Algorithms for Atmospheric Correction (AC) and retrieval of water constituents from ocean colour satellite imagery have advanced considerably during the last decade but still they are subject to ongoing improvements. These improvements take place on many fronts, comprising the characterization of the atmosphere’s optical properties, i.e., aerosols, the characterization of the optical properties of the water body, i.e., scattering and absorption properties of optically active substances as well as the water itself, and the mathematical methods to invert the radiative transfer equation.

In the 1990’s, Schiller and Doerffer (1999) have introduced a methodology to resolve the problem of AC and in-water retrieval based on machine learning (ML) with artificial neural nets trained on large datasets derived by radiative transfer modelling. It was implemented in the Case2Regional processor, which later evolved into the C2RCC processor (Brockmann et al., 2016). While the ML aspects are most often mentioned when the C2RCC approach is discussed, the proper optical characterization of the water and the atmosphere is probably one of the most important parts. This consists of the parameterization of the radiative transfer models, and the specification of the optical model for the training dataset (co-variance of water constituents, their ranges and frequency distribution). Specific optical properties of algae assemblages, sediments, and coloured dissolved organic matter (CDOM) were used to derive special neural nets for extreme scattering or extreme absorbing waters (C2X nets) or for regional waters such as the Baltic Sea (Baltic+ nets).

The C2RCC processor is open source (https://github.com/senbox-org/s3tbx/tree/master/s3tbx-c2rcc) and publicly available through the SNAP toolbox (https://step.esa.int/main/toolboxes/snap/). The neural nets of C2RCC are included in the Sentinel-3 OLCI ground segment processor, but it is also capable of processing data from Sentinels–2 MSI, MERIS, VIIRS, MODIS, and Landsat-8 OLI.

In the past, the C2RCC development benefitted from community feedback and contributions by various scientists through collaborations in various international projects.

With the beginning of 2022 the water colour community will be entrusted (completely) with C2RCC as a community project. C2RCC offers several options to modify or further develop it. The elements will be made available to the community, so that C2RCC will hopefully evolve by the inputs provided by the community:
• optical properties of aerosols and atmospheric gases
• specific inherent optical properties of water constituents (SIOP, ranges)
• models that link optical with biogeochemical parameters
• optical water type classification (OWT)
• phytoplankton diversity
• validation data (link to in-situ databases such as Limnades, EUM OCDB, NASA SeaBASS)
• training datasets derived by RT modelling (well referenced to above IOPs)
• trained neural nets for
o atmosphere inversion (water leaving reflectance)
o path reflectance
o atmospheric transmittance
o auto-associated nets for out-of-scope testing
o water forward
o water inversion (IOP retrieval)
o kd spectral
o uncertainties of IOPs
• applicability to new sensors

Documentation for the parameters, the models used, and training of the nets will be provided. Webspace for the community to exchange their results and present their finding will be another part of the community project. Four main partners (Brockmann Consult Germany, Helmholtz-Zentrum Hereon, Brockmann Geomatics Sweden, and Ocean Obs Norway) will drive and maintain the project. However, it will be an unfunded activity whose success will depend on the uptake and contribution by the water colour science and application community. Of course, this platform may help the community to develop ideas for further innovative R&D work and projects.


References:
Brockmann, Carsten; Doerffer, Roland; Peters, Marco; Kerstin, Stelzer; Embacher, Sabine; Ruescas, Ana. Evolution of the C2RCC Neural Network for Sentinel 2 and 3 for the Retrieval of Ocean Colour Products in Normal and Extreme Optically Complex Waters. Living Planet Symposium, Proceedings of the conference held 9-13 May 2016 in Prague, Czech Republic. Edited by L. Ouwehand. ESA-SP Volume 740, ISBN: 978-92-9221-305-3, p.54

Helmut Schiller & Roland Doerffer (1999) Neural network for emulation of an inverse model operational derivation of Case II water properties from MERIS data, International Journal of Remote Sensing, 20:9, 1735-1746, DOI: 10.1080/014311699212443

