Precipitation Type Classification of Micro Rain Radar Data Using an Improved Doppler Spectral Processing Methodology

This paper describes a methodology for processing spectral raw data from Micro Rain Radar (MRR), a K-band vertically pointing Doppler radar designed to observe precipitation profiles. The objective is to provide a set of radar integral parameters and derived variables, including a precipitation type classification. The methodology first includes an improved noise level determination, peak signal detection and Doppler dealiasing, allowing us to consider the upward movements of precipitation particles. A second step computes for each of the height bin radar moments, such as equivalent reflectivity (Ze), average Doppler vertical speed (W), spectral width (σ), the skewness and kurtosis. A third step performs a precipitation type classification for each bin height, considering snow, drizzle, rain, hail, and mixed (rain and snow or graupel). For liquid precipitation types, additional variables are computed, such as liquid water content (LWC), rain rate (RR), or gamma distribution parameters, such as the liquid water content normalized intercept (Nw) or the mean mass-weighted raindrop diameter (Dm) to classify stratiform or convective rainfall regimes. The methodology is applied to data recorded at the Eastern Pyrenees mountains (NE Spain), first with a detailed case study where results are compared with different instruments and, finally, with a 32-day analysis where the hydrometeor classification is compared with co-located Parsivel disdrometer precipitation-type present weather observations. The hydrometeor classification is evaluated with contingency table scores, including Probability of Detection (POD), False Alarm Rate (FAR), and Odds Ratio Skill Score (ORSS). The results indicate a very good capacity of Method3 to distinguish rainfall and snow (PODs equal or greater than 0.97), satisfactory results for mixed and drizzle (PODs of 0.79 and 0.69) and acceptable for a reduced number of hail cases (0.55), with relatively low rate of false alarms and good skill compared to random chance in all cases (FAR < 0.30, ORSS > 0.70). The methodology is available as a Python language program called RaProM at the public github repository.