives and platforms, which aim to support analysis at various spatial scales, storing multi-modal data, using distinct infrastructure and software implementations, as well as different interfaces.
Acknowledging the inherent advantages and the need to harness the potential of Earth Observation and facilitate cross-border transferability of knowledge and information, the Centre for Research and Technology Hellas (CERTH) deployed PONTOS Data Cube (PDC). It is built on CEOS Systems Engineering Office (SEO) Open Data Cube software suite, release 2.21, and it constitutes a full-stack web application adopted for the analysis implementation of large raster datasets. PDC enables cross-regional and intersectoral monitoring of the Black Sea Basin. It was envisioned and is realized within the framework of ENI CBC Joint Operational Programme Black Sea Basin 2014 – 2020 «Copernicus assisted environmental monitoring across the Black Sea Basin - PONTOS» under Grand Agreement BSB 889.
PONTOS Data Cube provides, as a web-service, a common analytical framework comprised by a series of easy-to-access and easy-to-use applications that promote the analysis of EO Big Data time series based on an open source geospatial data management and analysis software and a user-orientated user interface. It aims to strengthen interregional cooperation and development by exploiting offered spaceborne data thesaurus at Black Sea and Mediterranean (initially pilot) regions. The PDC brings along competitive features both for the public and the private sector. These are a) open access to shareable information, b) decreased pre- and processing time, c) minimization of required specialized knowledge to access and process satellite data, d) access to historical data, e) coupling with the latest available tools in the open source and access communities worldwide, and f) validated online services.
The PDC accumulates analysis ready data for four (4) selected areas across Ukraine, Georgia, Armenia and Greece. Environmental and land use practices resulting to coastal line shifting, water quality and related eutrophication events, floating vegetation cover, and maintaining agricultural water balance are of major concern for the region. Focusing on Sevan Lake Basin in Armenia, Rioni River Delta and Kolkheti National Park in Georgia, Nestos River Delta in Greece, and the coastline from Odessa city to the Danube river delta including the Dniester river delta area and adjacent estuary in Ukraine, PDC serves already part of their monitoring needs. The PDC analysis ready data thesaurus comprises of 36 complete years (1984 - 2020) of Landsat-5 ETM, -7 ETM, -8 OLI, and 6 years (2015 -2021) Sentinel-2 MSI images. This archive contains approximately 99,000 images, occupying a total volume of 20.82 TB.
Open Data Cube ready accompanying service tools are embedded in the PDC. The CERTH team further adjusts and aims to incorporate proven online services, such as the WaterMasks and HydroPeriod, which are the legacy of successful research results from the completed H2020 ECOPOTENTIAL project. In close cooperation with local actors PDC’s capacity, along with accompanying tools and components of the PONTOS platform, offered services are being tested for the quality of the products (via in situ validation campaigns) and of the user experience. Status and overall progress will be demonstrated and discussed.

Traditionally, sea level is observed at tide gauge stations, which usually also serve as height reference stations for national leveling networks and therefore define a height system of a country. Thus, sea level research across countries is closely linked to height system unification and needs to be regarded jointly. One of the main deficiencies to use tide gauge data for geodetic sea level research and height systems unification is that only a few stations are connected to permanent GNSS receivers next to the tide gauge in order to systematically observe vertical land motion. As a new observation technique, absolute positioning by SAR using active transponders on ground can fill this gap by systematically observing time series of geometric heights at tide gauge stations. By additionally knowing the tide gauge geoid heights in a global height reference frame, one can finally obtain absolute sea level heights at each tide gauge. With this information the impact of climate change on the sea level can be quantified in an absolute manner and height systems can be connected across the oceans.
The paper presents the results of a project, which was conducted in the years 2019 to 2021 in the frame of ESA´s Baltic+ initiative. Within this project a test network of electronic corner reflectors (ECR) as targets for Sentinel-1 was realized in the Baltic Sea area. The ECR locations were either co-located with tide gauges or with permanent GNSS stations in order to observe systematically the ellipsoidal heights of the tide gauges and possibly also any vertical land motion at the stations. Data for the year 2020 were collected at 10 stations in Estonia, Finland, Poland and Sweden and jointly analyzed with GNSS data, tide gauge records and regional geoid height estimates. The obtained results are promising, but also exhibit some problems related to the ECR´s and their performance. At co-located GNSS stations the estimated ellipsoidal heights agree in a range between about 2 and 50 cm between both observation systems. From the results it could be identified that most likely variable systematic electronic instrument delays of the ECR´s are the main reason for these differences and that each instrument needs to be calibrated individually. Nevertheless, the project provides a valuable data set, which offers the possibility to enhance methods and procedures in order to develop the geodetic SAR positioning technique towards operability. All data and reports are accessible at the following web site: https://www.asg.ed.tum.de/iapg/baltic/

