Dynamic Residual Attention UNet for Precipitation Nowcasting Based on WGAN

In recent years, using radar echo maps for precipitation nowcasting has been a research hotspot. How to use deep learning methods to forecast precipitation is a challenge. Radar echo map contains rich temporal and spatial information, capturing the location distribution and intensity characteristics of radar echo is a key problem in precipitation prediction. To tackle these challenges, the paper presents a novel approach called the Dynamic Residual Attention UNet model(DRA-UNet). This model incorporates Decoupled Dynamic Filter(DDF) and Dynamic Residual Attention Modules(DRAM) while leveraging the Wasserstein GAN training strategy to perform generative adversarial training. A decoupled Dynamic Filter can adaptively adjust the convolution kernel in the feature extraction stage, effectively reducing blank areas in the feature maps. By exploring the correlation between residual paths and input image statistics, and appropriately weighting each residual path, the model's focus on precipitation positions is enhanced. Moreover, the utilization of the Wasserstein GAN(WGAN) strategy during model training enhances the image generation quality when facing the discriminator in adversarial training. This advancement ensures that the model's outputs closely approximate real results, leading to further improvements in overall model performance. We comprehensively evaluate the performance of our model on the KNMI dataset, and a large number of experimental results show that our method achieves remarkable results on the precipitation prediction task.