Adaptive Multi-Branch Generative Adversarial Network With Parallel Attention for Robust Tropical Cyclone Forecasting

Guangxin He, Xiaoyu Li, Xiaoyu Li, Jie Feng, Jingjia Luo, Xun Li, Xun Li

Tropical Cyclone Research and Review · 31 July 2026 · Fudan University (FuXi), Nanjing University of Information Science & Technology (NUIST)

Paper

In one sentence. MAST-GAN fuses environmental fields and storm attributes with parallel attention to improve tropical cyclone wind field forecasts, especially for extreme typhoons.

Abstract

Accurate prediction of tropical cyclone (TC) wind fields is crucial for disaster mitigation, yet remains challenging, particularly for extreme events. Existing deep learning models often struggle to capture the diverse dynamics across the TC lifecycle and tend to under-predict intense storms. To address these issues, we propose a Multi-Attention Spatio-Temporal Generative Adversarial Network (MAST-GAN), which couples a multi-source feature encoder that jointly processes environmental fields via an CNN,Transformer backbone and TC attributes via an MLP, then fuses them with adaptive gating, a triple parallel attention fusion module that simultaneously emphasizes channel, spatial and temporal cues, and a multi-branch dynamic prediction head that adapts across the lifecycle by balancing global and TC-core forecasts while using an auxiliary branch to enhance extreme-event prediction. Experiments on a large dataset built from ERA5 and IBTrACS show that MAST-GAN improves TC wind field reconstruction and forecasting accuracy over strong baselines, with clear gains on an extreme-typhoon test set measured by RMSE and MAE, and case studies on Typhoons Bebinca and Pulasan further confirm realistic and robust wind field evolution.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.