A Dual-Branch Encoder–Decoder Network with Convolutional Long Short-Term Memory and Efficient Channel Attention for Forecasting Summer Marine Heatwaves in the East China Sea

Frequent marine heatwaves (MHWs) threaten marine ecosystems and economies. To accurately anticipate MHW occurrence probability in the East China Sea (ECS), this work proposes MHW-NET, a deep learning model. The model uses an encoder–bottleneck–decoder dual-branch architecture, ConvLSTM, efficient channel attention (ECA), and a composite loss function with focal Tversky and SSIM losses. The model is trained on data from 1982 to 2011, with an independent test period of 2016–2022. Over lead days 1 to 7, MHW-NET performs relatively well in the Yellow Sea, while exhibiting substantially weaker performance in complex nearshore regions such as Bohai Bay, the Zhejiang coast, and the waters off northern Taiwan. The model accurately characterizes the spatial distribution of MHWs and reproduces the spatial patterns of the two extreme MHW events in 2016 and 2022, with spatial correlations exceeding 0.85. Feature importance analysis indicates that sea surface temperature and 2-m air temperature are the main driving factors of MHWs in the ECS. This study’s ablation experiments show that ConvLSTM and ECA are core modules and noise and refinement are auxiliary. Benchmark comparisons with baseline models further confirm the superiority of MHW-NET, demonstrating its potential for improving MHWs forecasting and early warning in the ECS.