Sea Surface Temperature Prediction Based on Adaptive Coordinate-Attention Transformer

Sea surface temperature (SST) serves as a critical indicator of oceanic thermodynamic processes and climate variability, exerting essential influence on ocean fronts, typhoon tracks, and monsoon evolution. Nevertheless, owing to the highly nonlinear and complex multi-scale characteristics of SST, achieving accurate spatiotemporal forecasting remains a formidable challenge. To address this issue, we proposed an enhanced Transformer architecture that incorporates a Coordinate Attention (CA) module and an Adaptive Fusion (AD) module, enabling the joint extraction and integration of temporal and spatial features. The proposed model is evaluated through SST prediction experiments over a localized region of the South China Sea with lead times of 1, 7, 15, and 30 days. Results indicate that our approach consistently outperforms baseline models across multiple evaluation metrics. Moreover, generalization experiments conducted on datasets from regions with diverse latitudes and climate regimes further demonstrate the model’s robustness and adaptability in terms of both accuracy and stability.