The evident effects of climate change require the systematic and timely monitoring of the arable lands and grasslands. The new era of the Common Agricultural Policy (CAP2020+) has introduced the concept of the exhaustive monitoring, as opposed to the sampled-based current operating model, and the optimization of its Integrated Administration and Control System (IACS). The new CAP will attempt to shift towards a more regional based ensemble of regulations, according to the requirements and the particularities of the local ecosystems. These are mentioned as Member State (MS) specific eco-schemes for Pillar I and agri-environmental measures for Pillar II. The Sentinel satellite missions have been acknowledged as a key component for the provision of reliable solutions for CAP monitoring, thanks to the high spatial and temporal resolution data that they freely offer.
There is a number of studies that have used Sentinel data, and particularly Sentinel-1 and Sentinel-2, in the context of CAP-specific grassland monitoring. Monitoring of grassland activity and the detection of abrupt changes, such as mowing and grazing events, is required for the compliance verification that is accordant to CAP’s direct payments rules (Pillar I), as well as for the conceptual design of targeted agro-ecological and climate-focused policies (Pillar II). In order to detect abrupt changes on grassland we need uninterrupted optical imagery time-series, i.e. from the Sentinel-2 mission. Nevertheless, the continuity of the Sentinel-2 time-series is often hindered by cloud coverage, which is even more of an issue for Northern European countries. In order to tackle this problem, many research studies suggest the complementary usage of Sentinel-1 Synthetic Aperture (SAR) data, which are mostly weather independent, when Sentinel-2 data are affected by clouds.
In this work, we propose deep learning architectures in order to i) generate dense cloud-free Normalized Difference Vegetation Index (NDVI) time-series at the pixel level and ii) precisely track the occurrences of abrupt changes that can be characterized as mowing events. Moreover, we propose a scheme for the quantification of the grassland use intensity that is based on the frequency of mowing events, predicted at the previous step. The area of interest for this study is the country of Lithuania for the year 2020. Specifically, in order to generate dense NDVI time-series, we implement a Recurrent Neural Network (RNN) architecture that takes as input Sentinel-1 time-series along with any available, namely cloud-free, Sentinel-2 images. Consequently, these artificially produced NDVI time-series are given as input to a similar RNN architecture, aiming to accurately detect mowing events.
Finally, the pixel-level decisions are aggregated indicating the agricultural activity at the parcel level. The overall evaluation of the proposed methodology was achieved using annotated event instances that were generated through a sophisticated photo-interpretation process, performed by three experts.
Acknowledgements: This work has been supported by the EU's Horizon 2020 innovation programme under grant agreement H2020-869366 ENVISION.

