Deep Learning-based Fusion of Analysis and Satellite Measurements for Ocean Surface Wind Downscaling for Tropical Cyclones

Accurate forecasting of tropical cyclones is crucial for safeguarding coastal areas against the loss of life and property. Near-real-time analysis data, such as the Cross-Calibrated Multi-Platform Ocean Surface Wind Vector (CCMP), is widely utilized for predicting tropical cyclones due to its comprehensive coverage and consistent temporal and spatial measurements. However, CCMP has a limited resolution of 25 kilometers and frequently underestimates wind speeds during tropical cyclones because of rain interference. In contrast, Synthetic Aperture Radar (SAR) can measure ocean surface winds under all weather conditions with a significantly higher resolution of approximately 0.5 kilometers, though it lacks extensive area and temporal coverage. We developed a novel deep learning approach that leverages the strengths of both CCMP and SAR data. By using SAR wind measurements for tropical cyclones globally as the ground truth, we trained our deep learning model on the corresponding CCMP data to enhance its accuracy and spatial resolution. We evaluated various deep learning architectures, including U-Net, DeepLabV3+, and TransUNet. Our results indicate that TransUNet performs the best, improving CCMP's accuracy by 45% for wind speeds over 20 m/s, 20% for overall wind field, 56% for the maximum wind speeds, and 64% for the radius of the maximum wind speed. Our method can create gap-free, high-resolution, and accurate ocean surface wind data for tropical cyclones.