Hydrometeor classification using clustering of polarimetric radar measurements

Hydrometeor classification aims at identifying the dominant type of hydrometeors in the region covered by a weather radar. Techniques documented in the literature are mostly based on scattering simulations and fuzzy logic. This involves the arbitrary selection of a set of categorical classes of hydrometeor types and the simulation of their scattering behavior. The information derived from the simulation is then applied to radar measurements by means of Fuzzy logic or as a prior information in Bayesian methods. This approach has some limitations: the number and type of the hydrometeor categories undergoing identification is arbitrarily selected and the scattering simulations can be based on unreliable assumptions, especially in case of solid particles. Furthermore, in presence of noises and uncertainties (unavoidably related to real observations) it is not guaranteed that different and distinct classes of particle types could be effectively and systematically identified.

In the present work we investigate the potential of a different approach to the classification problem, based on observations instead of simulations. The goal is to determine the number of hydrometeor classes that can be reliably identified. Hierarchical unsupervised classification, based on a correlation metrics, is used to group the observations in clusters in the multi-dimensional space of the polarimetric variables. The optimal number of clusters is estimated iteratively as an optimal tradeoff between spatial smoothness of the classified domains, and complexity of the partitions. These clusters are then related to hydrometeor type, compared with the output of known classification algorithms and a supervised classification approach is used to identify the type of hydrometeor in a given polar radar image.

Two polarimetric datasets, collected by an X-band weather radar are employed in the study. The two datasets cover weather conditions ranging from Alpine precipitation collected in the Swiss Alps to Mediterranean orographic events, collected during the special observation period (SOP) 2012 of the HyMeX campaign.