European rural landscapes are under pressure due to rural-urban migration, change of livelihood strategies, socio-economic and institutional changes regarding land use, climate change, and invasion of alien species (IAS). Agricultural land abandonment is probably, the most widespread land change process in Europe and affects both croplands and grasslands. Agricultural land abandonment may result in the formation of novel ecosystems with positive but also negative impacts on the environment. Evidence shows that IAS may spread over abandoned agricultural lands (Masemola et al., 2020), resulting in diminishing the ecosystem services and thus, have significant negative implications for human livelihoods and human well-being. Among IAS common in Europe is Giant Hogweed (Heracleum Sosnowskyi), further, H. Sosnowskyi. It was introduced early 20th century in some European countries as an ornamental plant originally from the Caucasus. During the 1950s, it has also been planted as a livestock fodder crop due to the substantive biomass potential of this plant. Following the demise of the Soviet Union, H. Sosnowskyi went out of control and is now rapidly spreading over in Eastern European countries. H. Sosnowskyi aggressively spreads, it is difficult to control H. Sosnowskyi. H. Sosnowskyi is also dangerous to humans and animals. 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- SAR Sentinel-1 and optical Sentinel-2 satellites, and therefore, extending the availability of optical Landsat observations dating back to 1980s. Our major goal was to document the agricultural land-cover change from 1990 to 2020 in temperate Eastern Europe (with a strong focus on European Russia and neighboring countries), namely transitioning of managed grasslands and croplands to abandoned lands at a different stage of natural succession, but also to reveal the spread of H. Sosnowskyi and urbanization. We tested the suitability of random forest classifier as well as U-Net architecture for convolutional neural networks. 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, attracting trained volunteers and open-access crowdsource information. We employed available Landsat-based land-cover maps circa 1990 from the University of Maryland (Potapov et al., 2014) and circa 2015 from Greenpeace-Russia (Glushkov et al., 2021) to attribute the transitions from managed cropland and grassland to different stages of agricultural land abandonment for a Russia (European part) and neighboring countries. Further, we complemented the analysis with detailed elaboration on land use and land cover by 2020 with the analysis of Sentinel-1 and Sentinel-2 time series and zoomed-in Moscow province of Russia to understand the spread of H. Sosnowskyi. Our broad-scale mapping of land-cover change using Landsat-based from 1990 to 2015 showed a widespread abandonment of croplands and grasslands, particularly in European Russia, albeit with remaining cultivation spots, such as in Vladimir and Kaluga provinces, but also in the forest-steppe zone. Many abandoned lands became naturally afforested, yet the capacity still exists for further natural afforestation. The results of broad-scale analysis correlated well with change observed from the official agricultural statics at the province level. The detailed zoom-in to Moscow province showed that only a small portion of managed grasslands in 1990 remained mown by 2020. Formerly managed croplands and grasslands transitioned to unmanaged grasslands, areas with early shrub succession and young forest, but were also contracted due to urbanization and the spread of H. Sosnowskyi. Out of 1,790 thousand hectares of managed agricultural lands, approximately 8% were encroached by H. Sosnowskyi by 2020. From a remote sensing aspect, Sentinel-2 served as a source of data to develop monthly cloud-free composites. Preliminary analysis showed, while the add-on of Sentinel-1 derivatives did not result in higher classification accuracies, once it has been added to Sentinel-2 composites, nevertheless VV, and VH polarizations were useful to distinguish mowing activities. Feature analysis with random forest showed a contribution of Sentinel-2 SWIR bands and simple seasonal metrics, such as standard deviation calculated from May 1st to October 1st. We also tested and did not find a significant contribution of the Hogweed Index (which is based on the difference between green ad NIR bands) because it is probably represented with uncondensed information already available in the classified layer stack. Our further 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.

In sum, our study showed a massive agricultural land-use change transformation after 1990, resulting in cropland and grassland abandonment with subsequent shrubs and young trees encroachment. However, alternative trajectories of grassland contraction, as the case study in Moscow province showed, primarily urbanization and invasion of H. Sosnowskyi, additionally reduces available grasslands and meadows in temperate Europe. Our study showed a great advantage of using Sentinel-1 and Sentinel-2 time series, particularly the latter one, to map shrubs and young trees encroachment as well as the spread of H. Sosnowskyi. The invasion of H. Sosnowskyi is worrisome as it already affects many European countries-for instance in Denmark, Germany, France, the Baltic States, Russia and Ukraine. Therefore, our study shows how H. Sosnowskyi and the transition of grasslands can be monitored timely. Last but not least, we postulate, the withdrawal of lands from agricultural land use and restoration of vegetation should be steered rather left for “nature”, as it may result in unintended invasions of alien species, such as H. Sosnowskyi.

Cited literature
Glushkov, I., Zhuravleva, I., McCarty, J.L., Komarova, A., Drozdovsky, A., Drozdovskaya, M., Lupachik, V., Yaroshenko, A., Stehman, S.V., Prishchepov, A.V., 2021. Spring fires in Russia: results from participatory burned area mapping with Sentinel-2 imagery. Environ. Res. Lett. 16, 125005. https://doi.org/10.1088/1748-9326/ac3287
Masemola, C., Cho, M.A., Ramoelo, A., 2020. Sentinel-2 time series based optimal features and time window for mapping invasive Australian native Acacia species in KwaZulu Natal, South Africa. International Journal of Applied Earth Observation and Geoinformation 93, 102207. https://doi.org/10.1016/j.jag.2020.102207
Potapov, P.V., Turubanova, S.A., Tyukavina, A., Krylov, A.M., McCarty, J.L., Radeloff, V.C., Hansen, M.C., 2014. Eastern Europe’s forest cover dynamics from 1985 to 2012 quantified from the full Landsat archive. Remote Sensing of Environment. https://doi.org/10.1016/j.rse.2014.11.027