Atmospheric correction (AC) is a critical step for retrieving spectral water reflectance — commonly referred to as water colour — from space-borne observations. It consists of removing the dominant and variable influence of the atmosphere and non-water contributions of the surface (e.g., Sun glint, white caps) from top of atmosphere (TOA) observations. Among many challenges, aerosol correction remains the most critical one due to the variability of particles in terms of types, their diverse horizontal and vertical distribution in the atmosphere, and their complex Sun light interactions.
As a legacy from developments for the past ENVISAT MERIS sensor, the current water reflectance Level 2 product from OLCI on Sentinel-3, processed at EUMETSAT, is obtained by applying an historical (also called “standard”) AC algorithm that essentially relies on the near-infrared (NIR) spectral range for determining the aerosol optical properties. The Spectral matching Atmospheric Correction for Sentinel Ocean colour measurements (SACSO) project has focused on developing a new AC prototype to improve the quality of the OLCI water reflectance Level 2 product. The SACSO strategy is based on an innovative approach, using a spectral matching technique, to address several issues not resolved with the standard approach, e.g., the high sensitivity to strong aerosol contamination, in particular absorbing ones, and thin or sub-pixel clouds, which cause reduced accuracy and spatial coverage, negatively impacting the usefulness of the products.
The SACSO team has built upon the POLYMER AC algorithm as a baseline for developing this alternative processor. The scheme still relies on iterative spectral optimization using the full solar spectrum, but the atmospheric reflectance model and the numerical optimization have been deeply reworked, leading to improved results and better algorithmic stability compared to POLYMER. The results of extensive validation studies are presented, including improvements with respect to the current AC algorithm.

We present the latest evolutions of the Dark Spectrum Fitting (DSF) atmospheric correction algorithm as integrated in the free and open source ACOLITE processor (Vanhellemont and Ruddick 2018, 2021, Vanhellemont 2019a, 2019b, 2020). ACOLITE/DSF can process imagery from dozens of optical satellite sensors from CubeSats (Planet Doves) to experimental hyperspectral missions (e.g. PRISMA/DESIS). The main purpose of ACOLITE is to enable aquatic applications in coastal and inland waters from a variety of satellite missions. The source code and binary versions for macOs, Windows, and Linux are freely available from https://github.com/acolite/acolite

The DSF assumes a given satellite scene or subscene contains targets for which the surface contribution is negligible in one of the sensor bands, and can hence be used to estimate the atmospheric path reflectance and aerosol optical depth (AOD). Unlike classical ocean colour atmospheric correction algorithms, the DSF has no a priori assumption of which bands to use to determine the AOD, and dynamically selects the band that gives the lowest AOD. As the DSF does not rely on NIR/SWIR bands, and can even use land targets for the retrieval of AOD, it can give reasonable estimates of the AOD even in difficult conditions, such as for water targets that have strong influence of sun glint or land adjacency. Experimental methods to correct for these effects after the atmospheric correction are included.

This presentation will focus on the processing of Copernicus Sentinel-2/MSI and Sentinel-3/OLCI data, and matchup performance using in situ measurements. PANTHYR (Vansteenwegen et al. 2019) and HYPSTAR autonomous hyperspectral radiometers (AHR) were installed at various inland and coastal sites. AHR data allow for the validation of any satellite spectral band within the spectral range of the instrument, and for this work the AHR data were resampled to the relative spectral response of the Sentinel-2 and -3 sensors (MSI-A, MSI-B, OLCI-A, OLCI-B). Satellite and AHR data were matched up, and the performance evaluated across a number sites and different targets. The resulting matchups allow for the evaluation of different ACOLITE/DSF configurations, and to recommend good general settings for ACOLITE/DSF processing of Copernicus optical satellite data.

References

Vanhellemont and Ruddick 2018, Atmospheric correction of metre-scale optical satellite data for inland and coastal water applications https://doi.org/10.1016/j.rse.2018.07.015

Vanhellemont 2019a, Adaptation of the dark spectrum fitting atmospheric correction for aquatic applications of the Landsat and Sentinel-2 archives https://doi.org/10.1016/j.rse.2019.03.010

Vanhellemont 2019b, Daily metre-scale mapping of water turbidity using CubeSat imagery. https://doi.org/10.1364/OE.27.0A1372

Vanhellemont 2020, Sensitivity analysis of the dark spectrum fitting atmospheric correction for metre- and decametre-scale satellite imagery using autonomous hyperspectral radiometry https://doi.org/10.1364/OE.397456

Vanhellemont and Ruddick 2021, Atmospheric correction of Sentinel-3/OLCI data for mapping of suspended particulate matter and chlorophyll-a concentration in Belgian turbid coastal waters https://doi.org/10.1016/j.rse.2021.112284

