Fusion of Dual Wind Component for Radar Echo Nowcasting Based on a Deep Learning Model

Nowcasting serves as a pivotal tool in predicting severe weather phenomena, which is essential for disaster prevention and mitigation. However, existing nowcasting models often struggle due to insufficient background information input. In this study, we propose the Late Fusion Wind Filed UNet (LFWF UNet) model, which integrates radar data and 3D wind field data to enhance the three-dimensional physical back-ground information necessary for the development of convective systems. This integration significantly improves forecasting capabilities. Additionally, we introduce a novel variable fusion method to address the crosstalk effects associated with direct multi-channel fusion. The results, evaluated using several metrics, show that the LFWF UNet model outperforms other experimental approaches in both point-by-point evaluations, binary diagnostic scores and computer vision evaluation metrics. Furthermore, the synergistic use of radar data with multilayer wind field data enhances forecasting accuracy. Therefore, the approach of combining multidimensional physical wind field information holds great promise for nowcasting and can effectively improve extrapolated forecast results.