The group has made research progress in the direction of urban flood risk assessment.

As a key component of rainfall estimation, the study of droplet size distribution (DSD) is a hotspot of meteorological and hydrological interest. Weather radar can observe precipitation microphysical processes at large spatial scales, so it is widely used in DSD estimation studies.The traditional polynomial regression algorithm, which correlates DSD parameters with radar features, is still widely used due to its simple structure and acceptable accuracy. However, the regression method has limitations on the application scenarios, and its relationship extrapolation and spatial and temporal prediction are currently not well understood, and there is great uncertainty in large-scale applications. To address this issue, Dr. Jingsuan Zhu published a research result entitled ‘Raindrop size distribution (DSD) retrieval from polarimetric radar observations using neural networks’ in Atmospheric Research in September 2024, which was published by Dr. Jingsuan Zhu in the journal Atmospheric Research.

Long Short-Term Memory (LSTM) network is an extended system to deal with the gradient vanishing problem and degree memory in sequence modelling, which can effectively deal with time series data with significant time lags between important events, and provides a new idea for radar inversion of DSD parameters. In this study, a new DSD retrieval method is proposed based on the LSTM network technique using dual-polarised radar observations, i.e., three schemes capable of retrieving the normalised gamma DSD parameters, namely, LSTM-D0, LSTM-Nw, and LSTM-μ, are proposed for different combinations of dual-polarised radar measurement inputs.In contrast to polynomial regression methods, all LSTM estimators show better performance than the polynomial regression methods.

The results indicated that the Nash efficiency coefficients for estimating the median diameter of raindrops (D0) and the intercept parameter (Nw) increased from 0.93 and 0.70 to 0.95 and 0.93, respectively, and the performance of estimation of the shape parameter (μ) was even more greatly improved at the Chilbolton station in the UK. In addition, this study estimates the temporal and spatial predictability, D0, Nw, and μ, as measured by the Nash coefficient increased by 0.08, 0.31, and 0.39 in time and 0.03, 0.19, and 0.23 in space, respectively.This study contributes to the improvement of quantitative estimation of precipitation from radar polarimetry for a better understanding of precipitation microphysical processes.