Statistical detection of anomalous propagation in radar reflectivity patterns

Abstract

A classical radar (MRL-5) station operates in central Poland performing volume scanning every 10 min. Two months of hourly data were chosen to create learning and verification samples for the AP detection algorithm. Each observation was thoroughly analyzed by an experienced radar meteorologist. The features taken into account were visually estimated local texture and overall morphology of echo pattern, vertical echo structure, time evolution (using animation), and the general synoptic information. For each 4 km × 4 km pixel of 933 observations the human classification was recorded resulting in a sample of 631 166 points with recognized echo type, 14.6% of them being AP echoes. The unsuppressed AP echo impact on monthly accumulated precipitation was 59% of the actual sum for the month of June and as much as 97% for September. Three Bayesian discrimination functions were investigated. They differ in selection of the feature vector. The AP echo recognition error was about 6% for the best-performing function, when applied to an independent (September) sample, which would make the method acceptable for operational use. -from Authors