Development of a 3D Precipitation Nowcasting Method Using U-Net with Phased Array Radar Data

Rapidly developing heavy rain causes a lot of damages. Therefore, rapid and precise forecasting of short-term developing convective rain cells, called nowcasting, is very important for disaster prevention. In this work, we develop a radar-based 3D precipitation nowcasting method using a neural network based on U-Net.

We utilized a dataset obtained by the Multi-Parameter Phased Array Weather Radar at Saitama in Japan, to learn the evolution of 3D convective rain cells during summertime. Each radar data had three dimensions, where the lengths in the x, y and z directions were 40 km, 40 km, and 16 km, respectively, and the spatial resolution was downsampled to 500 m due to the maximum memory size. The total duration of the dataset was about 5 hours, with a time resolution of 30 seconds.

We predicted radar reflectivity after 10 minutes by the conventional advection nowcasting and the proposed U-Net nowcasting. In the advection nowcasting, we obtain a motion vector from radar reflectivity at two points in time, and linearly extrapolate the current radar reflectivity with the obtained motion vector. In the proposed U-Net method in Fig. 1, we input 1-minute-interval observation data up to 4 minutes before the current time as one 3D data having 5 channels, and the 3D radar reflectivity after 10 minutes is generated. Each data was normalized before being processed by the network. We used the mean squared error as the loss function to train our U-Net and incorporated the residual learning with the advection nowcasting results, where the optimizer was RMSprop of learning rate 0.001.Our U-Net uses 3D spatial convolutions and considers the time-steps of the input observations as channels.

Prediction results of the two methods were compared with the correlation coefficient (CC) and the critical success index (CSI). CSI is one of the indicators to evaluate the estimation performance for rare events. We employed 37.5 dBZ and 10 dBZ as reflectivity threshold values to calculate CSI, which correspond to 10.1 mm/h and 0.2 mm/h in rainfall intensity, respectively. In the advection nowcasting, CSI was 0.31 for 37.5 dBZ and 0.40 for 10 dBZ, and CC was 0.54 for test data. In the proposed U-Net, CSI was 0.23 for 37.5 dBZ and 0.41 for 10 dBZ, and CC was 0.58 for the same test data. By using U-Net, we got better scores in CSI for 10 dBZ and CC than the advection nowcasting.