Submerged aquatic vegetation (SAV) maps are of primary importance for the sustainable management of coastal areas and serve as a basis for the fundamental ecological studies. There is also an extensive list of shallow water benthic parameters (e.g. bathymetry, SAV class, SAV cover, species richness, moving underwater sand dunes) that different users would like to have in order to make different spatial planning decisions. Coastal waters of the Baltic Sea are often turbid and the whole sea contains relatively high amount of coloured dissolved organic matter (CDOM). These factors decrease significantly the water depths, where benthic parameters, as well as direct physical impact on sea bottom made by human activities can be detected by remote sensing.
Pärnu Bay in Estonia, Baltic Sea, is highly affected by flux of dissolved and particulate materials from adjacent land transported to the sea by river. As such, Pärnu Bay represents the water body, which has one of the lowest water transparencies in Estonian coastal waters. Field campaigns were carried out to test whether remote sensing can be used for bathymetry and benthic habitat mapping in the area. Concentrations of water column constituents ranged between 2.37-9.68 mg/m3 for Chl-a, 4.45-16.4 g/m3 for total suspended matter (TSM) and 0.82-2.03 for aCDOM(400) m-1. The assessment of benthic substrate detectability limits revealed that the depth restriction for various SAV classes remained between 0.5 and 2.0 m. However, it has to be noted that the fieldwork was carried out in relatively clear parts of the bay, not next to the mouth of the Pärnu river (with huge CDOM inflow) or sandy beaches (with lots of sediment resuspension). Therefore, an average depth restriction in the study area for Sentinel-2 benthic substrate detection was 1.5-2.0 m for benthic vegetation and 2.0-3.5 m for sand.
The benthic parameter we tried to estimate from the Sentinel-2 imagery was submerged aquatic vegetation percentage cover (SAV %cover). IDA image processing software allows to perform several image pre-processing steps (such as atmospheric correction, glint correction) and allows to retrieve water depth and benthic habitats using adaptive lookup table approach. IDA was used for SAV %cover assessment. Empirical regression method, where relationships were established between in situ measured SAV %cover and spectral reflectance (surface reflectance, bottom reflectance), was used as the second method for vegetation coverage mapping. Single band, multiple bands and PCA component (Principal Component Analysis components) were used in regression models. Multiple band regression model, applied to water surface reflectance data, gave the highest R2 (0.66) and lowest RMSE (21.76) values in the 0-2 m water depth range. However, both methods (IDA and empirical regression model) provided generally similar SAV %cover patterns. The regression model was trained to the current image and may not perform so well on other images. IDA image processing software provided slightly lower accuracy (R2=0.64, RMSE=24.10), but it has more potential to work well with other images.

Our study demonstrated the applicability of supervised classifier ensembles to map floating vegetation and beach cast with Sentinel-2 and PlanetScope data at the German Baltic coast.
Drifting vegetation and beach cast create overlays at the otherwise sandy or stony beaches and are particularly important for sandy beach ecology. Drifting vegetation is an aggregation of detached macroalgae (sometimes mixed with seagrass) floating close to the water surface or near the bottom of the coastal sea. Beach cast is unattached macrovegetation that is washed ashore by waves or tides and accumulated on the beach. The species composition of the washed-up vegetation reflects the species assemblage in the nearshore habitats. Beach cast, therefore, provides information on drifting and sessile aquatic habitats that are difficult to access or highly dynamic over time. This is why the sampling of beach cast is an alternative option for investigating the species composition of macrovegetation in coastal habitats. These overlays influence the morphodynamics and structures of the beaches.
To better understand the influence of these patchy drifting vegetation and beach cast habitats, a regular monitoring is necessary. Field surveys are the traditional mapping methods, but they are time-consuming and cost-intensive. Spaceborne remote sensing can provide frequent recordings of the coastal zone. Drifting vegetation and beach cast, however, pose a particular challenge for a remote sensing based monitoring, because they show high temporal dynamics and often a small spatial extent. Our study therefore aims at the monitoring of drifting vegetation and beach cast on spatial scales between 3 and 10 m. In a study area along the Western Baltic coastline of Schleswig-Holstein (Germany), we developed an automated coastline separation algorithm and tested six supervised classification methods and various classification ensembles for their suitability to detect small-scale assemblages of drifting vegetation and beach cast in a study area at the coastline of the Western Baltic Sea using multispectral data of the sensors Sentinel-2 MSI and PlanetScope. The shoreline separation algorithm shows very high accuracies in masking the land area while preserving the sand-covered shoreline.
Generally, Sentinel-2 and PlanetScope data are suitable for the detection of beach cast and drifting vegetation. We could achieve the best classification results using PlanetScope data with an ensemble of different classifiers. This ensemble accomplished a combined f1-score of 0.95. The accuracy of the Sentinel-2 classifications was lower but still achieved a combined f1-score of 0.86 for the same ensemble.

