Super-resolution downscaling of WRF-based typhoon wind fields over complex coastlines using a decoupled physics-guided U-Net

Accurate kilometer-scale typhoon wind fields are important for coastal hazard assessment, but Weather Research and Forecasting (WRF)-based dynamical downscaling remains computationally expensive, while traditional statistical methods often fail to capture localized topographic effects. This study develops a decoupled physics-guided U-Net to emulate WRF-based downscaling of hourly maximum 10-m typhoon wind fields from 25 km to 3 km resolution. The model integrates 12-channel inputs describing large-scale circulation, typhoon structure, and terrain modulation; depthwise separable convolutions decouple spatial-gradient extraction from cross-variable interactions, attention gates enhance land–sea boundary features, and a Huber loss improves robustness under sparse high-wind samples. Trained on six typhoons and validated on two independent WRF-simulated landfall events, the model reduces MAE by about 30% compared with the Standard U-Net and by approximately 30% relative to Kriging and Random Forest. It also maintains a weighted-average Gradient Root Mean Square Error (G-RMSE) below 3.0 m/s and a Structural Similarity Index Measure (SSIM) above 0.82 across the validation scenarios, indicating improved spatial-gradient and structural fidelity. The framework provides a computationally efficient WRF-based emulator for reconstructing complex coastal wind gradients and supporting rapid wind-hazard assessment.