Leveraging Deep Learning to Extract Raindrop Size Gamma Distribution Parameters from Wind Profile Radar Data

The raindrop size distribution (DSD) is crucial in study of precipitation microphysics, but the parameters that characterize the DSD is not easy to be accurately derived. Wind profile radar (WPR) provides rich data at high spatiotemporal resolution and low attenuation, and previous studies using WPR power spectral distribution to extract DSD parameters demonstrate certain limitations. In this study, a deep-learning architecture that combines neural network (NN) and long short-term memory (LSTM) network was adopted to take advantage of the high-resolution WPR and disdrometer data collected from 2017 to 2021 in Beijing as training data, and a WPR Gamma DSD parameter estimation network (WPR-DSDnet) was further developed by using the WPR reflectivity factor and velocity spectral width as inputs and the disdrometer data derived Gamma parameters as labels. The WPR-DSDnet model is capable to derive Gamma DSD parameters, including normalized intercept parameter (Nw) and mass-weighted average diameter (Dm). Three precipitation cases were used to validate the performance of this deep-learning model. The results demonstrate that WPR-DSDnet is capable to effectively derive the Gamma DSD parameters. The ratio between the estimated and measured values of both lgNw and Dm exceeds 98%, indicating accurate estimation of the DSD parameters. The absolute deviation between the estimated and true values of lgNw is larger than that of Dm, but the relative deviation is smaller, probably due to a larger distribution range of lgNw that results in a larger absolute deviation; overall, the estimation of lgNw is slightly more accurate. The vertical distribution of lgNw closely aligns with variations in precipitation intensity, making it a useful indicator of precipitation intensity change. On the other hand, the distribution of Dm is closely associated with the degree of convection, making it valuable for precipitation recognition and classification. The two parameters are successfully and efficiently extracted by leveraging the deep learning model, facilitating further in-depth analysis of precipitation characteristics and mechanisms.