Coastal nutrient retrieval using multi-dimensional synergistic MSI/OLCI data with a group-specific feature fusion neural network

Current satellite remote sensing methods for coastal nutrient monitoring face dual challenges of low accuracy and poor interpretability due to the non-optically sensitive nature of dissolved inorganic nitrogen (DIN) and soluble reactive phosphorus (SRP); furthermore, single-sensor approaches cannot fully exploit multi-dimensional spectral information. To address these challenges, this study develops a group-specific feature fusion neural network (GSFNN) integrating Sentinel-2 Multispectral Instrument (MSI) and Sentinel-3 Ocean and Land Colour Instrument (OLCI) data to achieve accurate and interpretable coastal nutrient retrieval. The framework employs: (1) co-attention inverse distance weighting for spatial enhancement, (2) spectral response function-based alignment for multi-sensor integration, and (3) group-specific random feature attention mechanism for dynamic feature fusion and enhanced interpretability. Validation in the Bohai Sea demonstrates superior performance with R2 values of 0.74 for DIN and 0.65 for SRP. Application to Lushunkou coastal waters reveals three-stage nutrient regulation patterns driven by algal growth cycles and quantifies spatiotemporal nutrient-phytoplankton coupling mechanisms, and identifies positive correlations between wind speed and nutrients as well as negative precipitation–nutrient relationships via dilution. This study provides the first interpretable multi-sensor neural network framework for coastal nutrient monitoring, offering robust technical support for marine environmental management and eutrophication assessment.