Análise espacial da precipitação extrema no sul e sudoeste de Minas Gerais

Abstract

The occurrence of extreme rainfall events can cause significant damage to urban infrastructure, the environment, and human activities in general. Thus, understanding the behavior of this phenomenon in a region can assist in the planning of activities subject to such damages. Therefore, this study aimed to spatially model the maximum rainfall in the southern and southwestern regions of Minas Gerais, Brazil, using two approaches: Bayesian inference combined with Kriging and Inverse Distance Weighting (IDW), and Max-Stable Processes (MSP) and Spatial Generalized Extreme Value (GEV) models. Daily rainfall data from 29 cities in the region were used for the study. For the IDW analysis, Ordinary Kriging (OK) and Log-Normal Kriging (LNK) predictions were employed, obtained through Bayesian inference for each location and return periods (RPs) of 2, 5, and 10 years. The predictions were obtained using the best prior structure (non-informative and informative) for each municipality. For the Kriging methods, the best semivariogram model was evaluated among Gaussian, Spherical, Exponential, and Wave models. Model evaluation was performed using cross-validation and the mean prediction error (MPE). The evaluation results showed that for the spatial prediction at the highest return period, the most suitable model was OK with the Wave semivariogram. Consequently, this model was used to obtain the prediction maps for the 50- and 100-year RPs. For the MSP analysis, the Smith model and the Schlather model with Bessel, Cauchy, Powered Exponential, and Whittle-Matérn correlation functions were employed. In the Max-Stable and GEV spatial models, trend surfaces for the GEV parameters were used. The analysis of the spatial dependence of extremes was conducted using the Extremal Coefficient, which indicated evidence of low spatial dependence for the data. The models were evaluated using the Takeuchi Information Criterion and the calculation of the MPE. The results showed similarity between the models; however, the Smith model proved to be the most suitable. Therefore, this model was selected to obtain the prediction maps for the 50- and 100-year RPs