A Spatio-Temporal Information Fusion Deep Learning Model for Daily and Monthly Sea Surface Temperature Prediction Based on ConvGRU

Deep learning has advanced sea surface temperature (SST) prediction. However, existing models encounter difficulties in extracting multi-scale spatio-temporal information and fully leveraging spatial position details. As a result, models fail to adequately comprehend the spatial correlations among grid points in SST data and struggle to capture the complex interactions between different geographical positions, which leads to low prediction accuracy. To address these issues, this study proposes the MSIFM-SMAM-ConvGRU model based on the Convolutional Gated Recurrent Unit (ConvGRU) to enhance SST prediction accuracy. The Multi-scale Spatial-temporal Information Fusion Module (MSIFM) extracts multi-scale spatio-temporal information. The Spatial-aware Multi-dimensional Attention Module (SMAM) captures complex SST patterns and inter-grid correlations from multi-dimensions. The Temporal Feature Extraction Module (TFEM) strengthens the extraction of temporal features and outputs prediction results. The model is trained and tested using SST data from the South China Sea. The experimental results indicate that the proposed model outperforms the comparison model. Specifically, for the next 1-30 days, the Root Mean Square Error (RMSE) ranges from 0.128 to 0.638°C and the Mean Absolute Error (MAE) from 0.086 to 0.441°C; for the next 1-12 months, the RMSE ranges from 0.464 to 0.530°C and the MAE from 0.331 to 0.381°C. Moreover, the spatial distribution of errors and scatter plots demonstrate the model's reliability and robustness. Ablation studies verify the effectiveness of each proposed module in enhancing model performance. The one-sided paired t-tests further confirm the statistical significance of the prediction results compared to comparison methods. Cross-regional tests exhibit the model's excellent generalization capability and adaptability.