Global estimation of phytoplankton pigment concentrations from satellite data using a deep-learning-based model

Based on a global matchup between satellite observations and high performance liquid chromatography (HPLC) measurements, we developed a deep-learning-based model (DL-PPCE model) for globally estimating concentrations of 17 different phytoplankton pigments. The model adopted a fusion architecture of residual and pyramid networks to achieve robust estimation performance. The model inputs include three different data types: essential ocean color parameters, satellite-derived environmental parameters, and the slope of above-surface remote-sensing reflectance (Rrs). We compared the model performances with various input parameters to determine the most effective inputs. The results showed that Rrs in the essential ocean color parameters and sea surface temperature (SST) in the environmental parameters were the most critical input parameters. The estimation of phytoplankton pigment concentrations was validated against HPLC data using the leave-one-out cross-validation method. Except for three pigments, 19′-butanoyloxy-fucoxanthin, prasinoxanthin, and lutein, the estimated pigment concentrations and in-situ observations were strongly correlated for all other pigments (an average relative root-mean-square error of 0.59, R2≥0.60, and regression slopes close to 1). In addition, a time series analysis was performed on the MODIS retrieved global pigment concentrations during 2003–2021 using the established DL-PPCE model to explore the relationship between the distribution of phytoplankton groups and El Niño in the western equatorial Pacific. Our findings revealed that the prokaryotes-dominated area extended eastward from180°E to 150°W during the 2015/2016 El Niño event. From 2003 to 2021, prokaryotic abundance was positively correlated with El Niño intensity (R=0.65,P≪0.01) but negatively correlated with the abundance of the entire phytoplankton community (R=−0.53,P≪0.01). These results demonstrate that the DL-PPCE model presents a novel approach for estimating the concentration of 17 pigments worldwide, and the estimated pigment concentrations are advantageous for analyzing the phytoplankton community dynamics on a large spatiotemporal scale.