Meteorological extreme events such as droughts have severe impacts on grassland ecosystems globally. Droughts can result in harvest loss, grassland degradation, and impacts on species composition. As a consequence of ongoing climate change, drought periods will become more frequent, more severe, and longer in the coming decades in many regions of the world. In Central Europe, grassland vegetation substantially deteriorated immediately in response to extreme droughts in recent years with major impacts on livestock farming. Time series of vegetation indices (e.g., Normalized Difference Vegetation Index) were most commonly used to analyze changes in grassland vegetation during drought periods from remote sensing data. Yet, a vegetation index value as such does not allow the differentiation of percent ground cover of different vegetation and non-vegetation land cover fractions. Fractional cover provides physically-based estimates of vegetation dynamics during droughts, thereby contributing to a better understanding of drought effects on grasslands and their spatial variability.
We developed a new approach to assess drought impacts in grasslands based on intra-annual time series of fractional vegetation cover. Specifically, our method produces intra-annual time series of the Normalized Difference Fraction Index (NDFI) from Sentinel-2. The NDFI is based on fractional cover estimates of photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV) and soil. It allows for an integrated assessment of the three fractional cover components of grasslands by contrasting NPV and soil relative to PV. The NDFI is specifically sensitive to intra-annual drought effects, as it takes a pre-season NPV base value into account, thereby capturing the drought-relevant changes of NPV fractional cover within each growing season. In our previous study in a major Central European grassland region (Kowalski et al., 2022), we showed that (i) fractional cover time series of PV, NPV and soil can be reliably derived from spectral unmixing using a constant set of library spectra representative of the entire time series and study site and (ii) that the calculated NDFI time series tracked vegetation changes during drought periods well in line with several meteorological and hydrological drought indices. Beyond that, the high spatial resolution of the Sentinel-2-based NDFI revealed strong spatial variations of drought impacts in Central European lowlands, highlighting the value of Sentinel-2 for capturing detailed grassland dynamics during extreme meteorological events.
In this study, we extend the applicability of the NDFI for grassland drought monitoring across the larger environmental and land use gradients in Europe. We investigate twelve regions in typical land management settings thereby covering common climatic (e.g., Atlantic to continental climate) as well as topographic (e.g., lowland to alpine elevation) gradients. We base our analysis on the full Sentinel-2A/B archive from 2017 to 2021. To estimate fractional cover time series from Sentinel-2, we first build a global spectral library of pure PV, NPV, and soil spectra based on Land Use/Cover Area frame Survey (LUCAS) grassland samples. We verify the consistency of the library spectra by investigating NDVI vs. SWIR ratio across feature space distributions confirming temporally consistent spectral feature spaces for the test regions. Using the global spectral library in a regression-based unmixing approach, we derive fractional cover time series to calculate intra-annual NDFI time series from 2017 to 2021. First results for the grassland regions distributed along environmental as well as land use gradients in Europe underpin the potential for wall-to-wall monitoring of grasslands with the drought-sensitive NDFI on national to continental scales. Further research is required to assess the NDFI for global mapping of grassland drought, e.g., in semi-arid to arid grassland systems with different climate conditions and drought regimes.

