U-Net Based Retrieval of Precipitation Microphysics from Polarimetric Weather Radar Data

Weather radar polarimetry has enabled more accurate estimation of rain microphysics parameters such as mass- or volume-weighted mean diameter (Dm) and liquid water content (W) by providing comprehensive information about the size, shape, composition, and phase of hydrometeors. Conventional microphysics retrieval methods, such as physics-based inversion and empirical formulas, often result in large biases and errors in parameter estimates due to observational measurement errors, model errors, and nonlinear relationships between radar variables and microphysics parameters. The deep neural network (DNN) has recently been introduced and applied to rain microphysics retrieval and has shown positive impact and potential in reducing estimation errors and improving retrieval performance because of its capability to handle error and nonlinearity effects. The DNN performs retrievals from polarimetric radar data (PRD) at single range gate independently and has limitations, such as not utilizing spatial or temporal information. To further improve microphysical retrievals from PRD and expand the study to multiple hydrometeor species in which there are more unknowns than independent measurements at every range gate, more robust machine-learning tools are needed. In this study, an advanced convolution neural network, the U-Net algorithm, is adapted to retrieve microphysics parameters for multiple species of hydrometeors, including rain, snow, graupel, hail, and their mixtures. Initial training and testing are conducted using simulated data generated by the Weather Research and Forecasting (WRF) model and recently developed forward polarimetric radar operators. The qualitative and quantitative evaluation of the performance of U-Net is conducted by comparing the distributions of Dm and W and their differences between the retrievals and the simulations. Preliminary results indicate that the spatial distribution of both Dm and W for all seven sets of hydrometeor mixtures matches reasonably well with the simulation results. The initial analysis and the potential applications of the Horus radar for precipitation microphysics retrievals using U-Net will also be presented.