A Novel Spatiotemporal Feature Decomposition Method and Fusion Strategy for Sea-Surface Density Prediction

Conventional sea-surface density (SSD) retrieval methodologies primarily depend on statistical analysis and numerical simulations driven by in situ measurements and satellite observations. These approaches exhibit inherent limitations in capturing the patterns of data variation, utilizing feature information, and coupling with oceanographic physical mechanisms, thereby constraining prediction accuracy. To address these challenges, we propose an integrated framework that synergistically combines satellite remote sensing with deep learning through a spatiotemporal feature decomposition method and fusion strategy. First, a variational mode decomposition (VMD)-based data processing method is developed to decouple the main feature into three components that align with the real-world SSD change patterns: seasonal trends, seasonal periods, and residuals. Subsequently, we construct a feature fusion strategy guided by SSD physical laws, utilizing three-dimensional convolution to hierarchically fuse multiphase features through the VMD constraint and feature polynomial fitting. The strategy further merges trend feedback based on auxiliary features to enhance or compensate for trend information, while leveraging the attention mechanism of the transformer to strengthen key spatiotemporal information and address long-term dependencies. Comprehensive experiments demonstrate the effectiveness of the proposed model, which integrates deep learning with oceanographic principles for enhanced SSD prediction. For the Bohai Sea area, compared with the state-of-the-art spatiotemporal models, the proposed model demonstrated a significant 16.76% decrease in the root mean square error (RMSE). Compared with the conventional inversion technique, the RMSE of the proposed model was dramatically reduced by 68.20%, further validating its superiority in improving prediction accuracy.