Item

Abstract

Traditional methods for precipitation nowcasting such as optical flow methods give prediction results by extrapolating the radar echo maps, which doesn’t consider well the generation and disappearance of convection. With the development of deep learning, extrapolating the radar echo maps using deep learning methods has become an important way to predict precipitation. The Global Navigation Satellite System (GNSS) can accurately sense water vapor with high temporal resolutions and short latency, which makes it possible to improve the accuracy of precipitation nowcasting by integrating GNSS and radar observations. In this study, we choose GNSS stations in the German region and make use of the radar products provided by the German Weather Service (DWD) to give a precipitation prediction with deep learning methods. The U-Net, which takes four consecutive radar composite grids and GNSS zenith total delays (ZTDs) as separate input channels (t-15, t-10, t-5 minutes, and t, where t is the time of the nowcast) to produce a nowcast at time t+5 minutes. Verification experiments on several rainfall events are carried out, and the results suggest that the proposed GNSS and radar fusion model reveals an enhanced accuracy of precipitation nowcasting when compared to the conventional method based on radar echo maps only, when the routine verification metrics of Mean Absolute Error (MAE) and Critical Success Index are taken into consideration.