Physics-enhanced deep learning for sea surface temperature forecasting via multi-scale feature integration

Accurate prediction of sea surface temperature (SST) is essential for marine environmental monitoring and climate forecasting. However, most existing deep-learning-based approaches rely heavily on data-driven methodologies and lack sufficient integration of physical mechanisms, thereby limiting their physical consistency and interpretability. To overcome this limitation, this study introduces a multi-source coupled prediction neural network (MSCPNN), which incorporates temperature, salinity, and current dynamics into a multi-scale feature learning framework. Built upon the multi-feature physical neural network (MFPNN), the proposed model integrates a convolutional block attention module (CBAM), where channel attention adaptively models the multi-factor coupling among temperature, salinity, and currents, and spatial attention captures multi-scale spatial patterns in SST. Using high-resolution reanalysis data from the South China Sea spanning 2011 to 2020, comprehensive experiments were conducted comparing MSCPNN with MFPNN and PCL-MFPNN. The results demonstrate that MSCPNN significantly outperforms the baseline models across multiple evaluation metrics—including RMSE, correlation coefficient, PSNR, and SSIM—achieving an average reduction in RMSE of 17% and an increase in correlation coefficient of 0.035, which reflects higher predictive accuracy and improved physical consistency. Ablation studies further validate the superiority of multi-factor coupling over single-factor alternatives and clarify the distinct contributions of salinity and currents to SST prediction. Overall, MSCPNN advances the accuracy and stability of long-term SST forecasting while providing a more interpretable framework for intelligent ocean prediction.