For sea level studies, coastal adaptation, and planning for future sea level scenarios, regional responses require regionally-tailored sea level information. Global sea level products from satellite altimeters are now available through the European Space Agency’s (ESA) Climate Change Initiative. However, these global datasets are not entirely appropriate for supporting regional actions. For the Baltic Sea region, complications such as coastal complexity and sea-ice restrain our ability to exploit altimetry data opportunities.
This presentation highlights the opportunities offered by such regionalised advances, through an investigation by the ESA-funded Baltic SEAL project (http://balticseal.eu/). We present the challenges faced, and solutions implemented, to develop new dedicated along-track and gridded sea level datasets for Baltic Sea stakeholders, spanning the years 1995-2019. Advances in waveform classification, altimetry echo-fitting, expansion of echo-fitting to a wide range of altimetry missions, and multi-mission cross calibration, enabled all available mission data to be integrated into a final gridded product.
The freely available gridded product provided new insights into the Baltic Sea’s mean sea level and its variability. A Mean Sea Surface dataset was also developed as part of the process, in addition to an analysis of sea level trends in the region (using both tide gauge and altimetry data). The Baltic SEAL absolute sea level trend at the coast better aligns with information from the in-situ stations, when compared to current global products. A pronounced sea level trend gradient which increases towards the North-East was found. A proportion of the SL trend gradient can be directly linked to enhanced southerly wind forcing and associated Ekman transport towards the Bothnian Bay. The spatial and temporal density of the data allows for a robust comparison to be made, between the sea level time series and relevant climate indices such as the North Atlantic Oscillation. This has implications for regionalising information on global climate change impacts.
These investigations highlight the potential of regionalised products for the Baltic Sea region, and beyond. The availability of multi-mission along-track data, and gridded data, offer a wide range of opportunities. These range from supporting local ocean circulation research, to storm surge monitoring. Such opportunities are presented to promote further exploitation, and identify synergies with other efforts focused on relevant oceanic variables, and societal applications.

The use of GNSS positioning is growing in various applications and the awareness of space weather impact on GNSS observations is increasing.
Improving the understanding and characterisation of the effects of space weather phenomena on the Earth and in the space can increase situational awareness, inform decision making, and enable missions to be carried out that depend on technologies and services susceptible to disruption from space weather [1].
In our study, performed at the Institute of Geodesy and Geoinformatics, University of Latvia (GGI), the Latvian CORS ground-based GPS observations were collected during the 24th solar cycle.
The main objective of the present study is to perform an analysis of the space weather impact on the Latvian CORS (Continuously Operating Reference Stations) GPS (Global Positioning System) observations, in situations of geomagnetic storms, sun flares and extreme TEC (Total Electron Content) and ROTI (Rate of change of TEC index) levels, by analysing the results, i.e., 90-s kinematic post-processing solutions, obtained using Bernese GNSS Software v5.2. To complete this study, the 90-s kinematic time series of all the Latvian CORS for the period from 2007 to 2017 were analyzed, and a correlation between time series outliers (hereinafter referred to as faults) and extreme space weather events was sought. Over 36 million position determination solutions were examined, 0.6% of the solutions appear to be erroneous, 0.13% of the solutions have errors greater than 1 m, 0.05% have errors greater than 10 m, and 0.01% of the solutions show errors greater than 50 m. The correlation between faulty results, TEC and ROTI levels and Bernese GNSS Software v5.2 detected cycle slips was computed. This also includes an analysis of fault distribution depending on the geomagnetic latitude as well as faults distribution simultaneously occurring in some stations, etc. This work is the statistical analysis of the Latvian CORS security, mainly focusing on geomagnetic extreme events and ionospheric scintillations in the region of Latvia, with a latitude around 57°N.
The statistical data of the results of the space weather impact on GPS observations are presented in this study. Conclusions on the security level of the Latvian CORS will be drawn on the basis of these statistics. At the end of this study, Pearson’s correlation analysis is performed on the relation characteristics of both the TEC and the ROTI to the impacted GPS positioning discrepancies. The assessment of the TEC and ROTI irregularities will be discussed.

