Seamless retrievals of chlorophyll-a from Sentinel-2 (MSI) and Sentinel-3 (OLCI) in inland and coastal waters: A machine-learning approach

Consistent, cross-mission retrievals of near-surface concentration of chlorophyll-a (Chl a ) in various aquatic ecosystems with broad ranges of trophic levels have long been a complex undertaking. Here, we introduce a machine-learning model, the Mixture Density Network (MDN), that largely outperforms existing algorithms when applied across different bio-optical regimes in inland and coastal waters. The model is trained and validated using a sizeable database of co-located Chl a measurements (n = 2943) and in situ hyperspectral radiometric data resampled to simulate the Multispectral Instrument (MSI) and the Ocean and Land Color Imager (OLCI) onboard Sentinel-2A/B and Sentinel-3A/B, respectively. Our performance evaluations of the model, via two-thirds of the in situ dataset with Chl a ranging from 0.2 to 1209 mg/m 3 and a mean Chl a of 21.7 mg/m 3 , suggest significant improvements in Chl a retrievals. For both MSI and OLCI, the mean absolute logarithmic error (MAE) and logarithmic bias (Bias) across the entire range reduced by 40–60%, whereas the root mean squared logarithmic error (RMSLE) and the median absolute percentage error (MAPE) improved two-to-three times over those from the state-of-the-art algorithms. Using independent Chl a matchups (n < 800) for Sentinel-2A/B and -3A, we show that the MDN model provides most accurate products from recorded images processed via three different atmospheric correction processors, namely the SeaWiFS Data Analysis System (SeaDAS), POLYMER, and ACOLITE, though the model is found to be sensitive to uncertainties in remote-sensing reflectance products. This manuscript serves as a preliminary study on a machine-learning algorithm with potential utility in seamless construction of Chl a data records in inland and coastal waters, i.e., harmonized, comparable products via a single algorithm for MSI and OLCI data processing. The model performance is anticipated to enhance by improving the global representativeness of the training data as well as simultaneous retrievals of multiple optically active components of the water column.