The accurate prediction of sea surface temperature (SST) is essential for studying marine phenomena, understanding climate dynamics, and forecasting environmental changes. However, developing a general SST prediction model is challenging due to significant regional variations and the impacts of diverse climate phenomena. To improve the performance of SST predictions, we propose a hybrid framework that effectively models the spatial and temporal dependencies of SST data with a Gaussian process-enhanced Long Short-Term Memory network. The LSTM module adaptively captures both long and short-term temporal trends in SST variation, while the Gaussian process incorporates the spatial dependency of neighboring data to further refine the predictions. Furthermore, our proposed framework estimates the uncertainty associated with SST predictions, providing crucial information for practical applications. Comprehensive experiments are conducted on the OISST dataset, with a focus on the Bohai Sea and the South China Sea. The results of our framework outperform state-of-the-art methods, validating its superiority in SST prediction.

Accurate prediction of Sea Surface Temperature (SST) is critical for understanding marine phenomena, studying climate dynamics, and forecasting environmental changes. However, SST time series exhibit varying trends across different regions due to differences in latitude, ocean currents, and depth, which complicates precise forecasting. To address this, our proposed framework, the Dynamic Scale Interaction Network (DSIN), dynamically adapts to these regional variations by selecting appropriate temporal scales in the frequency domain. This process is facilitated by a dynamic routing module that does not require the predefinition of scale numbers, enabling our model to generalize across diverse geographic areas with varying SST trends. Additionally, SST is significantly influenced by local climate phenomena, such as storms, making the modeling process more complex. DSIN addresses this challenge by incorporating multivariate interactions and leveraging a cross-attention mechanism to capture the correlations between SST and exogenous variables, such as storm data. This enhances the model’s ability to accurately forecast SST even under the influence of localized climate events. Moreover, to enhance the model’s adaptability to complex spatiotemporal dynamics and improve its robustness across different SST prediction tasks, the framework employs a Graph Neural Network (GNN) to capture intricate correlations. By incorporating a probabilistic loss function, we estimate the uncertainty of SST predictions, thereby providing valuable information for more informed decision-making in climate modeling, marine navigation, and fisheries management. Extensive experiments conducted on the OISST SST dataset demonstrate that DSIN consistently outperforms state-of-the-art methods, proving its superiority in SST prediction.