Multi‐Scale Spatio‐Temporal Evidence Learning for Tropical Cyclone Unusual Track Forecasting Over the Western North Pacific

Abstract

Despite significant advances in tropical cyclone track forecasting, unusual tracks remain a major challenge, primarily due to their substantial prediction uncertainty and inadequate representations. In this paper, we present the Tropical Cyclone Unusual Track Forecasting (TCUTF) model, a novel artificial intelligence‐based approach specifically designed for predicting unusual tracks over the Western North Pacific. To address these aforementioned challenges, TCUTF introduces two key innovations: uncertainty quantification and enhanced representation learning. For uncertainty quantification, TCUTF enables simultaneous prediction and uncertainty estimation in a single forward pass by employing an evidential network, thus allowing the model to quantify prediction uncertainty. For enhanced representations, TCUTF incorporates two components to better capture the complex physical characteristics and spatiotemporal features underlying unusual track patterns: (a) physics‐guided feature engineering that utilizes satellite cloud imagery differences as the primary input to represent cyclone movement patterns, augmented by satellite‐derived pseudo‐relative vorticity and intensity‐aware heatmap differences to comprehensively capture rotational dynamics and center distribution characteristics; and (b) a multi‐scale spatio‐temporal analysis module that employs spatial and temporal wavelet transforms to effectively model the multi‐scale spatiotemporal features of unusual tracks. Experimental validation demonstrates that TCUTF achieves F1‐scores of 0.274, 0.353, and 0.878 for northward deflection, westward deflection, and no deflection predictions, respectively, while maintaining accuracy for high‐confidence predictions and appropriate uncertainty calibration for challenging cases.