Vision Transformer applied for Time Series based Prediction of Sea Surface Temperature

Sea surface temperature (SST) is a critical variable for ocean monitoring and ecosystem protection, particularly in the context of climate change. Although LSTM-based models have been widely adopted for SST prediction, they often struggle to capture localized spatial patterns effectively. To address this, we propose TS-ViT-SST, a Token Sparsification-based Vision Transformer that selectively attends to informative regions within SST data, thereby enhancing the modeling of fine-grained spatial variations. Furthermore, we introduce a physics-inspired temperature difference loss to suppress unrealistic fluctuations and improve prediction accuracy. Model performance is evaluated primarily using RMSE, with additional comparisons of parameter counts, computation times, AIC, BIC, and other relevant metrics across different models. Experiments on the SST-RSS dataset demonstrate that TS-ViT-SST surpasses state-of-the-art (SOTA) methods, achieving a 0.11 reduction in RMSE while requiring fewer computational resources. Additional validation on the ICAR-ENSO dataset for El Niño–Southern Oscillation (ENSO) prediction confirms the robustness and generalization capability of our approach. By integrating sparsified attention mechanisms with physical constraints, TS-ViT-SST establishes a new benchmark for SST forecasting. The code is available at: https://github.com/zhzhao2020/TS-ViT-SST.