Marine heatwave prediction method for the South China Sea

Abstract

Marine heat waves (MHWs) are prolonged anomalous sea temperature phenomena that affect marine ecosystems and their accurate prediction is of significance. In this study, a multivariate multi-granularity spatiotemporal graph neural network (MMS-GNN) based on a transformer is proposed. First, time series are sorted by period. The transformer is then used to capture the long-term dependency of various temporal granularities. Second, the entire marine region is divided into several subregions. The spatial dependency between different locations is then established by studying the changing trends between neighboring nodes. Third, a GNN is combined with a temporal convolutional network and employed to predict the collected information. Finally, experiments are conducted in the South China Sea using 40 years of observational data. Test results demonstrate that the proposed MMS-GNN can outperform other methods, meaning high effectiveness and flexibility of MMS-GNN for long-term MHW prediction.