Tropical Cyclone Intensity Prediction by Spectral-Temporal Dislocation and Attention-Based Networks

Accurate prediction of Tropical cyclone (TC) intensity using multispectral images (MSIs) is critical to avoid economic loss and life casualty. Although existing methods have achieved good prediction results, they neglect changes in cloud patterns such as cyclone eyes and cloud spirals, which are closely related to TC intensity. How to leverage temporal-spatial-spectral features of MSIs to improve prediction accuracy is challenging task. In this paper, we propose a novel framework with Spectral-Temporal Dislocation and Attention-Based Networks (STD-AN) to predict MSW speed values near cyclone centers. The STD technique allows the framework to learn temporal-spatial-spectral features of TC. Meanwhile, the Self-Attention Modules (SAM) enable global attention feature extraction and Cross-Attention Modules (CAM) fuse different band features to improve prediction accuracy. Experimental results show that the proposed framework outperforms several state-of-the-art methods for TC intensity prediction.