Study on Remote Sensing Inversion of Subsurface Chlorophyll in the South China Sea Based on Convolutional Neural Networks

Understanding the vertical structure of chlorophylla (Chl-a) is essential for assessing ocean primary productivity, biogeochemical cycles, and carbon fluxes, yet this information cannot be directly obtained from satellite observations. In this study, we developed a novel convolutional neural network-based model (CNN-SCMs) to estimate the vertical distribution of subsurface Chl-a concentration from satellite-derived surface parameters. CMEMS Chl-a model data and NOAA OISST data from the South China Sea (SCS) were fused to construct the training dataset. Using only sea surface Chl-a concentration, sea surface temperature, and month as inputs, CNN-SCMs predicts Chl-a profiles within the $0-150 \mathrm{~m}$ depth range. The model achieved an RMSE of $0.11 \mathrm{mg} / \mathrm{m}^{3}$ and a correlation coefficient of 0.84 on the test dataset. Temporal inversion results at the SEATS station and vertical distributions along the $116^{\circ} \mathrm{E}$ and $18^{\circ} \mathrm{N}$ transects further confirmed the model’s effectiveness and accuracy.