Machine Learning Approach for Polarimetric Retrieval of Raindrop Size Distributions

Retrieving DSD parameters is a necessary step toward comprehending precipitation processes. Conventionally, this is achieved through the use of polynomial or power-law regression models, which relate the DSD parameters with measurable radar parameters such as ZH and ZDR. However, due to the inherent limitations of linear additive models, accurately capturing the intricate nonlinearity of this relationship can be difficult. In contrast, machine learning (ML) algorithms offer a promising solution to overcome these limitations. These techniques can effectively handle the complex non-linear relationship between the independent variables and the response variable without making any distributional or modeling assumptions.
In this study, we have examined the capability of the use of an ML algorithm to retrieve the DSD parameters (generalized number concentration, N0′, and generalized mean diameter, Dm′) using polarimetric radar variables (ZH, ZDR, KDP, ρHV, and AH). The ML algorithm was trained using simulated dual-polarization variables based on DSDs from a two-dimensional video disdrometer. We utilized warm season data collected over the past ten years from different locations in South Korea.
The use of the retrieval algorithm in analyzing the DSD evolution of the storm from polarimetric radar observation will be further demonstrated. In order to understand how DSD changes over time and space, a Lagrangian cell tracking method is employed to track the precipitation system. Then the trained ML algorithm was applied to the polarimetric radar variables of the tracked storm. This study highlights that the ML algorithm can be an effective tool in understanding the DSD characteristics from polarimetric radar measurements.

ACKNOWLEDGMENT
This work was funded by the Korea Meteorological Administration Research and Development Program “Observing Severe Weather in Seoul Metropolitan Area and Developing Its Application Technology for Forecasts” under Grant (KMA2018-00125). This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2021R1A6A3A13042215)