Partial Beam Blockage Correction for Improving Radar Quantitative Precipitation Estimation

Weather radar observations often suffer from missing or low-quality data regions, particularly due to beam blockage. Correcting radar observations affected by partial or full beam blockage is crucial for data quality control and quantitative applications, especially in complex terrain environments like the western United States. In this paper, we propose a deep learning framework based on generative adversarial networks (GANs) to restore partial beam blockage regions in polarimetric radar observations. Two types of precipitation data from S-band Weather Surveillance Radar – 1988 Doppler (WSR-88D) stations, KFWS in northern Texas and KDAX in northern California, are utilized for training the GAN model. We manually simulate partial beam blockage situations in both radar data during the training phase. The trained model is tested using independent precipitation events in the DFW area and northern California to demonstrate its capability in repairing missing data with both similar and different precipitation events. Results show that the deep learning-based inpainting approach significantly improves the continuity of precipitation systems for both domains. For KFWS radar data, the deep learning model outperforms conventional interpolation methods, while for KDAX test dataset, the trained model for heavy precipitation achieves comparable results to the model for similar precipitation events. To validate the effectiveness of the repaired images for the blocked KDAX radar data in Northern California, we employ quantitative precipitation estimation (QPE) based on Drop Size Distribution (DSD) data. DSD data provides information on raindrop size distribution, which directly relates to rainfall intensity. By quantitatively comparing the QPE data obtained from the repaired radar display scenes with the ground truth DSD data, we can verify the accuracy and reliability of the CGAN-based repair approach. This comprehensive evaluation demonstrates the efficacy of the proposed deep learning method for restoring radar observations impacted by partial beam blockage, enhancing radar data usability for weather forecasting and hydrological modeling applications.