References:
Kowalski, K., Okujeni, A., Brell, M., Hostert, P., 2022. Quantifying drought effects in Central European grasslands through regression-based unmixing of intra-annual Sentinel-2 time series. Remote Sensing of Environment 268, 112781. https://doi.org/10.1016/j.rse.2021.112781

Livestock are an important source of proteins (meat and milk) for large groups of pastoralists in Africa. Quantitative data, with high spatiotemporal resolution, on biomass and biomass changes are vital to support management of grazing resources in Africa. The RAMONA (RAngeland MONitoring for Africa using Earth Observation - Continental Demonstrator) project will produce several spatial resource related products for rangelands in Africa based on Earth Observation data from Sentinel 1, 2 and 3. Biomass and biomass anomalies, two of the products, are briefly described here.

Herbaceous rangeland biomass (a RAMONA core product) [t DM ha-1] is the net accumulation of the photosynthetic gain of carbohydrates (gross primary productivity, GPP, [g C m-2 day-1]) and losses through autotrophic respiration (Ra). Whereas GPP can be estimated by remote sensing, Ra can’t. In limited areas it may be possible to estimate grassland biomass productivity by linear models directly from GPP, or by using simple scalars to estimate net primary productivity (NPP) from GPP as in the current Copernicus Global Land Service Dry Matter Productivity product where dry matter productivity is assumed static and equal 0.5 x gross dry matter production. However, in Africa, there is in general insufficient ground data for calibration and validation. An alternate solution is to combine remote sensing and ecosystem modelling. Using an ecosystem model to convert GPP to NPP and biomass provides better constraints to the relationships across an extensive geographical area. The seasonal and spatial dynamic in the model is provided by GPP and originating from Sentinel based Earth observation data. GPP is estimated via a state-of-the-art implementation of the light use efficiency (LUE) methodology based on time series of fAPAR from fused Sentinel 2 and 3 data. The general algorithm quantifies GPP [g C m-2] as a product of the absorbed photosynthetic radiation (APAR=fAPAR x PARin, [MJ]) and the efficiency by which APAR is converted into carbohydrates, εmax [g C MJ-1] and environmental scalars that downregulates εmax based on resource limitation or environmental constrains:

GPP = εmax • fAPAR • PARin • scalars (1)

GPP and NPP will be calculated every five days as accumulated GPP – autotrophic respiration (Ra). Biomass dry weight is assumed to equal NPP/0.45 assuming a carbon content of 45%. εmax denotes the maximum rate of conversion of light energy and varies with photosynthetic pathway (C3/C4), may vary spatially and temporally. Candidate values for εmax will be derived from; 1) current eddy covariance flux measurements from African rangelands sites, 2) historical flux measurements from the FLUXNET2015 data base and additional African sites, 3) extracted from literature and 4) extracted from existing LUE models such as MOD17 and similar models based on Earth observation.

fAPAR can be accurately estimated based on satellite data and is a key variable for the methodology. Candidate variables for deriving fAPAR include several vegetation indices (NDVI, EVI, kNDVI) and the S2ToolBox Level 2 fAPAR product. Candidate variables will be temporally gap-filled using fused S1/S2/S3 data with additional noise removal by time series processing using the TIMESAT package (https://web.nateko.lu.se/timesat/timesat.asp). TIMESAT-smoothed data from the S1/S2/S3 fusion provides regular and less noisy (due to clouds aerosols etc) input for downstream processing. NDVI has been frequently used in to estimate fAPAR, and this relationship has been shown to be linear in the Sahel. However, across large areas the influence of several factors affects the relationship, particularly variations in soil background colour. Several studies have found that the EVI has better relationship with fAPAR than other vegetation indices. These properties relate strongly to the broader dynamic range and lower sensitivity to soil background of EVI, but there is also evidence that the efficiency of EVI is related to its response to fAPARchlorophyll while NDVI is more related to fAPARcanopy. These reasons help explaining why EVI has been successfully used for GPP estimation across a range of biomes, including grasslands, and has been applied in African drylands.

PARin, the incoming photosynthetic radiation quantifies the energy available for conversion to biomass. PARin is normally derived as a fraction (0.48) of incoming shortwave radiation and are depending on position, time, cloudiness and aerosols. Candidate variables to estimate PARin include ERA5-Land reanalysis data set (Surface net solar radiation) and the Daily Downward Surface Shortwave Flux provided by EUMETSAT.

