A new approach to detect extreme events: a case study using remotely-sensed precipitation time-series data

Abstract

Detecting and predicting extreme events are of major importance for socioeconomic, healthcare and ecological purposes. This study proposes an alternative model to detect extreme events based on analyses of probability distribution functionffnss (f(X)), called Optimum Probability Distribution Function Searcher Model (Opt.PDF-model). The Opt.PDF-model involves the optimization of a fitness function between an histogram and a set of theoretical f(X), and the subsequent evaluation of the Probability Point Function (PPF) of the fittest theoretical (f(X)) to assess threshold values for the classification of extreme events. Any occurrence in the dataset with a PPF value equal to or greater than 90% was considered an extreme event candidate. A satellite-derived precipitation time-series (Climate Hazards Group InfraRed Precipitation with Station data) was used to calibrate and validate the proposed model, with data on accumulated precipitation from more than 30 years (Jan.1981 to Dec.2018) of the Brazilian Amazon region. The proposed method was pairwise cross-validated with two other extreme event models based on Gamma and Gaussian distributions, as applied by the European Drought Observatory of the European Environment Agency. Aditionally, all three extreme event classification models were cross-validated relative to the El Niño Southern Oscillation (ENSO). By means of the Opt.PDF-model, it was possible to evidence two positive temporal trends for the area of study: one for more intense precipitation events, and another for less intense events. The pairwise cross-validation analysis returned specific Kappa’s similarity indices, and the highest similarity was observed between the Gamma and the Opt.PDF models (48% for PPF(97.7%)). This analysis indicated that extreme event detection models are highly sensitive to distribution family priors and to threshold definitions. The proposed approach and the results obtained here are potentially useful for climate change warnings, and can be extended to other scientific areas that involve time-series analyses.