Sea surface temperature correction model based on spatiotemporal axial attention

Sea surface temperature (SST) is a core indicator of the ocean’s thermal state and plays a crucial role in fields such as ocean dynamics, climate prediction, and marine ecology. Currently, numerical models serve as the primary tool for SST forecasting; however, their outputs often exhibit systematic biases relative to observed values, necessitating effective post-correction. To address this issue, this study proposes a spatiotemporal correction model for SST forecasts based on an enhanced Transformer architecture. The proposed model explicitly captures the unique spatiotemporal dependencies inherent in SST data by introducing a learnable spatial positional encoding within the embedding layer, thereby enhancing spatial sensitivity. Subsequently, a spatiotemporal self-attention mechanism is designed for both the encoder and decoder, which sequentially extracts features along the latitude, longitude, and time dimensions. The feature extraction is performed independently along each dimension, ensuring consistency and effectively addressing the insensitivity of standard Transformers to spatial information when processing spatiotemporal sequences, while overcoming their inherent limitation in capturing correlations across time and space. Furthermore, the Laplace equation is incorporated as a physical constraint into the loss function, reinforcing the spatial smoothness and physical consistency of the corrected SST fields. Experimental evaluations conducted in the Sea of Japan demonstrate that the proposed model substantially reduces the forecasting error of numerical SST predictions for a seven-day horizon, yielding a 63.23% improvement in accuracy. Both quantitative and qualitative analyses demonstrate a substantial reduction in prediction errors after correction. Compared with uncorrected predictions, the root mean square error (RMSE) of the proposed method decreases from 1.181 ℃ to 0.434 ℃, significantly outperforming existing mainstream deep learning methods. Moreover, cross-region generalization tests show that the model maintains a monthly mean RMSE consistently in the range of 0.350–0.450 ℃ across different marine backgrounds, including the Bohai Sea, Yellow Sea, and South China Sea, without systematic failure or seasonal bias. These results fully validate the model’s effectiveness, stability, and technical advantages, demonstrating its great potential for operational marine forecasting.