Reconstruction of Radar Composite Reflectivity Based on Satellite Observations and Deep Learning

Ground-based weather radar provides crucial information for severe weather monitoring and forecasting, but it faces coverage limitations in regions with complex terrain, especially for oceanic and mountainous regions. To address the limitation, this study proposes "Echo Reconstruction UNet (ER-UNet)", a novel deep learning approach that reconstructs radar composite reflectivity (CREF) using Fengyun-4A geostationary satellite observations with broad coverage. The proposed ER-UNet enhances the UNet architecture by integrating wavelet transforms and multi-scale feature extraction mechanisms, significantly improving the network's capacity to capture detailed radar echo characteristics. Experimental results demonstrate that ER-UNet achieves superior performance compared to UNet, with improvements in statistical metrics that include root mean square error (RMSE), mean absolute error (MAE), and structural similarity index measure (SSIM), as well as categorical verification scores including probability of detection (POD), false alarm rate (FAR), critical success index (CSI), and Heidke skill score (HSS). Case studies further reveal ER-UNet's enhanced capability in reconstructing strong echo features, particularly in terms of intensity distribution and spatial structure. The proposed method shows potential for providing reliable radar reflectivity estimates in areas with limited radar coverage, offering valuable support for severe weather monitoring and early warning services.