Enhancing sea surface salinity short-term prediction using physically informed deep learning

Abstract

Sea surface salinity (SSS) is a crucial variable in understanding ocean dynamics, climate change, and marine ecosystems. However, purely data-driven deep learning models often lack integration with physical laws, which limits their interpretability and prediction accuracy. To address this issue, this study develops a physically informed deep learning framework that incorporates spatiotemporal patterns from Empirical Orthogonal Function (EOF) analysis of SSS into deep learning architectures. Two distinct methods are proposed: one integrates EOF-based constraints into the loss function (LEOF), and the other employs a Siamese network branch to explicitly embed physical constraints within the model (SEOF). The models utilize multi-source satellite data and various factors (such as sea surface temperature) to produce daily SSS prediction. The results demonstrate that the inclusion of EOF constraints significantly improves predictive accuracy, reducing Root Mean Square Error and Mean Absolute Error by up to 25.51% and 53.63%, respectively. The SEOF-RAUnet model, which combines EOF-based constraints with an attention mechanism, achieves the best performance, particularly in high-variability regions such as the eastern Pacific under the influence of the Equatorial Countercurrent. The proposed framework establishes a basis for forthcoming research focused on the integration of physical constraints with AI-driven SSS prediction, thereby enhancing applications in marine environment monitoring and prediction.

The authors' abstract, as published at the source. Applied Ocean Research, 2025 · DOI ↗