Decision Tree Ensemble Classifiers for Anomalous Propagation Echo Detection

摘要

Several types of non-precipitation echoes such as permanent, spurious, and anomalous propagation significantly disturb weather radar observation process. Specifically, the anomalous propagation echo is one of main issues due to its similar characteristics compared with precipitation. It occurs by refracted radar beam, and makes irregular shape echo with random size and reflectivity. For solving problems caused by the anomalous propagation echo, studies and researches have been conducted for years in various fields such as quantitative precipitation estimation, quality control, operational hydrology, and so forth. In order to implement a reliable automatic detection system of anomalous propagation echo, we compared decision tree ensemble classifiers for finding the most efficient classifier to detect the anomalous propagation echo. We carefully select representative ensemble classifiers such as Breiman's random forest, extremely randomized trees, adaptive boosting, gradient boosting, and CART algorithm. By comparing these classifiers, it is confirmed that boosting can provide higher accuracy than others, while ensemble methods are superior to the CART algorithm.