Investigation of Physics-Informed Methods for Improving Sea Surface Height Prediction Based on Neural Networks in the South China Sea

Sea surface height (SSH) derived from satellite altimetry is essential for oceanographic research and marine monitoring. Although artificial intelligence (AI) models show considerable potential in forecasting, their application in oceanography remains constrained by several limitations. To address these challenges, we propose a set of physics-informed methods to improve SSH prediction based on neural networks in the South China Sea (SCS). The key strategies include: (1) incorporating land mask information to mitigate artifacts induced by the presence of land in marine data; (2) introducing a geostrophic constraint into the loss function; and applying latitude-dependent weighting to this constraint to account for the breakdown of geostrophic balance near the equator. On the test dataset, the physics-informed SimVPv2 (Phys-SV) model achieves an RMSE of 0.0173 m, a 13% improvement over the baseline SimVPv2 (Base-SV). The PredRNNv2 (PR) model also benefits significantly from the inclusion of land mask input, with RMSE reduced by 12% (from 0.0280 m to 0.0246 m). To the best of our knowledge, this study is the first to identify the artifact issue in AI models caused by land points in ocean data and to reveal the limitations of directly concatenating heterogeneous oceanic variables as model inputs.