Skillful deep learning-based precipitation nowcasting based on new AI-synthetic radar data

Accurate and timely rainfall nowcasting is important for protecting the public from heavy rainfall-induced disasters. In recent years, deep-learning models have been demonstrated to significantly outperform traditional methods in heavy rainfall nowcasting. However, the performance of existing deep-learning-based nowcasting models is still limited by short effective prediction time (< 3 hours), insufficient training data, and the rapid growth of blurriness increases in forecast time. In this work, we propose a novel heavy rainfall nowcasting model based on an innovative task-segmented architecture, namely the TS-RainGAN, consisting of two modules: the MaskPredNet predicts the spatial coverage of different rainfall categories to provide bounding for rainfall with various intensities, and the IntensityGAN predicts the intensity of rainfall based on the rainfall coverage produced by the MaskPredNet. To overcome the data scarcity, we develop another novel model called RainMaker which can generate huge amounts of new radar data based on limited observed radar data. In addition, in order to improve the typhoon-induced precipitation nowcasting, we design a new typhoon structure segmentation method, which distinguishes the typhoon structure into typhoon eye, principal rainband, and outer rainband. The TS-RainGAN training with new AI-synthetic radar data produced by RainMaker can accurately capture the spatiotemporal features and evolutions of rainfall systems and provide skillful precipitation prediction with high skill scores compared with the results of the widely used baseline models. The performance of typhoon-induced precipitation nowcasting shows significant improvement by our innovative typhoon structure segmentation method. Meanwhile, the blurriness of the predicted images is significantly reduced. This enables district-level heavy rainfall nowcasting with competitive forecast skills for up o 6 hours.