Finnish Environment Institute SYKE has long experience in estimating and providing a wide range of Earth Observation (EO) data. Among these are chlorophyll a, turbidity, sea surface temperature and surface algal blooms in the Baltic Sea with medium resolution instruments. The high spatial resolution observations of Sentinel-2 MSI instrument (10-20m) and Landsat-8 OLI (30m) has allowed SYKE to extend the provision to Finnish coastal regions and thousands of lakes. Other novel products such as Secchi depth and absorption of CDOM (colored dissolved organic matter) have also been added to SYKE’s EO catalogue. Together with water quality products interest for true color imaginary (RGB) has also grown fast. A web application TARKKA (www.syke.fi/tarkka/en) was created to meet the data provision needs of this information source. It visualizes and provides various kind of optical satellite data to end users.

TARKKA is based on open-source solutions and utilizes SYKE’s open data, as well as EO data and open GIS data. Depending on the satellite overpasses and cloudiness, the water quality data products provided in SYKE’s Web Map Service (WMS) service cover only part of Finnish coastal waters and lakes. Cloud cover prevents some of the useful observations, but still the amount of data is huge. In the future as a new version of TARKKA will be revealed, the development will be emphasized on user orientation, easier access to cloudless observations from large number of observations and to provide more understandable combinations of data.

SYKE’s processing environment (CalFin, build on Calvalus parallel processing system) enables calculation of water quality data from every overpass, although only demonstration products based on user interest are visualized on map. CalFin have also been utilized to calculate daily statistical water quality data for many kinds of areas of interest on Finnish lakes and coastal regions as well as for the whole Baltic Sea. This data is available for users in the national environmental administration via STATUS database and user interface. TARKKA provides part of this statistics as validation information to give confidence on EO product quality for users. The validation information is presented as time series of EO and station observations at reference stations (currently over 200 lake stations and 30 coastal stations). The amount of reference stations is growing steadily. As the time series of new instruments was getting longer, more flexible visualization of time series graphs in TARKKA was introduced. In the future more of this statistical data which currently is available only in STATUS will be made publicly available in a new version of TARKKA.

TARKKA has been improved in many ways during recent years. For example, as a new feature a front page was introduced to TARKKA. The purpose of the front page is to highlight some of the interesting data and phenomena. Many new datasets were also included, high resolution sea surface temperature data as latest of them. A gallery for algae monitoring on Finnish lakes have lately been included as a part of TARKKA service. The most utilized part of TARKKA is still the true color image catalogue. It has increased citizen and expert interest in EO data and TARKKA service. As public interest grew fast, a new daily true color image service based on the data interfaces of Sinergise Company was included in TARKKA in the beginning of summer 2018.

The complete true color imagery of Sentinel-2 and Landsat-8 has proven to be effective for observing various changes over water areas but also rapid changes in land cover. In the beginning the main interest was algal bloom monitoring of the Baltic Sea but soon also lake and costal ice cover also become a target for public interest. For several years true color imagery has been used for monitoring and visualizing various phenomena, examples include dredging, resuspension events, aquatic vegetation of lakes, pollen, effects of stormy weather on Finnish coastal waters, seasonal variations of lake’s water quality, riverine impact areas etc.

The development of TARKKA web map application and the EO data it contains have been funded by a chain of national projects and international co-operation: TARKKA is one of the user interfaces utilized in EU Horizon 2020 EOMORES project. The basics of the data products and technical solutions were formed in national projects, mainly VESISEN&VESISEN II, and ENVIBASE. TARKKA’s future development is part of national SYTYKE and CorEO projects as well as international co-operations within ESA BalticAIMS and EO-Crowd projects.
In this poster presentation practical examples of using TARKKA for environmental monitoring of water quality and land cover will be shown. In addition, the use of the system for Marine Spatial Planning purposes will be demonstrated. This is done by allowing the user to combine various EO, model, GIS and in situ data sources for analyzing the impact of human activities (agriculture, coastal construction, dredging, shipping etc.) on the coastal environment.

