Radar Partial Beam Blockage Correction for Improving Precipitation Mapping

Weather radar observations often encounter challenges related to missing or suboptimal data segments, notably arising from beam blockage. Correcting radar observations affected by partial or complete beam blockage holds paramount importance in ensuring data quality control and enabling quantitative applications, particularly in complex terrain regions such as the western United States. In this paper, we introduce a sophisticated deep learning framework rooted in generative adversarial networks (GANs) to rectify partial beam blockage regions in polarimetric radar observations. To train the GAN model, we leverage two distinct types of precipitation data gathered from S-band Weather Surveillance Radar - 1988 Doppler (WSR-88D) stations, namely KFWS in northern Texas and KDAX in northern California. During the training phase, different scenarios of partial beam blockage in both radar datasets are simulated. The model is tested using independent precipitation events in both the Dallas-Fort Worth (DFW) area and northern California. This evaluation aims to showcase the model's adeptness in restoring missing data stemming from both similar and dissimilar precipitation events. Our findings reveal that the deep learning-based inpainting approach significantly enhances the continuity of precipitation systems within both geographical domains. In addition, we have employed a quantitative precipitation mapping scheme to the radar data before and after partial beam blockage correction to further demonstrate the effectiveness of the corrected radar data for quantitative applications. Thorough comparison between precipitation estimates derived from the repaired radar data and the ground truth (i.e., rain gauge data) has shown that the enhanced radar data quality can lead to dramatic improvement on precipitation estimation.