Physics-Informed Deep Learning for 3D Wind Field Retrieval of Open-Ocean Typhoons

Accurate retrieval of three-dimensional (3D) typhoon wind fields over the open ocean remains a critical challenge due to observational gaps and physical inconsistencies in existing methods. Based on multi-channel data from the Himawari-8/9 geostationary satellites, this study proposes a physics-informed deep learning framework for high-resolution 3D wind field reconstruction of open-ocean typhoons. A convolutional neural network was designed to establish an end-to-end mapping from 16-channel satellite imagery to the 3D wind field across 16 vertical levels. To enhance physical consistency, the continuity equation, enforcing mass conservation, was embedded as a strong constraint into the loss function. Four experimental scenarios were designed to evaluate the contributions of multi-channel data and physical constraints. Results demonstrate that the full model, integrating both visible/infrared channels and the physical constraint, achieved the best performance, with mean absolute errors of 2.73 m/s and 2.54 m/s for U- and V-wind components, respectively. This represents significant improvements over the baseline infrared-only model (29.6% for U, 21.6% for V), with notable error reductions in high-wind regions (>20 m/s). The approach effectively captures fine-scale structures like eyewalls and spiral rainbands while maintaining vertical physical coherence, offering a robust foundation for typhoon monitoring and reanalysis.