The present work is developed within the MedEOS project, an application development project funded by the European Space Agency (ESA) as part of the Mediterranean Regional Initiative. Its main objective is to develop and produce daily high-resolution, gap-free water quality products based on Earth observation (EO) data for the whole Mediterranean coastline. This is achieved through combining the high temporal resolution of Sentinel-3 Ocean and Land Colour Imager (S3 OLCI) and the high spatial resolution of Sentinel-2 Multispectral Instrument (S2 MSI) in a process of data fusion.
Within MedEOS, five EO directly derived water quality products are to be developed: Total Suspended Matter, Turbidity, Chlorophyll-a Concentration, Secchi Depth and Colored Dissolved Organic Matter. In addition, EO indirectly derived water quality products shall also be produced, largely relying on combination of satellite and in situ data with numerical model results: Faecal Bacterial Contamination Indicators, Eutrophication Indicators, Harmful Algal Blooms, and Global Environmental Anomaly Detection. Finally, a river plume monitoring algorithm will provide a systematic detection of plumes related to major rivers discharging freshwater into the Mediterranean basin.
Given state-of-the-art approaches for spatiotemporal data fusion available in the litterature (see review in ), the proposed strategy for data fusion on EO derived water quality products is a combination of reconstruction-based and learning-based approaches.
The reconstruction-based STARFM (Spatial and Temporal Adaptive Reflectance Fusion Model)  is widely used in the litterature for spatiotemporal data fusion for land applications. Reconstruction-based approaches such as the STARFM relies on the hypothesis of reflectance similarity between both high and coarse-resolution sensors. If this hypothesis is often verified for land applications, it is less true on water. Lower values of water surface reflectances in comparison to land reflectances makes the reflectance similarity much more delicate to reach between two optical sensors. The higher sensitivity to atmospheric correction, higher impact of BRDF effects, heterogeneous sunglint effects and differences in spectral response for the same spectral band are all sources of discrepancies between water reflectance levels estimated from each sensor. It was therefore decided not to apply STARFM at the reflectance level but rather at the water quality products level to ensure a better consistency between products derived from both optical sensors.
Besides, STARFM is also strongly based on the assumption that objects do not drastically change between the date at which a pair of S2/S3 images exists (T0) and the date for which only a S3 image is available (T1). As it is most of the time not the case on highly dynamic ocean waters, an alternative approach is needed to enable the use of algorithms such as STARFM on water applications.
Our proposed strategy is to trick the STARFM algorithm by feeding it with an artificial pair of S2 and S3 data. In order to do so, a reference database made of small image patches is built from a S2/S3 water quality product database derived from a large time series of data. The S2/S3 pair is then created artificially through a coarse and fine resolution matching analysis between the reference database and the S3 water quality product observed on the targeted day. The matching analysis is first carried at the coarse resolution and further refined at the fine resolution. Each matching patch of coarse S3 pixels and its corresponding group of fine S2 pixels are further used to create an artificial S3 T0 product at a coarse resolution and an artificial S2 T0 product at fine resolution. Both the pair of S2/S3 artificial data at T0 and the S3 gap-filled product at T1 are given as input to the STARFM algorithm to produce the simulated product at S2 fine resolution.
The gap-filling part of the process is performed using the DINEOF algorithm . The database of water quality products derived from Sentinel-3 images are combined into multiple overlapping time series of approximately 2 months. Masks are extracted from each daily data to differentiate clouds from land pixels and perform the gap-filling only on desired pixels. The daily S3 products are also subsetted to fit the extent of a targeted S2 tile.
A first run of DINEOF is performed on the target time series. Using an approach proposed in , outlier pixels, mainly located near the edges of clouds, are detected and further removed from the dataset before applying the second round of DINEOF. The target day gap-free product is further extracted from the gap-free time series and used as input to the data fusion part of the processing chain.
An example of results of data fusion is given in Figure 1.
Next steps in the MedEOS project will be dedicated to the integration of all the different algorithms, developed by different service providers, in a single EO Exploitation Platform solution. This solution allows for the deployment of different components of the services supply chain in separated ICT cloud providers, making all service outputs available in a unique archive and catalogue module and visible in tailor made web-based geoportal. MedEOS implementation, therefore, follows the latest guidelines from the data science community, setting up the service processors where the input data is readily available, thus avoiding unnecessary download and transfer of data and increasing the performance of the services.
Production of water quality products and application of the proposed data fusion approach shall be performed initially for one complete year (2020) within five Pilot Areas in Egypt, Greece, France, Spain and Tunisia. Validation of these delivered products will be performed by comparing results with CMEMS products OCEANCOLOUR_MED_BGC_HR_L3_NRT_009_205  and with in-situ data collected by engaged users from each pilot. In a second phase, those products shall be derived and validated over the entire Mediterranean Sea coasts for a 3,5 year period, from March 2019 to September 2022.
For more information please visit the MedEOS website: https://medeos.deimos.pt/
 Belgiu, M., & Stein, A. (2019). Spatiotemporal image fusion in remote sensing. Remote sensing, 11(7), 818.
 Gao, F., Masek, J., Schwaller, M., & Hall, F. (2006). On the blending of the Landsat and MODIS surface reflectance: Predicting daily Landsat surface reflectance. IEEE Transactions on Geoscience and Remote sensing, 44(8), 2207-2218.
 Beckers, J. M., & Rixen, M. (2003). EOF calculations and data filling from incomplete oceanographic datasets. Journal of Atmospheric and oceanic technology, 20(12), 1839-1856.
 Alvera-Azcárate, A., Sirjacobs, D., Barth, A., & Beckers, J. M. (2012). Outlier detection in satellite data using spatial coherence. Remote Sensing of Environment, 119, 84-91.