A Spatiotemporal Multimodal Fusion Network for Tropical Cyclone Intensity Forecasting

Tropical cyclones (TCs) pose severe threats, yet accurate forecasting of its intensity remains challenging due to limitations in feature utilization, spatiotemporal sequence modeling, and multi-time-step forecasting consistency. To address these issues, a SpatioTemporal MultiModal Fusion Network for multi-step intensity forecasting, termed ST-MMFN, is proposed based on the TCIR (TC image-to-intensity regression) dataset. ST-MMFN builds a three-dimensional (3D) feature joint modeling architecture that combines TC temporal, spatial, and frequency domain data by fusing satellite images, intensity statistics, and wavelet frequency domain information. More importantly, to meet the forecasting requirements across different timescales, a hierarchical forecasting mechanism is designed. The immediate prediction head directly captures recent intensity changes, while the extended prediction head integrates immediate outputs to perform recursive extrapolation for longer horizons. Experimental results demonstrate that ST-MMFN significantly outperforms existing baseline models, achieving an average mean absolute error (MAE) of 2.90 m s^-1 and an average root mean squared error (RMSE) of 4.17 m s^-1 (R² = 0.90) across multiple time steps including 3, 6, 12, and 24 h. This provides an effective framework for TC intensity forecasting.