A Novel RFAt-UNet3+ Learning Model for Rainfall Forecast With Meteorological Radar

Abstract

Rainfall forecast is generally defined as the prediction of precipitation or severe convective weather in a specific region over a short time interval, recognized as playing an extremely important role in daily meteorological disaster prevention. Traditional precipitation forecasting methods, which primarily rely on numerical weather prediction, have limited the capability to utilize the latest information for short-term precipitation nowcast. A novel deep learning model based on the RFAt-UNet3+, composed of the attention and receptive field modules additionally, is proposed for precipitation nowcasting. Compared with existing models (UNet3+, CBAM_UNet3+, RFBs_UNet3+), a more accurate precipitation forecasting is accomplished by the proposed learning model.