Modeling non-stationarity in extreme rainfall data and implications for climate adaptation: A case study from southern highlands region of Tanzania

Abstract

The Southern Highlands region of Tanzania has witnessed an increased frequency of severe flash floods. This study examines rainfall data of four stations (Iringa, Mbeya, Rukwa, and Ruvuma) spanning 30 years (1991–2020) to investigate drivers of extreme rainfall and non-stationarity behavior. The Generalized Extreme Value (GEV) model, commonly used in hydrological studies, assumes constant distribution parameters, which may not be true due to climate variability, potentially leading to bias in extreme quantile estimation. Recent studies have introduced a technique for constructing non-stationary Intensity-Duration-Frequency (IDF) rainfall curves. The method incorporates trends in the parameters of the GEV distribution, only using time as a covariate. However, uncertainty exists about whether time is the most suitable covariate, highlighting the need to explore all potential covariates for modeling non-stationarity. The aim of this study is to assess the influence of other time-varying covariates on extreme daily rainfall events, considering seasonality and climate change in the rainfall data. Specifically, five processes (i.e., local temperature changes (LTC), urbanization, annual Global Temperature Anomaly (GTA), the Indian Ocean Dipole (IOD), and the El Niño-Southern Oscillation (ENSO) cycle) were studied as drivers of extreme rainfall events. Sixty two non-stationary GEV models are developed based on these covariates and their combinations, alongside two non-stationary GEV models using the time covariate to capture the seasonality of the unimodal rainfall in the region, and one stationary GEV model (S0). With the use of corrected Akaike Information Criterion (AICc), the best model for each duration (i.e., 1-, 3-, and 5-days) of rainfall series is chosen. Results indicate that local processes (i.e., LTC and urbanization) are the optimal covariates for 1 day-duration rainfall, while global processes (i.e, IOD, ENSO cycle, and GTA) are identified as the most suitable covariates for 3, and 5 day-duration rainfall across all stations. The identified best non-stationary model (with their best covariates) are then used to develop non-stationary rainfall IDF curves for all stations. According to the analysis of non-stationary extreme values, the return periods of extreme rainfall events concluded a notable decrease in comparison to the stationary approach. The study also revealed strong correlations between global climate indices (ENSO, IOD, GTA) and long-duration extreme rainfall in Tanzania’s Southern Highlands. Local factors like Urbanization and temperature changes also show significant associations with 1-day duration events. These findings emphasize the need for integrated climate forecasting to inform effective adaptation strategies. Finally, the study addresses associated uncertainties in our predictions of forthcoming extreme rainfall events through rigorous analysis. The study demonstrated that return levels for extreme rainfall events exhibit a rising trend with increasing return period, indicating heightened intensity over longer time spans, whereas, a relative uncertainty analysis illustrate escalating uncertainty with increasing return periods, emphasizing challenges in long-term prediction.