Data-Driven Spatiotemporal Early-Warning Framework for Marine Heatwaves Based on Intensity–Duration Schemes

Abstract

A rapid and accurate risk early warning system of marine heatwaves (MHWs) is crucial for mitigating the potential compound ecological and socioeconomic impacts. However, research on early warning is limited, in particular for China's marginal seas and adjacent offshore waters (CMSOW). This study proposes an innovative early warning system—reversible residual Transformer and bidirectional gated recurrent unit (RTG), for predicting the risk level of MHWs in CMSOW. The framework (1) constructs a duration–intensity MHW-risk taxonomy from historical observations and (2) integrates Bi-GRU–Transformer networks with reversible residual layers and the Fast Attention via Positive Orthogonal Random Features (FAVOR+) attention mechanism to forecast sea surface temperature (SST) and identify potential marine heatwave risks. Extensive experiments show that (1) pairing a Bi-GRU with a Transformer cuts the SST prediction root mean squared error (RMSE) by 40% compared with either model alone; (2) the proposed RTG framework lowers the overall SST prediction RMSE by 8.08% relative to these benchmark models (RNN, LSTM, GRU, Bi-LSTM, Bi-GRU, TCN and Transformer); (3) for 15-day lead forecasts, RTG maintains superior stability, achieving an additional 32% reduction in RMSE and (4) in the early warning context, it lowers the false alarm rate by 3.8% and raises the Dice coefficient by 15.6%.