An Informer-based prediction model for extensive spatiotemporal prediction of sea surface temperature and marine heatwave in Bohai Sea

Abstract

Marine heatwaves (MHWs) are prolonged events of extreme sea surface temperature (SST) that have adverse effects on marine ecosystems and socio-economic aspects. Therefore, accurately and effectively predicting SST and MHW events in advance is crucial for mitigating their adverse effects. However, prediction of extreme events over the long term remains a significant challenge. In this study, we apply a deep-learning technique based on Informer, combined with the Empirical Orthogonal Function (EOF) and Empirical Mode Decomposition (EMD), named EOF-EMD-Informer, to improve the prediction of MHWs' occurrence on spatiotemporal scales. Extensive experiments in the Bohai Sea, based on daily satellite-observed SST, shows that the proposed Informer-based model outperforms recurrent neural networks and their variants in medium-term prediction of SST and MHWs' occurrence. The model performs well in predictions up to 30 days ahead, with a root mean square error of about 0.96 °C and an F1 score of about 0.93. About 51 % of the spatial grids have a root mean square error smaller than 0.55 °C with EOF-EMD-Informer model, representing an improvement of approximately 5 % and 27 % compared to the EMD-Informer and Informer models, respectively. This study serves as a proof of concept, demonstrating the potential applications of Informer-based methods in medium-term (up to at least 30 days) predictions of daily SST and MHWs and highlighting their effectiveness in extensive spatiotemporal predictions.