The sea surface salinity (SSS) has a critical role in determining ocean state, marine pollution, and climate change. Hence, persistent monitoring of the SSS is needed to detect any unusual changes that can endanger marine ecosystems.
Currently, SSS is measured globally by multiple space-borne sensors, allowing to obtain a higher spatial and temporal coverage, as compared to the conventional observation techniques (such as in-situ sampling). Accordingly, the focus of this study is to examine the performance of SSS data that are based on the Copernicus Sentinel-2 and NASA’s soil moisture active passive (SMAP) missions. A numerical study is conducted over estuarine waters of the Baltic Sea (located in northern Europe). The main source of fresh water of the Baltic Sea originates from the continental rivers and salty water infiltrates from the Atlantic Ocean via the narrow connection (via Danish Straits) with the North Sea. This often leads to the formation of strongly stratified waters both horizontally and vertically. This sea area is one of the most challenging to retrieve satellite SSS data from mainly due to the basin's highly stratified nature. Previous studies have also shown disparities amongst the various sources (satellites, in-situ and hydrodynamic models) of SSS in the Baltic Sea.
Several studies have been conducted to analyze the performance of modeled and satellite-derived SSS in the Baltic Sea, but only a few have examined the performance of sentinel-2 and SMAP based SSS variations. Moreover, SMAP and sentinel-2 research on SSS is still in its primary stages, although there are numerous potential and application possibilities. This study aims at taking a closer examination on this by also comparing it with in-situ measurements and hydrodynamic model.
There are two Sentinel-2 SSS data products accessible for users: Level-1C (L1C) Top-of-Atmosphere (TOA) and Level-2A (L2A) Bottom-of-Atmosphere (BOA), which allow retrieval of SSS. The SMAP L3-V5 SSS product used in this study was obtained from NASA, Jet Propulsion Lab. In this study for the year 2020 monthly Sentinel-2 SSS products shall be compared from a statistical perspective with those of SMAP along with in-situ measurements and that of the Nemo Nordic ocean-sea ice model. The results of this study will be helpful to understand the seasonal trends of SSS in the Baltic Sea and also the validation of hydrodynamic models. These results are expected to be instrumental in enhancing our understanding of the marine dynamics for the Baltic Sea.

References
Gambau, V., González, Olmedo, E., Martínez, J., 2019. Challenges of Retrieving Sea Surface Salinity over the Baltic Sea, in: Geophysical Research Abstracts 21. Presented at the EGU2019, European Geosciences Union, Vienna, Austria.
Medina-Lopez, E., Ureña-Fuentes, L., 2019. High-Resolution Sea Surface Temperature and Salinity in Coastal Areas Worldwide from Raw Satellite Data. Remote Sens. 11.
Rajabi-Kiasari, S., Hasanlou, M., 2020. An efficient model for the prediction of SMAP sea surface salinity using machine learning approaches in the Persian Gulf. Int. J. Remote Sens. 41, 3221–3242. https://doi.org/10.1080/01431161.2019.1701212
Wu, Q., Wang, X., Liang, W., Zhang, W., 2020. Validation and application of soil moisture active passive sea surface salinity observation over the Changjiang River Estuary. Acta Oceanol. Sin. 39. https://doi.org/10.1007/s13131-020-1542-z

In the frame of the ESA Regional Initiative, the Baltic+ contracts aim at developing research activities to advance the use of ESA and non-ESA Earth observations missions towards the achievement of major scientific challenges identified by Baltic Earth community for the next decade[1]. This includes, in particular:
• Dedicated products for the Baltic: ocean colour, sea level, coastal altimetry, salinity, and new dedicated Sentinel-2 products.
• Characterisation of biochemical exchanges (land-sea and air-sea) including salinity dynamics.
• Characterising and closing the water cycle of the Baltic.
• 4D reconstruction of ocean dynamics by integration of EO and modeling of the Baltic Sea.
• Characterising and predicting major Baltic inflows.
Baltic+ Salinity Dynamics and Baltic+ SEAL (Sea Level) projects have contributed to these challenges by developing dedicated products of sea surface salinity (SSS) and sea surface height (SSH), respectively. Here, we will explore the potential synergy between both products with a twofold purpose: to perform an inter-validation of both products, and to explore the potentiality of both datasets to address some of the scientific challenges identified by ESA and the Baltic Earth community.
We will first explore the detection and monitoring of the Atlantic salinity inflow and its recirculation inside the basin by presenting a preliminary assessment of the consistency between structures detected in Baltic+ Salinity SSS maps and circulation patterns derived from Baltic+ SEAL altimetry observations [2,3]. As an example of application, we will analyse how SSS and SSH reflect the mean flow condition across the Danish straits and how they react by local wind conditions and larger atmospheric circulation patterns[4]. While ideally the characteristics of the full water column would be needed, the combination of SSS and altimetry data can help monitoring the inflow and the distribution of surface waters characterised by different densities.

[1] https://eo4society.esa.int/regional-initiatives/baltic-regional-initiative/baltic-regional-initiative-science/ (last access 24th November 2021)
[2] https://eo4society.esa.int/2021/03/22/expanded-capability-to-monitor-sea-level-at-higher-latitudes-with-new-ssh-products/ (last access 24th November 2021)
[3] Passaro M, Müller FL, Oelsmann J, Rautiainen L, Dettmering D, Hart-Davis MG, Abulaitijiang A, Andersen OB, Höyer JL, Madsen KS, Ringgaard IM, Särkkä J, Scarrott R, Schwatke C, Seitz F, Tuomi L, Restano M and Benveniste J (2021) Absolute Baltic Sea Level Trends in the Satellite Altimetry Era: A Revisit. Front. Mar. Sci. 8:647607. doi: 10.3389/fmars.2021.647607
[4] Lehmann, A., W. Krauß, and H.-H. Hinrichsen (2002), Effects of remote and local atmospheric forcing on circulation and upwelling in the Baltic Sea, Tellus, Ser. A, 54(3), 299–316.

