Assessing the effects of multivariate functional outlier identification and sample robustification on identifying critical PM2.5 air pollution episodes in Medellín, Colombia

Abstract

Identification of critical episodes of environmental pollution, both as a outlier identification problem and as a classification problem, is a usual application of multivariate functional data analysis. This article addresses the effects of robustifying multivariate functional samples on the identification of critical pollution episodes in Medellín, Colombia. To do so, it compares 18 depth-based outlier identification methods and highlights the best options in terms of precision through simulation. It then applies the two methods with the best performance to robustify a real dataset of air pollution (PM2.5 concentration) in the Metropolitan Area of Medellín, Colombia and compares the effects of robustifying the samples on the accuracy of supervised classification through the multivariate functional DD-classifier. Our results show that 10 out of 20 methods revised perform better in at least one kind outliers. Nevertheless, no clear positive effects of robustification were identified with the real dataset.