Interpreting optical-biogeochemical associations underlying satellite nitrate retrieval in coastal waters using Explainable AI

Sea surface nitrate (SSN) monitoring is essential for understanding coastal eutrophication and anthropogenic water cycle. However, SSN is optically inactive, and the mechanisms that enable its retrieval from satellite observations remains unclear, limiting the reliability of large-scale and early diagnosis of coastal eutrophication from space. To address this gap, we employ an explainable artificial intelligence (XAI) model to identify the latent associations between SSN and individual satellite spectral bands. By estimating daily SSN distributions in the East China Sea, the Pearl River Estuary, and Chesapeake Bay, the XAI model explicitly quantifies the spectral contributions and reveals the optical-biogeochemical association underlying SSN retrieval. According to optical complexity, the coastal waters are classified as low-turbidity water and high-turbidity water. In low-turbidity water, XAI reveals a distinct spectral fingerprint characterized by positive contributions from the red band near 673 nm and negative contributions from the violet band near 400 nm. Further analysis demonstrates that the red-band contribution reflects the biological nitrate uptake by phytoplankton, whereas the negative violet-band contribution is associated with shortwave absorption by CDOM and detrital material that co-vary with nitrate transport and enrichment. In high-turbidity water, strong particle scattering obscures the spectral fingerprint, indicating that satellite-derived SSN estimates should be interpreted with greater caution under highly turbid conditions. These findings highlight the optical-biogeochemical associations underlying satellite nitrate retrieval and provide both an explainable basis and practical guidance for future monitoring of coastal nutrient dynamics from space.