This study investigates whether data mining methods can be applied to satellite data, weather data and field measurements to create proxies for forecasting bacterial (vibrio ssp.) abundances along the Baltic coast of Schleswig-Holstein, Germany.
Vibrios are gram-negative, bacillary shaped, polar-flagellated bacteria that are moderately to markedly halophilic and pervasive in water bodies. The most popular kind will be V. cholerae O1/O139 causing severe cholera disease. These kinds of V. cholerae are not endemic in German coastal waters but several, non-cholerae species occur, which are also human-pathogenic. This study investigates the abundance of the V. vulnificus in the Baltic Sea region which can cause wound infections to fatalities. Massive population increase in bathing areas constitutes a high risk especially for immunodeficient people hence the detection is really important. Vibrios are pervasive not only in the Baltic Sea but also in the North Sea, brackish waters and haline estuaries. Recent studies show massive population increase of V. vulnificus in areas with water temperatures above 20°C and a prevalent salinity of 0.5 to 2.5% (5 to 25 PSU, Practical Salinity Unit; Baltic Sea at Schleswig-Holstein has a prevalent salinity of 13 to 18 PSU). In the last years reported incidents of vibrio infections increased along the coast of Schleswig-Holstein from June to September. Recent studies show the connection of increasing summer temperatures and the resulting warming of the Baltic Sea water to the strong increase of V. vulnificus populations. Sampling for vibrios by authorities in bathing areas only takes place when there is a strong suspicion i.e. hospitalized patients with typical symptoms. In addition to that there is no threshold signaling a higher pathogenicity. In our study, we apply a different approach, i.e., data mining techniques, to assess the risks of V. vulnificus infections. For that, we analyze time series Earth observation data and derived products (sea surface temperature, salinity), weather data (e.g. air temperature, rainfall) and in situ samples of V. vulnificus to search for causal relationships and patterns. We also re-evaluate the spectrum of parameters put into consideration for investigating V. vulnificus population increase. Other than traditional approaches, machine learning and process analytics give novel insights into the use of remote sensing and other environmental data for a better understanding of V.vulnificus increase in coastal bathing areas. Such outcomes can also assist to develop a forecasting of V.vulnificus abundance, which is of high interest for public authorities and seaside resorts to prevent hospitalizations.

Clouds are essential components of the climate system and significantly affect the available solar radiation. They greatly influence Earth's energy budget by increasing the albedo and absorbing its longwave radiation. Besides that, solar radiation is an essential part of future energy production. Its share will have to rise from the current few percentages to fulfil the Paris Climate Agreement objective to keep global warming below 2 degrees. Although cloud coverage is constantly changing, the long-term changes affect the energy budget the most and must be considered while planning for changes in energy production. It is therefore essential to study these changes. Besides summer, It is crucial to look at changes in CFC in spring and autumn when the incoming radiation is weaker in higher latitudes, and the role of clouds is, therefore, more significant.
We have used the EUMETSAT's satellite-based cloud climate data record CLARA-A ed. 2.1 from Satellite Application Facility on Climate Monitoring (CM SAF). Daily mean cloud fraction cover (CFC) with 0.25° x 0.25° grid from 1982 to 2019 was chosen to study the changes in cloud cover in March. The region included an area from 52 to 70° N and 5 to 32° E, covering the Baltic Sea basin and most surrounding countries. It extends our previous work covering the same area but shorter time scale and different aspects. In the initial study, we also used cloud observations from the Tartu-Tõravere meteorological station, Estonia. CFC data from CLARA-A agreed with in situ data regarding the maxima and minima years and a downward trend in March over the 1982–2015 period.
In March, the mean CFC for the entire region from 1982 to 2019 was 71.8 % (std. 4.9 %, median 70.8%). The mean values are higher over the Norwegian Sea and lower over the Baltic Sea. Geographically, the statistically significant decrease in CFC is observed over most of Sweden, Finland and Denmark, Southern Norway, and the middle part of the Baltic Sea.
To analyse the changes in cloud cover, five different groups of days were looked at, based on the CFC values: 1) clear (CFC = 0); 2) almost clear (CFC < = 5); 3) low cloud cover (5 < CFC < = 30); 4) almost overcast (CFC > = 95) and 5) overcast (CFC = 100). There are more clear or almost clear days over the sea area than land and more in the South. On the other hand, days with low cloud cover are more common over land, especially in southern parts of Finland, Sweden and Norway. There are fewer days with overcast or fully overcast clouds covering Norway's mountainous region, and the highest number of overcast days are over the Norwegian Sea.
The change in total CFC comes mainly from the decrease of almost overcast and overcast days since 1982. The number of days with overcast or almost overcast days has decreased practically everywhere in the region, with maximum change -3.7 and -4.5 % per decade accordingly. However, since 2013, the number of almost overcast and overcast days has increased in the study area. No such rapid change is visible for the other groups. The number of clear and almost clear days are practically stable, while there is a slight increase in the number of days with low cloud cover for the entire period.

