Reconstructing Subsurface Ocean Temperature From Sea Surface Multivariate Remote Sensing Data Using CSSP-ConvLSTM

The three-dimensional ocean temperature field plays a critical role in studies of global climate change, carbon cycling, and ocean dynamics. Subsurface temperature reconstruction, particularly within the thermocline, remains insufficient due to limitations in observational coverage and modeling frameworks. To address this challenge, we propose CSSP-ConvLSTM (CNN–Swin Transformer Spatial Parallel ConvLSTM), a subsurface temperature reconstruction model that integrates multi-source sea surface remote-sensing data. The model employs a spatial dual-branch structure and a temporal enhancement mechanism to achieve multi-scale characterization of local details, global dependencies, and nonlinear evolution. When applied to the northwestern Pacific (30–1000 m), reconstruction accuracy is significantly improved compared with single-branch or single-model baselines, with an average R² of 0.9811 and RMSE of 0.4557 °C reconstruction errors are markedly reduced within the thermocline (100–600 m) and deep layers (800–1000 m). Independent evaluations using ARMOR3D and EN4 datasets are conducted, confirming strong temporal generalization, stability, and physical consistency. Interpretability analysis based on SHAP indicates that shallow temperatures are primarily influenced by SST and SSH, whereas subsurface and deep layers are more affected by SSS and SSH. These results demonstrate that CSSP-ConvLSTM provides an effective approach for reconstructing high-resolution three-dimensional (3D) ocean temperature fields, and facilitates improved understanding of thermocline dynamics and their potential climatic implications.