Ocean temperature reconstruction from satellite observations in the North Atlantic using an explainable deep learning framework

In this study, we reconstructed the North Atlantic Ocean temperature field by combining an interpretable deep learning framework with multi-source satellite data, including sea surface temperature (SST), sea surface height (SSH), sea level anomaly (SLA), sea surface wind field (SSW), and SST gradient (GRAD). The optimal model configuration achieved a root mean square error (RMSE) of 0.49 °C. This study extends the SHapley Additive exPlanations (SHAP) method to spatial, temporal, and vertical dimensions to systematically analyze the model’s error sources and feature contributions. The results show that the reconstruction error is concentrated in regions with large physical gradients, such as the thermocline and Gulf Stream, exhibiting a significant seasonal bias pattern. SHAP analysis reveals a clear vertical transition in feature importance that SST dominates at all depths with a relative contribution of 34.60%. The contributions of SSW and GRAD were mainly concentrated at the sea surface, whereas the contributions of SSH and SLA began to increase in the lower mixed layer and peaked in the thermocline, becoming crucial in connecting the surface and ocean interior. This study highlights the depth and spatial dependence of feature contributions, providing a physical basis for the future optimization of remote sensing-based ocean temperature reconstruction models.