Scalars, quantifying environmental downregulation of the LUE include, temperature and moisture availability. The scalars vary over space and time and within rangeland types and plant functional types. For down-regulation due to temperature we will mimic the MOD17 algorithm with further calibration from flux measurements. Candidate variables to estimate temperature include ERA-5 Land, 2m temperature.

Water availability can be expressed in several ways. Vapour pressure deficit (VPD) is commonly used to assess moisture stress and has been applied successfully in Africa even if the advantage of using soil moisture have been stressed. Candidate variables to estimate water availability include: VPD will be derived from temperature and relative humidity, available from ERA5-Land. Soil moisture available to plants can be obtained at 20 m spatial resolution through the modelling of soil evaporation and plant transpiration (evapotranspiration – ET) using optical and thermal data as successfully exemplified in ESA Sen-ET (https://www.esa-sen4et.org/) and ET4FAO (https://et4fao.dhigroup.com/#/) projects and as have been applied for quantifying soil moisture stress on crops. Alternatively, soil moisture estimates are available from ERA-5-Land. Evaporative fraction (LE/(LE+H), where LE = latent heat and H = sensible heat, has been reported strongly correlated to GPP and, similarly to soil moisture, could be obtained through own modelling or derived from ERA5-Land.

NPP will be derived from GPP based on the carbon used efficiency (CUE, NPP/GPP) derived from one or several state-of-the-art dynamic vegetation models, as CUE has been shown to vary substantially over time and space. This approach will provide a temporally and spatially distributed CUE per plant functional type in African rangelands.

Biomass [DM] is assumed to equal NPP/carbon content on monthly basis, where carbon content of vegetation is on average 0.45. Additional correction will be applied for the shoot:root ratio and biomass turnover.

Calibration of the LUE model to estimate herbaceous rangeland biomass will be performed using data collected at test sites distributed in African rangelands and additional available existing data. Candidate variables and their combinations will be evaluated at the test sites in order to find the most suitable and robust combination for continental upscaling of herbaceous rangeland biomass.

Herbaceous rangeland biomass is one of three core products (the other are Rangeland Extent and Rangeland Type) of the RAMONA projects. The core products will be produced at 10 m resolution for all rangelands in Africa. RAMONA also include three experimental products (Biomass Anomalies, Phenology and Carrying Capacity and three intermediate products (Fractional Woody Cover, SAR Backscatter and Surface Reflectance).

Herbaceous Rangeland Biomass Anomalies will be generated from standardized z-scores of fAPAR derived from medium-term Copernicus Sentinel-3/Proba-V at 300 m resolution as outlined by earlier studies. These data will supply a minimum of a 7-year baseline which is expected to be appropriate in the highly variable and rapidly changing drylands of Africa, at least for on-the-ground users more interested in comparing present conditions to the recent past than to conditions over past decades. If 7-year baselines turn out unsatisfactory it is possible to fall back on MODIS NPP data (MOD17A3hgfv061, 500 m resolution, going back 20 years). Harmonization of data to ensure consistency with the biomass originating from the fused S1/S2/S3 data will be performed in the pilot areas. The data will be seasonally and annually integrated and z-scores for each season and year computed per rangeland pixel. Harmonization between the Proba-V and Sentinel-3 time-series may be required to ensure consistency. The Herbaceous Biomass Anomalies is a experimental product.