The spatio-temporal sea level variation is a key indicator of our climate and applicable in many fields. It can be deduced from the tide gauge (TG) data which are providing continuous information about the sea level at one constraint specific location. Satellite Altimetry (SA) technique, on the other hand, allows the instantaneous along-track sea level observations. The SA data accuracy however needs to be continuously assessed considering the SA limitation to retrieve precise estimates of sea level. To accomplish this Hydrodynamic Models (HDM), which can reproduce the state of the sea and help to improve our understanding of interrelations between different phenomena, can be utilized. Furthermore, high resolution geoid models maximize opportunities that allows improvements in deriving realistic sea-level using SA.
This study provides a thorough assessment of the performance of ESA funded Baltic+ SEAL project SA along track data for Sentinel-3A, Sentinel-3B and Jason-3 over entire Baltic Sea (BS) by utilizing tide-gauge calibrated Nemo-Nordic HDM and the NKG2015 marine geoid model. The HDM data have been corrected at the location and time of SA passes using TG records. The 2017-2019 SA data performance is then monitored via both calibrated HDM and TG datasets utilizing the NKG2015 quasi-geoid model.
The comparison provides specific insight into the validation of the SA derived dynamic topography (DT). Moreover, it allows an estimation of the HDM accuracy. This is a necessary precondition if the model is employed in geodetic applications, or if a prediction of the model-based DT is performed. Finally, retrieved accurate offshore DT measurements in local region could be beneficial to demonstrate the marine geoid and lead to more precise HDM computation. The statistics spatio-temporal performance of the comparisons performed between the SA and the ground truth in the study area. The analysis of different datasets has been conducted for the entire Baltic Sea. Our assessment intention is to identify the spatio-temporal variations which SA have relatively better performance in comparison with the ground truth, it is evident that the performance of the satellite radar altimetry missions has been improving
The inter-comparison of SA missions outcomes confirmed that better results are obtainable by SAR mode (S3A and S3B) comparing to LRM mode mission, although much larger discrepancies were found between filtered SA data and the calibrated HDM with respect to those obtained for the S3A and S3B mission. This fact emphasizes that the JA3 dataset is affected by a higher measurement noise than S3A, and that SRAL instrument onboard of the Sentinel-3 missions solves better the signal in the coastal

The increasing availability of remote sensing data allows the quantification of biodiversity in space and time. In particular, remote sensing of spectral diversity, defined as the variability of electromagnetic radiation reflected from plants, is linked to biodiversity. However, phenology and land management in grasslands lead to variations in remotely sensed data and, in turn, remotely estimated plant biodiversity. Nevertheless, our knowledge on how to use datasets sampled at multiple points in time to quantify biodiversity is scarce, although such an approach might extend the accuracy and informative power way beyond the use of a dataset obtained for a single point in time. Therefore, we introduced a new spatio-temporal approach based on a dissimilarity measure accounting for spectral diversity in space and time. The so calculated spectral diversity of a region includes the distinct contribution of space, time, and their interaction. From the total spectral diversity between plants, regardless of their space and time allocation, the approach allows to allocate the overall diversity to different sources of variability, e.g., level of ecological organization or stages of the growing season.
We illustrated our methodological approach with a case study that estimated taxonomic beta-diversity from Sentinel-2 data for managed grasslands which differ across a large gradient of environmental properties. In particular, the contribution of different plant communities to beta-diversity was assessed, and the results were validated against a dataset of in-situ measured beta-diversity from plant surveys. Compared to spatial dissimilarities from distinct stages of the growing season, using spatio-temporal dissimilarities between communities produced a more accurate estimation of the uniqueness of a community. In particular, the root-mean-square error could be reduced by up to 74% when estimating community contributions to beta-diversity from space by using multi-temporal datasets that cover the entire growing season compared to using mono-temporal datasets. Overall, our study shows how to account for temporal variations in the spectral diversity of plant communities and demonstrates that this improves the estimation of plant biodiversity through remote sensing. Grasslands are dynamic ecosystems and building on multiple datasets in time covering phenological variations as well as management events is necessary for a proper estimation of biodiversity.

