Impact of physical and attention mechanisms on U-Net for SST forecasting

Abstract

Sea surface temperature (SST) prediction is a subject of great significance to the marine environment and human society. Changes in SST not only impact marine ecosystems and fishery resources but also trigger extreme weather events and disastrous consequences. Therefore, the precise prediction of SST is essential to avoiding these problems. Although numerous data-driven SST prediction models have emerged in recent years, these models are characterized by a lack of physical mechanisms related to sea temperature changes as well as insufficient generalization capabilities and interpretability. In our work, attempts were made to integrate physics-related convection phenomena into deep learning models, and traditional deep learning models were improved by incorporating time and space attention modules. The results of a series of experiments showed that the incorporation of physical mechanisms enhanced the performance of data-driven models. Furthermore, attention mechanisms were similarly helpful, of which temporal attention proved to be more important. The modules proposed in this work also improved the baseline model’s accuracy by 22%. In addition, seven-day SST predictions were carried out for the world’s five major fishing grounds. The results demonstrated that the application of transfer learning strategies yielded superior performance, further improving prediction accuracy by 1%–5%.