Reply on CC1

Abstract

<strong class="journal-contentHeaderColor">Abstract.</strong> Precipitation nowcasting has important implications for urban operation and flood prevention. Radar echo extrapolation is the 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 real rainfall evolution and generated images. The evolution information is easily lost during extrapolation, which is reflected as echoes attenuation. Existing models, including Generative Adversarial Network (GAN)-based models, are all difficult to reduce loss and curb attenuation, which results in insufficient rainfall prediction accuracy. Aim to the problem, a Spatiotemporal Process Intensification Network (GAN-argcPredNet v2.0) based on GAN-argcPredNet v1.0 is designed. GAN-argcPredNet v2.0 reduces the loss by intensifying the influence of the previously input evolution information. A Spatiotemporal Information Changes Prediction (STIC-Prediction) network is designed as generator. With the intensification of echo feature sequence, the generator focuses on the spatiotemporal variation and generates more accurate images. Furthermore, discriminator is a Channel-Spatial Convolution (CS-Convolution) network. The discriminator intensifies the discrimination of echoes information by strengthening spatial information of single image. Identification results are fed back to the generator, which reduces the loss of important evolutionary information. The experiments are based on the radar dataset of South China. The results show that GAN-argcPredNet v2.0 performs better than other models. In heavy rainfall prediction, compared with baseline, the Probability of Detection (POD), the Critical Success Index (CSI), and the Heidke Skill Score (HSS) increase by 24.8 %, 22.2 % and 21.5 % respectively. The False Alarm Ratio (FAR) decreases by 3.76 %.