GAN-argcPredNet v2.0: a radar echo extrapolation model based on spatiotemporal process enhancement

Abstract

Abstract. Precipitation nowcasting has important implications for urban operation and flood prevention. Radar echo extrapolation is a common method in precipitation nowcasting. Using deep learning models to extrapolate radar echo data has great potential. The increase of lead time leads to a weaker correlation between the real rainfall evolution and the generated images. The evolution information is easily lost during extrapolation, which is reflected as echo attenuation. Existing models, including generative adversarial network (GAN)-based models, have difficulty curbing attenuation, resulting in insufficient accuracy in rainfall prediction. To solve this issue, a spatiotemporal process enhancement network (GAN-argcPredNet v2.0) based on GAN-argcPredNet v1.0 has been designed. GAN-argcPredNet v2.0 curbs attenuation by avoiding blurring or maintaining the intensity. A spatiotemporal image correlation (STIC) prediction network is designed as the generator. By suppressing the blurring effect of rain distribution and reducing the negative bias by STIC attention, the generator generates more accurate images. Furthermore, the discriminator is a channel–spatial (CS) convolution network. The discriminator enhances the discrimination of echo information and provides better guidance to the generator in image generation by CS attention. The experiments are based on the radar dataset of southern China. The results show that GAN-argcPredNet v2.0 performs better than other models. In heavy rainfall prediction, compared with the baseline, the probability of detection (POD), the critical success index (CSI), the Heidke skill score (HSS) and bias score increase by 18.8 %, 17.0 %, 17.2 % and 26.3 %, respectively. The false alarm ratio (FAR) decreases by 3.0 %.