STAW-Net: Spatio-Temporal Attentional Windmill Network for Tropical Cyclone Intensity Prediction

Tropical cyclone (TC) intensity prediction using remote sensing images plays an important role in weather forecast and disaster prevention. Though having achieved great success in image analysis, deep learning methods lack feature extractors specifically designed for TC structures, resulting in limited representation capabilities for TC structure. Moreover, most of these methods overlook that TC is a temporal system with strong time dependencies. To this end, we propose a spatiotemporal attentional windmill network (STAW-Net) for TC intensity prediction. Specifically, we extract temporal-spatial-spectral features from multi-temporal TC multispectral images (MSIs). Notably, we design a novel Windmill Convolution (WMConv) to simulate the dynamic characteristics of TC occurrence and to extract structural features related to the TC eye region. Finally, we combine these temporal-spatial-spectral features with structural features to predict the maximum sustained wind (MSW) speed. Experimental results show that STAW-Net outperforms many baselines and state-of-the-art methods on various cyclone datasets. The errors of 24-hours MSW prediction by STAW-Net can be as low as 6.37m/s and 5.64m/s for the TCIR and the HURSAT-B1 datasets, respectively.