Abstract
Tropical cyclone (TC) intensity prediction is a critical task in disaster prevention and mitigation. Due to the high complexity and incomplete evolution mechanism, the accurate prediction of TC intensity remains a challenge. This study proposes a Few-Shot Spatiotemporal Fusion Network (FS-STFNet) to improve short-term TC intensity prediction under limited data. A spatial enhancement module is designed to adaptively fuse and enhance features from multi-source data (i.e. trajectories, atmospheric reanalysis, and satellite imagery). Then, an autoregressive deep learning model with enhanced spatiotemporal transformation equation is designed for short-term intensity prediction of single TC without relying large number of temporal training samples. A test applying the FS-STFNet to TCs in Northwest Pacific from 2019 to 2021 demonstrate the effectiveness and superiority compared to other methods. The model achieves optimal performance during stable TCs structures while effectively reducing biases for weak TCs or topography-affected cases through adaptive adjustments. This work offers a scalable technical framework for extreme weather forecasting in temporal data-scarce scenarios, supporting disaster preparedness.