Abstract: SST (sea surface temperature) is a critical physical parameter in ocean-atmosphere interactions. Accurate prediction of SST is of great significance for disaster prevention and mitigation, and marine ecological conservation. This study integrates multi-source ocean environmental parameters—including SST, DOS (density of sea surface), NO₃ (mole concentration of nitrate), and SWH (significant wave height)—based on the U-Net architecture to construct a sea surface temperature prediction model (SP-UNet). By comparing different parameter combinations, an optimal configuration scheme is derived. Results indicate the following: Incorporating SWH significantly enhances prediction accuracy in both the East China Sea and South China Sea regions, with particularly notable improvements in the East China Sea. The MAE (mean absolute error) decreased by 53% compared to using SST input alone. The predictive performance of DOS and NO₃ varies by sea area: introducing DOS yields superior results in the South China Sea, while incorporating NO₃ performs better in the East China Sea. Concurrently, model errors fluctuate seasonally, with the lowest errors occurring in the South China Sea during spring and summer, and in the East China Sea during autumn and winter. Furthermore, comparative results between SP-UNet and ConvLSTM models demonstrate superior predictive performance, validating the model's effectiveness in marine temperature forecasting tasks. This study provides an effective method for regional ocean temperature forecasting, holding significant implications for marine meteorological early warning and climate change research.

Key words: sea surface temperature prediction, multi-source data, Chinese coastal waters, deep learning, U-Net