Vansteenwegen, D., Ruddick, K., Cattrijsse, A., Vanhellemont, Q., Beck, M., 2019. The pan-and-tilt hyperspectral radiometer system (PANTHYR) for autonomous satellite validation measurements—prototype design and testing https://doi.org/10.3390/rs11111360

This talk presents a novel technique to improve visualisation and interpretation of harmful algal bloom (HAB) risk by combining Lagrangian trajectories and satellite observations. The algorithm can predict changes not only to chlorophyll-a concentration but also to spectral properties of the water, which are important for discrimination of different algal species from satellite ocean colour. The hydrodynamics are predicted by an operational setup of the Finite Volume Community Ocean Model (FVCOM) on an unstructured grid to allow higher resolution around complex coastlines; and the ocean colour data (300m Sentinel-3 OLCI) are advected hourly by the particle tracking model PyLAG. The prediction scheme was applied to regions along the coast of England to verify its applicability. It was shown that the Lagrangian methodology can significantly improve the coverage of satellite products, and the unique enhanced colour animations are effective for interpretation of the development of HABs over the next few days (example animation: https://rsg.pml.ac.uk/shared_files/junl/paper_animation/example2/rgb.gif ).

A comparison between chlorophyll-a predictions and satellite observations further demonstrated the effectiveness of this approach: r^2 = 0.81 and a low mean absolute percentage error of 36.9%. Although uncertainties from modelling and the methodology affect the accuracy of predictions, this approach offers a powerful tool for monitoring the marine ecosystem and for supporting the aquaculture industry with improved early warning of potential HABs. We have engaged with shellfish and finfish farms through Interreg Atlantic Area project PRIMROSE and BBSRC project “Safe and Sustainable Shellfish”.

Since the recent publication of the approach (Lin et al., 2021), we have been developing animated short-term predictions of the risk for certain high-biomass HAB species based on their multispectral characteristics (based on Kurekin et al., 2015). This will enable targeted early warnings to finfish farms that can be serious impacted by bloom species such as the dinoflagellate Karenia mikimotoi.

Lin, J., Miller, P.I., Jönsson, B.F. & Bedington, M. (2021) Early warning of harmful algal bloom risk using satellite ocean color and Lagrangian particle trajectories. Frontiers in Marine Science, 8(1642). https://dx.doi.org/10.3389/fmars.2021.736262

Kurekin, A.A., Miller, P.I. & Van der Woerd, H.J. (2014) Satellite discrimination of Karenia mikimotoi and Phaeocystis harmful algal blooms in European coastal waters. Harmful Algae, 31, 163-176. http://dx.doi.org/10.1016/j.hal.2013.11.003

We contribute to this Session with a description of our “QAA-RGB”, recently published in Remote Sensing of Environment. The QAA-RGB spectral retrieves absorption, backscattering, and the diffuse attenuation coefficient (Kd), as well as the Secchi disk depth (zSD), from the remote-sensing reflectance (Rrs) at only three bands (RGB). It is a minimal version of the Quasi-Analytical Algorithm (QAA), therefore keeping its robustness and general validity across different water types. It works with eighteen metre and decametre satellite sensors, including all the present and heritage Landsats, Sentinel 2 at 10 m and an array of commercial satellites such as PlanetScope, Pléiades and Worldview. Careful calibration for these sensors, having relatively wide bandwidths, required to take into consideration the full sensor relative spectral responses and not only the center wavelengths. The result is a multi-sensor algorithm that provides unbiased retrievals across all sensors, in view of future multi-sensor time series generation. In terms of statistics, retrievals of non-water absorption at 555 nm (anw(555)) are satisfactory up to anw(555) < 2 m−1, with a standard deviation of σ ~ 50%, based on in situ datasets. Validation to independent in situ data is also provided. To ease the uptake by a larger community of users, the QAA-RGB has been implemented in the ACOLITE generic processor, and is openly accessible. Consistency of near-coincident observations from different satellite sensors is demonstrated. Future activities will include the extension to very highly absorbing waters by the inclusion of a near-infrared band, which will likely extend the applicability range by an order of magnitude. That will require data collection in some extreme environments such as turbid estuaries and highly CDOM-laden lakes. The Secchi disk model for these waters is likely to be reviewed as well.