Modeling long-term monthly rainfall variability in selected provinces of South Africa using extreme value distributions

Abstract

Several studies indicated a growing trend in terms of frequency and severity of extreme events. Extreme rainfall could cause disasters that lead to loss of property and life. The aim of the study was to model the monthly rainfall variability in selected provinces of South Africa using extreme value distributions. This study investigated the best-fit probability distributions in the five provinces of South Africa. Five probability distributions: gamma, Gumbel, log-normal, Pareto and Weibull, were fitted and the best was selected from the five distributions for each province. Parameters of these distributions were estimated by the method of maximum likelihood estimators. Based on the Akaike information criteria (AIC) and Bayesian information criteria (BIC), the Weibull distribution was found to be the best-fit probability distribution for Eastern Cape, KwaZulu-Natal, Limpopo and Mpumalanga, while in Gauteng the best-fit probability distribution was found to be the gamma distribution. Monthly rainfall trends detected using the Mann–Kendall test revealed significant monotonic decreasing long-term trend for Eastern Cape, Gauteng and KwaZulu-Natal, and insignificant monotonic decreasing longterm trends for Limpopo and Mpumalanga. Non-stationary generalised extreme value distribution (GEVD) and non-stationary generalized Pareto distribution (GPD) were applied to model monthly rainfall data. The deviance statistic and likelihood ratio test (LRT) were used to select the most appropriate model. Model fitting supported stationary GEVD model for Eastern Cape, Gauteng and KwaZulu-Natal. On the other hand, model fitting supported non-stationary GEVD models for maximum monthly rainfall with nonlinear quadratic trend in the location parameter and a linear trend in the scale parameter for Limpopo, while in Mpumalanga the non-stationary GEVD model, which has a nonlinear quadratic trend in the scale parameter and no variation in the location parameter fitted well to the maximum monthly rainfall data. Results from the non-stationary GPD models showed that inclusion of the time covariate in our models was not significant for Eastern Cape, hence the bestfit model was the stationary GPD model. Furthermore, the non-stationary GPD model with a linear trend in the scale parameter provided the best-fit for KwaZulu-Natal and Mpumalanga, while in Gauteng and Limpopo the nonstationary GPD model with a nonlinear quadratic trend in the scale parameter fitted well to the monthly rainfall data. Lastly, GPD with time-varying thresholds was applied to model monthly rainfall excesses, where a penalised regression cubic smoothing spline was used as a time-varying threshold and the GPD model was fitted to cluster maxima. The estimate of the shape parameter showed that the Weibull family of distributions is appropriate in modelling the upper tail of the distribution for Limpopo and Mpumalanga, while for Eastern Cape, Gauteng and KwaZulu-Natal, the exponential family of distributions was found to be appropriate in modelling the upper tail of the distribution. The dissertation contributes positively to the body of knowledge in extreme value theory application to rainfall data and makes recommendations to the government agencies on the long-term rainfall variability and their negative impact on the economy.