Transfer Learning-Based Generative Adversarial Network Model for Tropical Cyclone Wind Speed Reconstruction From SAR Images

Synthetic aperture radar (SAR) plays a crucial role in monitoring the fine structure of tropical cyclones, but its effectiveness is constrained by limitations such as signal degradation and saturation. To address this challenge, we proposed a transfer learning-based generative adversarial network (GAN) framework with a dilated convolution and attention mechanism for reconstructing inner-core high winds from SAR images. We have employed the principles of transfer learning to adapt pretrained models developed by the Hurricane Weather Research and Forecasting (HWRF) model winds to SAR images during tropical cyclone events for reconstruction. The proposed model can effectively capture the relationship between features in the low-precision areas and global features from SAR images, facilitating tropical cyclone wind speed reconstruction. The utilization of Global Precipitation Measurement (GPM) Level 3 rainfall data facilitates the identification of rainfall regions in 89 SAR images obtained from Radarsat-2 and Sentinel-1A/B missions. Comparison with stepped frequency microwave radiometer (SFMR) data reveals that the model exhibits a bias of −0.69 m/s, a root-mean-square error (RMSE) of 4.08 m/s, and an R value of 0.91 under heavy rainfall conditions (>7.62 mm/h). Remarkably, the GAN model exhibits excellent performance compared with measurements from the soil moisture active passive (SMAP) L-band radiometer, achieving an RMSE of 3.78 m/s. Our findings indicate that deep learning (DL) technology holds significant promise for the reconstruction and monitoring of tropical cyclones through the utilization of SAR imagery.