3D Precipitation Nowcasting from Phased Array Radar with Uncertainty Estimation Using a Quantile Regression Neural Network

Paper i proceeding, 2025

In Japan, Multi-Parameter Phased Array Weather Radars (MP-PAWR) have been deployed to enhance the observation of sudden localized heavy rainfall. Leveraging these dense 4D observations, an AI-based system delivers real-time precipitation nowcasts (very short-term forecasts with high-resolution) with a 10-minute lead time, outperforming conventional methods for predicting convective rainfall onset. Using an adversarial training method, the system is capable of predicting fine spatial details of rainfall. In this study, a quantile regression neural network technique has been integrated into the model to estimate nowcast uncertainty. It is shown that the quantile levels effectively define credibility intervals but further investigations are required to fully understand the interaction between adversarial loss and quantile loss. Additionally, the model demonstrates its ability to nowcast at different altitudes despite be trained specifically for 2.125km. It successfully generates precipitation nowcasts within the 1-4km range, allowing for 3D prediction of rapid convective precipitation, another significant model enhancement compared to the first version. Future work for improving nowcast reliability, accuracy and lead-time is presented in the concluding remarks.