Raindrop Size Distribution (DSD) Retrieval from Polarimetric Radar Observations Using Neural Networks

As a key component of rainfall estimation, the understanding of raindrop size distribution (DSD) is a longstanding goal in meteorology and hydrology. Given that weather radar can observe the precipitation microphysics over large spatial and temporal scales, it has been broadly applied in DSD estimation. Traditional polynomial regression algorithms that correlate DSD parameters and radar signatures are still widely applied due to their simple structure and acceptable accuracy. This study proposes a new DSD retrieval model using dualpolarization radar observations based on long short-term memory (LSTM) network techniques. Three schemes able to retrieve the parameters of a normalized gamma DSD (LSTM-D0, LSTM-Nw, and LSTM-mu) are proposed with different combinations of polarimetric radar measurement inputs. All LSTM estimators exhibit better performance than the polynomial regression method. The Nash-Sutcliffe efficiency coefficient for estimates of drop median diameter (D0) and intercept parameter (Nw) increases from 0.93 and 0.70 to 0.95 and 0.93 respectively at Chilbolton station. Poor estimates of the shape parameter (mu) using the polynomial regression estimator complicate real applications, whereas the remarkable improvement of LSTM model estimation facilitates practical applications. The temporal and spatial predictability is then estimated to investigate long-term estimator performance for various radars, or at least for all radar pixels of a single radar. The predictability, measured by the Nash coefficient, increases temporally by 0.08, 0.31, and 0.39 and spatially by 0.03, 0.19, and 0.23 for the parameters D0, log10Nw, and mu respectively. This study contributes to improving quantitative precipitation estimates from radar polarimetry, enabling a better understanding of precipitation microphysics.