A Data-Driven Approach for Automatic Classification of Extreme Precipitation Events: Preliminary Results

Abstract

Even though there exists no universal definition, in the South America Andean Region, extreme precipitation events can be referred to the period of time in which standard thresholds of precipitation are abruptly exceeded. Therefore, their timely forecasting is of great interest for decision makers from many fields, such as: urban planning entities, water researchers and in general, climate related institutions. In this paper, a data-driven study is performed to classify and anticipate extreme precipitation events through hydroclimate features. Since the analysis of precipitation-events-related time series involves complex patterns, input data requires undergoing both pre-processing steps and feature selection methods, in order to achieve a high performance at the data classification stage itself. In this sense, in this study, both individual Principal Component Analysis (PCA) and Regresional Relief (RR) as well as a cascade approach mixing both are considered. Subsequently, the classification is performed by a Support-Vector-Machine-based classifier (SVM). Results reflect the suitability of an approach involving feature selection and classification for precipitation events detection purposes. A remarkable result is the fact that a reduced dataset obtained by applying RR mixed with PCA discriminates better than RR alone but does not significantly hence the SVM rate at two- and three-class problems as done by PCA itself.