Physically Consistent Tropical Cyclone Monitoring From Infrared Satellite Imagery via Knowledge‐Guided Deep Learning

Abstract Tropical cyclones (TCs) pose significant threats to coastal communities and ecosystems underscoring the need for accurate and timely monitoring of their structure and intensity. Although satellite infrared (IR) imagery provides continuous global coverage, traditional methods for estimating TC parameters, such as center location, intensity, and wind structure, often rely on empirical relationships or manual analysis, limiting their accuracy, scalability, and responsiveness. These approaches struggle to capture the nonlinear dynamics of TCs and cannot deliver consistent multiparameter estimates from a single data source. To address these limitations, we propose a unified, knowledge‐guided deep learning framework, KG‐TCM, that simultaneously estimates TC center, intensity, and 34 kt wind radius (R34) from geostationary IR imagery. By incorporating prior physical knowledge into a multitask learning framework, KG‐TCM aligns IR data with best‐track records and reanalysis‐derived environmental variables enabling physically consistent and automated inference. Tested on 6,790 IR images of western North Pacific cyclones, KG‐TCM achieves a center location error of 26.73 km, a mean absolute error of 5.24 m/s in intensity, and a R34 error of 33.82 km. The model demonstrates strong generalization to unseen TCs and outperforms baseline methods in accuracy and robustness. Its physically meaningful outputs and low computational requirements highlight its potential for real‐time operational TC monitoring systems integration.