Decoupled Dynamic Spatial–Temporal Graph Neural Network for Sea Surface Temperature Prediction

Accurate sea surface temperature (SST) prediction is essential for advancing the understanding of marine ecosystems and global climate dynamics. SST variations arise from a dynamic interplay of multiple factors. Specifically, we model changes in SST as a combination of diffusion processes, representing the spread of thermal energy across the ocean, and inherent signals that capture localized and intrinsic patterns of temperature variation. This perspective recognizes that SST is influenced not only by diffusion but also by factors such as ocean currents, seasonal cycles, and localized climatic events. Thus, we propose DDSTGNN, a novel model designed to decouple diffusion and inherent SST signals using data-driven methods. The model incorporates an estimation gate and a residual decomposition mechanism to effectively achieve this separation. Additionally, it integrates a dynamic graph learning module to model the evolving spatial and temporal dependencies of SST networks. Extensive experiments on two real-world SST datasets demonstrate that DDSTGNN outperforms state-of-the-art methods, highlighting its superior ability to model spatial-temporal SST data.