Evaluation of deep learning-based precipitation nowcasting methods during the 2023 flood season in Jiangsu Province

Compared to the traditional nowcasting methods based on extrapolation, deep learning-based precipitation nowcasting methods can effectively forecast the trigger, development, and dissipation of heavy rainfall events. Currently, there are three widely used deep learning methods, which are PhyDNet, PredRNN-v2, and GAN. The first two are spatiotemporal convolutional neural networks under the deep learning framework, while the third is the generative adversarial network with PhyDNet as the generator and multilayer convolution as the discriminator. This study evaluates the application of the above three deep learning-based precipitation nowcasting methods during the flood season of 2023 (April to July) in Jiangsu region. The applicability of these methods in different precipitation types was also discussed through the representative case studies. The results are as follows. (1) From the entire evaluation period, PhyDNet and PredRNN-v2 show better TS performance than GAN. However, GAN performs the best in terms of BIAS for the main precipitation bands, which helps eliminate the "blurry" issue observed in the first two methods with the increase of forecast lead time. (2) During the systematic heavy precipitation periods, the evaluation results for all three methods are consistent with the overall period assessment. However, during the localized heavy precipitation periods, GAN outperforms PhyDNet and PredRNN-v2 in both TS and BIAS. (3) The results of the typical case analysis show that all three methods can capture the evolution of precipitation systems. In the case of systematic heavy rainfall events, PredRNN-v2 exhibits better performance in precipitation intensification compared to the other two methods. In localized heavy rainfall events, GAN not only overcomes the issue of precipitation overestimation caused by "blurriness" but also provides a better representation of the intensity and locations of the heavy precipitation centers. The above results suggest that each of the three forecasting methods has its advantages for different scenarios in heavy precipitation forecasting and warning. PhyDNet performs the best overall, PredRNN-v2 can be used for forecasts with longer lead time, while GAN is suitable for localized heavy rainfall events. Therefore, in practical operational applications, the choice of method should be based on their respective strengths and characteristics.