Understanding the spatial distribution of habitat and how it varies over time constitutes essential information for managing and conserving wildlife. Human transformation, fragmentation, and modification of habitats, as well as natural disturbances are key drivers of habitat dynamics and through this, wildlife population dynamics. Spatially-explicit and temporally continuous information on species-specific habitat dynamics, however, is commonly lacking, as habitat assessments typically provide only single, one-time snapshots of habitat conditions (Franklin 2010). Coupling spatially explicit habitat models with satellite time series provides unique opportunities to overcome this limitation and allow for a continuous monitoring of habitat dynamics in space and time (Randin et al. 2020).
We will present a novel approach for monitoring wildlife habitat dynamics based on multi-decadal Landsat satellite time series. Specifically, we combined animal tracking data obtained from long-term monitoring with time series of Landsat-based spectral-temporal metrics in habitat models for mapping dynamics in habitat suitability over three decades (Oeser et al. 2019, Oeser et al. 2021). We applied this approach for assessing forest disturbance-related habitat dynamics for different European large herbivores (Red deer, roe deer, and European bison). As natural forest disturbances, such as bark beetle outbreaks, windstorms, and fires, are increasing in many forests across the globe as a consequence of climate change (Seidl et al. 2017), understanding their impact on wildlife is critical for devising appropriate management and conservation strategies. In this context, large herbivores are particularly important, as they are strongly affected by forest disturbances and in turn, through their foraging activities, themselves can affect forest structure and successional dynamics (Kuijper et al. 2009).
Our research demonstrates the large potential of Landsat time series for monitoring wildlife habitat dynamics continuously across space and time, yet also points to methodological challenges, particularly in terms of ensuring the consistency of mapped habitat changes due to noise inherent to satellite image time series. To overcome these challenges, we propose an approach for harmonizing spectral-temporal metrics across time through the application of a time series segmentation algorithm (Kennedy et al. 2010). Applying this approach considerably improved the consistency of the derived habitat time series and improved model transferability in time. Ecologically, our approach allows us to provide novel insights into how large herbivores – both widespread ones and species of conservation concern – respond to forest disturbance and successional dynamics over multiple decades, providing crucial information for forest management and wildlife conservation.




Franklin, J. 2010. Moving beyond static species distribution models in support of conservation biogeography. Diversity and Distributions 16:321-330.
Kennedy, R. E., Z. Yang, and W. B. Cohen. 2010. Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr — Temporal segmentation algorithms. Remote Sensing of Environment 114:2897-2910.
Kuijper, D. P. J., J. P. G. M. Cromsigt, M. Churski, B. Adam, B. Jędrzejewska, and W. Jędrzejewski. 2009. Do ungulates preferentially feed in forest gaps in European temperate forest? Forest Ecology and Management 258:1528-1535.
Oeser, J., M. Heurich, C. Senf, D. Pflugmacher, E. Belotti, and T. Kuemmerle. 2019. Habitat metrics based on multi-temporal Landsat imagery for mapping large mammal habitat. Remote Sensing in Ecology and Conservation 0.
Oeser, J., M. Heurich, C. Senf, D. Pflugmacher, and T. Kuemmerle. 2021. Satellite-based habitat monitoring reveals long-term dynamics of deer habitat in response to forest disturbances. Ecological Applications 31:e2269.
Randin, C. F., M. B. Ashcroft, J. Bolliger, J. Cavender-Bares, N. C. Coops, S. Dullinger, T. Dirnböck, S. Eckert, E. Ellis, N. Fernández, G. Giuliani, A. Guisan, W. Jetz, S. Joost, D. Karger, J. Lembrechts, J. Lenoir, M. Luoto, X. Morin, B. Price, D. Rocchini, M. Schaepman, B. Schmid, P. Verburg, A. Wilson, P. Woodcock, N. Yoccoz, and D. Payne. 2020. Monitoring biodiversity in the Anthropocene using remote sensing in species distribution models. Remote Sensing of Environment 239:111626.
Seidl, R., D. Thom, M. Kautz, D. Martin-Benito, M. Peltoniemi, G. Vacchiano, J. Wild, D. Ascoli, M. Petr, J. Honkaniemi, M. J. Lexer, V. Trotsiuk, P. Mairota, M. Svoboda, M. Fabrika, T. A. Nagel, and C. P. O. Reyer. 2017. Forest disturbances under climate change. Nature Climate Change 7:395-402.