Application of Dual-Polarization Radar Data Assimilation Via a Deep UNet Network Model

The assimilation of dual-polarization (dual-pol) radar data plays a crucial role in enhancing the simulation of hydrometeors and improving the short-term precipitation forecasts of numerical models. However, current dual-pol radar data assimilation (DA) methods exhibit limitations in terms of computational efficiency and data utilization. In this study, a new dual-pol radar DA approach is developed. This approach includes the application of a UNet-based model, which aims to retrieve mixing ratio information of four types of hydrometeors from dual-pol radar data. The validation results for the UNet-based model indicate that the distributions of the hydrometeor mixing ratios provided by the model are generally consistent with those based on the labeled data, with a reasonable range of root mean square errors (RMSEs). On this basis, the hydrometeor analysis increments retrieved by the UNet-based model are incorporated into the model integration process through the incremental analysis update (IAU) scheme, achieving a complete dual-pol radar DA process for the CMA-MESO model. To evaluate the effect of this DA scheme, comparative simulation experiments are conducted for Typhoon Lekima (2019). The verification results indicate that the use of the hydrometeor DA scheme generally improves the threat scores (TSs) for 3-hour accumulated precipitation during medium- and heavy-rainfall events. Additionally, the 24-hour accumulated rainfall TSs for the medium-, heavy-, and extreme-precipitation categories in the DA experiment are all better than those in the control experiment. The DA method can also be used to better predict the spatial distribution of extreme-rainfall events. These results demonstrate that the application of the proposed dual-pol radar DA approach effectively enhances the precipitation forecasting capabilities of numerical weather models.