SWINAT-UNET: A NEW BACKBONE FOR PRECIPITATION NOWCASTING

Conventionally, Weather forecasting mainly relies on numerical weather prediction. A downside of numerical weather prediction is that it has high computational and time requirements. With the explosive growth of observational weather data, data-driven deep learning models show impressive potential in precipitation nowcasting tasks. In this paper, we introduce SwinAt-UNet, an efficient data-driven predictive model. The SwinAt-UNet model which combines the UNet and Swin Transformer models can adaptively capture the short-term and long-term dynamic evolution law of radar echo. The proposed model is further equipped with depthwise-separable convolutions and attention modules to improve the generalization ability and forecast accuracy. We evaluate our approaches using weather radar detection data with a high spatiotemporal resolution of Shanghai City. The experimental results show that the forecast accuracy of the SwinAt-UNet model is higher than that of other test models under different reflectivity thresholds. The proposed model has potential advantages and application value in precipitation nowcasting tasks.