TAASRAD19 Radar Echo Extrapolation Based on SmaAt-UNet Neural Network Equipped with Attention Modules and Depthwise-Separable Convolutions

Weather radar detection data with high temporal and spatial resolution is the main tool of weather forecast. The precipitation forecast can be realized by extrapolating the radar echo image. The traditional radar echo extrapolation method has the defects of short forecasting time and insufficient ability to use radar data. Based on the data-driven neural network method, the data can be fully mined and its inherent law can be learned from the data, and accurate precipitation forecast can be produced. In this paper, SmaAt-UNet neural network is equipped with attention modules and depthwise-separable convolutions. The training and test data sets were constructed by using the radar echo map of the TAASRAD19 dataset, and the radar images of the first 1h were successfully used to forecast the precipitation of the next 1h.