Fire behaviour is well described by a fire’s direction, rate of spread, and its energy release rate. Fire intensity as defined by Byram (1959) is the most commonly used term describing fire behaviour in the wildfire community.
In frequently burning savannas and rangelands, fire intensity exerts a major influence on tree cover, tree growth, biodiversity, greenhouse gas emissions and the carbon cycle in general. Recent research indicates that fire intensity may play a major role in schemes to reduce emissions from savanna burning. However, fire intensity is difficult to observe from space or over large areas.
Here, we determine Byram's fire intensity for the first time from space by assessing fire rate of spread and fire radiative power using infrared sensors with different spatial, spectral and temporal resolutions. The sensors used offer either high spatial resolution (Sentinel-2) for fire detection, but a low temporal resolution, moderate spatial resolution and daily observations (Sentinel-3 and VIIRS), and high temporal resolution with low spatial resolution and fire radiative power retrievals (Meteosat SEVIRI). We extracted fire fronts from Sentinel-2 (using the short-wave infrared bands) and use the available fire products for Sentinel-3, S-NPP VIIRS and Meteosat SEVIRI. Rate of spread was analysed by measuring the displacement of fire fronts between the mid-morning Sentinel-2 overpasses and the early afternoon VIIRS overpasses. We retrieved FRP from 15-min Meteosat SEVIRI as well as Sentinel-3 and VIIRS observations and estimated total fire radiative energy release over the observed fire fronts. This was then converted to total fuel consumption, and, by making use of Sentinel- 2-derived burned area, to fuel consumption per unit area. Using rate of spread and fuel consumption per unit area, Byram’s fireline intensity could be derived. We tested this approach on fires in a frequently burning West African savanna landscape. Field experiments to determine fuel consumption, rate of spread and fire intensity were also carried out in the area and compared to models. Comparison to our own experiments and other studies in the region show similar numbers between field observations and remote-sensing-derived estimates. We also demonstrate differences between early and late season fires and discuss the implications that production of a larger satellite-derived dataset on fire intensity could have for mitigation projects striving to reduce emissions from savanna fires through changing current fire regimes, e.g. towards late season fires. Shortcomings of the presented approach and foundations of an error budget, and potential further development, also considering upcoming sensor systems, are discussed.

Livestock herders in Mali and Burkina Faso live under the twin threat of drought and armed conflict. Moving their herds to find pasture and water depends critically on access to reliable information on rangelands. This talk discusses a call center that uses Copernicus Earth Observation imagery and field data to provide herders with information on pasture, water and markets but with a focus on rangeland conditions. The talk will go over the architecture of the data treatment, demo the interface, talk about successes and failures and show how you can play with the data yourself.

Transhumance, or the seasonal movements of livestock herds to find pasture and water, is a centuries-old tradition in Mali and Burkina Faso. The process of selecting routes for movement hinges on a complex network of factors including customary access rights, pasture growth, rainfall, surface water, among others. However, years of climate change and armed conflict have made herding more precarious and prone to rapid changes. As a result, access to data on environmental and market conditions is critical for pastoralists. While satellite imagery has made much of this information readily accessible to the spatial community, few channels exist to transmit this information to herding communities. As a result, rangeland monitoring has become more powerful than ever before, yet mostly inaccessible to the pastoralists who depend on this information for their livelihoods.

In 2015, the GARBAL call center was built to provide this data to herders in Mali and Burkina Faso. The call center is powered by an open platform GIS built from Copernicus remote sensing data on vegetation and water and field data on market prices and animal conditions. Herders calling the center are connected to an agent who uses a dashboard to respond to their questions: Is pasture available near me? Is it crowded by other herds? Can I sell my goats for a good price? The call center’s goal is to provide herders with decision-making support in planning their routes.

The interface is built on mapserver and uses automated scripts to download and treat imagery from Sentinel 2 and Meteosat which then display information on pasture conditions and water availability. Field data is routed through a network of local data collectors who provide weekly updates on livestock conditions and market prices. In addition to an interactive map, the interface provides user-friendly textual outputs that summarize all the layers for any area of interest on the map, which allows call center agents to quickly provide data to callers.

This talk will share a number of the lessons learned from the STAMP project and provide a demonstration of the platform (which is openly accessible). Specific topics of discussion will include:
The architecture of the system- what worked and what didn’t
Maintaining regular field data collection in areas of ongoing active conflict
Building and translating GIS data for communities with low literacy
Examining the call records to see what data matters the most to users

You can use the platform at www.stamp-map.org. For more information, contact Alex Orenstein (alex@orensteingis.com / @oren_sa)