Physics-guided multi-task network for tropical cyclone intensity prediction based on satellite images

Abstract The prediction of tropical cyclone (TC) intensity remains a challenging task. Although satellite imagery is suitable for time-sensitive TC intensity prediction tasks due to its low-latency acquisition, it is underutilized in existing studies because of the complexities involved in modeling dynamic TC features from satellite imagery. In this study, a multi-task neural network framework that leverages satellite imagery as predictors for forecasting future TC intensity is proposed. Specifically, an encoding block is designed to integrate prior information into satellite imagery, and a multi-task strategy is employed to enhance modeling capabilities by learning multiple objectives from shared representations. Furthermore, a physics-guided constraint loss is constructed to exploit the cross-fusion features and mutual constraints of the predicted indicators. The experimental results demonstrate that models built upon the proposed framework achieve competitive performance in TC intensity prediction, which validates the generalization and effectiveness of the framework. Additionally, the ablation studies confirm that both the multi-task strategy and the physics-guided constraint loss effectively improve the prediction performance.