Machine Learning for Hydrometeor Particle Identification Using Dual-Polarization Radar Observations

Advancements in hydrometeor particle identification (PID) using dual-polarization radar observations are crucial for quantitative applications such as precipitation estimation and data assimilation. While traditional PID methods such as fuzzy-logic approaches are widely used, the integration of machine learning to resolve different hydrometeor particles and their distributions in polarimetric radar measurements is yet to be explored. This study presents a novel approach to PID that combines a T-matrix scattering model, a random forest classifier, and K-mean clustering for noise reduction, thereby addressing the increasing demand for hydrometeor classification. The T-matrix model provides simulated polarimetric measurements, i.e., Zh, Zdr, Kdp, and ρhv, for different types and distributions of hydrometeors under different environmental conditions, including rain, rain mixed with hail, hail, graupel, and snowflakes, among others. The Random Forest classifier is trained on the simulated radar observables and corresponding hydrometeor labels. To enhance the classification accuracy, we employ K-mean clustering to refine the classification boundaries by reducing the noise. This PID methodology can enhance the identification of intricate hydro-meteorological substances, such as graupel and mixed-phase particles, which are frequently misclassified in conventional methodologies. In addition, a traditional fuzzy logic method is employed as a comparative baseline to evaluate the efficacy of the proposed method, thereby offering insight into its relative accuracy performance. The PID models are validated under a range of meteorological conditions to demonstrate the performance of distinguishing challenging particle types, such as graupel and hail. The results demonstrate that this innovative approach can not only enhance hydrometeor classification accuracy, but also potentially improve the microphysics parameterization schemes in the numerical weather prediction models.