Rad3DGFM: A Deep Learning Network for Completing Data Gaps in 3D Radar Reflectivity Mosaics

Three-dimensional (3D) radar reflectivity mosaic products generated from weather radar networks, such as the Severe Weather Automatic Nowcasting (SWAN) system of the China Meteorological Administration, provide essential information for severe convective weather nowcasting, short-term forecasting, and quantitative precipitation estimation. However, these products often contain lower-level missing-data gaps caused mainly by Earth curvature and terrain blockage of radar beams. To address this problem, this study develops a Radar 3D Gap-Filling Model (Rad3DGFM) to reconstruct missing reflectivity information. Rad3DGFM is built upon a 3D U-Net architecture integrated with an attention mechanism and an enhanced conditional generative adversarial network (CGAN). The CGAN is improved by using a least-squares adversarial loss, label smoothing, and Gaussian-noise-perturbed discriminator training, together with MAE and MSE reconstruction constraints in the gap regions. This design improves training stability and helps the model generate radar reflectivity fields that are both structurally realistic and quantitatively accurate. The experimental results based on 6,562 test samples demonstrate that Rad3DGFM effectively reconstructs the missing data for all precipitation intensities. For reflectivity values exceeding 35 dBZ, the model achieves POD, FAR, TS, and FSS values of 0.71, 0.13, 0.67, and 0.81, respectively. Furthermore, the model performs robustly when applied to regions with different terrain conditions and data gaps from the training area, suggesting that Rad3DGFM has the potential to be applied in different regions without additional training.