Generative Tropical Cyclone Wind Model by Artificial Intelligence Based on SAR Measurement

This is the authors' abstract. We don't add key points for this paper.

Remote sensing provides indispensable observations for monitoring tropical cyclone (TC) surface winds, yet traditional numerical or parametric models often underestimate structural details and remain under-sampled. Leveraging more than 300 Sentinel-1 (S-1) synthetic aperture radar (SAR)-derived CyclObs wind products and the International Best Track Archive for Climate Stewardship (IBTrACS) best-track data from more than 200 TCs (2018–2024), we trained eXtreme Gradient Boosting (XGBoost) and Transformer models to reconstruct two-dimensional TC wind fields, and evaluated them on 100 independent TCs (2024–2025) against measurements from microwave radiometers onboard Soil Moisture Active Passive (SMAP), Advanced Microwave Scanning Radiometer-2 (AMSR-2) and Stepped-Frequency Microwave Radiometers (SFMR) measurements. The Transformer achieves the highest point-wise accuracy, with root-mean-square errors (RMSEs) of 2.85 m/s against SMAP/AMSR-2 and 3.06 m/s against SFMR, as well as superior correlation and structural similarity compared with XGBoost and the Generalized Asymmetric Holland Model (GAHM). However, its along-track reconstructions exhibit pronounced small-scale oscillations absent in the physically smooth GAHM, revealing that numerical closeness to observations does not guarantee physical authenticity. This purely data-driven approach lacks dynamical constraints and risks unphysical artifacts between tracks, thereby demanding independent two-dimensional validation. Structural validation against IBTrACS shows excellent correlations ( r ≥ 0.90) for 34-kt (R34), 50-kt (R50), and 64-kt (R64), with RMSE decreasing from 42.71 to 11.03 km and biases converging toward zero, confirming faithful reproduction of inner-core gradients despite slight smoothing of the outer rainbands. Across 2–10 km resolutions, RMSE remains nearly constant (∼2.35 m/s) and maximum wind speed shows no systematic change, indicating that finer grids only sharpen asymmetry without improving error metrics, likely due to overfitting or interpolation rather than real oceanic signals. Performance degrades sharply above 50 m/s (sparse extreme-wind data) and for translation speeds larger than 11 m/s (inadequate learning of motion-induced asymmetry), yet these limitations are purely statistical, lacking mechanistic insight. Thus, while the Transformer outperforms conventional models and is viable for routine operational TC warning applications, its application to extreme wind speeds (>50 m/s) or fast-moving storms (>11 m/s) is inadvisable without rigorous physical validation, ideally coupled with dynamical models and bias-corrected training data.