Deep learning improves sub-seasonal marine heatwave forecast

Abstract

Marine heatwaves (MHWs) are extreme anomalously warm water events, which are projected to cause increasing numbers of disastrous impacts on ecosystems and economies under global ocean warming. Our ability to forecast MHWs determines what effective measures can be taken to help reduce the vulnerability of marine ecosystems and human communities. In this study, we combine a deep learning model, the convolutional neural network, with a real-time sub-seasonal to seasonal physical forecast model, improving MHW forecast skills by nearly 10% of the global average in leading two weeks by correcting the physical model bias with observational data. This improvement has a nearly consistent influence (∼10%–20%) on a global scale, reflecting the wide-coverage promotion by deep learning. This work reveals the advantages and prospects of the combination of deep learning and physical models in ocean forecasts in the future.