Global Spatiotemporal Graph Attention Network for Sea Surface Temperature Prediction

Accurately predicting sea surface temperature (SST) plays an important role in the study of marine ecosystems and global climate. The SST prediction problem is usually formulated as a time-series regression problem; i.e., the future SST is predicted based on the historical SST. However, the existing methods are typically devoted to modeling the highly nonlinear temporal correlations in SST data. They often ignore the dynamic spatial correlations that exist. This can limit the performance of these models, making accurately predicting SST challenging. To address this challenge, by combining graph neural networks (GNNs) that have a clear advantage in modeling spatial correlations, we propose a global spatiotemporal graph attention network (GSTGAT). Specifically, we capture the global dynamic spatial correlations of nodes through a global graph attention network (GGAT) module that fuses the static adjacency matrix learned adaptively by the graph learning (GL) module with the dynamic attention coefficients. A gated temporal convolutional network (GTCN) module is used to capture the nonlinear temporal correlations. Then, the above modules are integrated into a unified neural network to predict SST. We conduct experiments on multiple time-scale datasets in the Bohai Sea and the South China Sea. The experimental results show that GSTGAT achieves the best performance and consistently outperforms other methods for different prediction horizons in the two sea areas.

Automated summary

Accurately predicting sea surface temperature is crucial for studying marine ecosystems and global climate. To address the limitations of existing methods in modeling dynamic spatial correlations, we propose a global spatiotemporal graph attention network (GSTGAT) that combines graph neural networks (GNNs). Our network captures global dynamic spatial correlations through a global graph attention network module and utilizes a gated temporal convolutional network module to capture nonlinear temporal correlations. Experimental results on datasets from the Bohai Sea and the South China Sea demonstrate the superior performance of GSTGAT compared to other methods for different prediction horizons.