A dynamic adaptive graph convolutional recurrent network model for efficient mid-short term prediction of global sea surface salinity

Accurate mid-short term prediction of sea surface salinity (SSS) is essential for operational ocean monitoring, particularly for capturing short-term salinity fluctuations that affect regional ocean dynamics and weather conditions. However, existing models struggle to extract complex spatiotemporal dependencies and are often limited to local regions, reducing their global applicability. To address these challenges, we propose a Dynamic Adaptive Graph Convolutional Recurrent Network (DAGCRN) for global SSS prediction. The DAGCRN employs an encoder–decoder architecture, where both the encoder and decoder integrate adaptive graph convolutional recurrent units (AGCRUs) and gated recurrent units (GRUs). AGCRUs dynamically construct topological relationships via graph convolution to model spatial variations, while GRUs capture temporal dependencies. This enables DAGCRN to effectively model the nonlinear and dynamic nature of global SSS variations. We evaluate DAGCRN's performance on the ESA Sea Surface Salinity CCI v3.21 dataset, which provides global gridded SSS observations from February 2010 to September 2020. Forecasting lead times range from 1 to 12 days. DAGCRN consistently outperforms LSTM, BiLSTM, ConvLSTM, and TCN. For 12-day prediction, RMSE is reduced by 36.0%, 24.4%, 13.0%, and 5.5%, respectively, demonstrating its effectiveness in modeling spatiotemporal dependencies for global SSS forecasting.