Decreasing flood hazard evaluated in Turkey using nonstationary models

• Published in: River Research and Applications
• Main Authors: Tosunoglu, F, Slater, L
• Format: Article in Journal/Newspaper
• Language: English
• Published: Wiley 2022
• Subjects: Kendall; North Atlantic
• Online Access: https://doi.org/10.1002/rra.3998; https://ora.ox.ac.uk/objects/uuid:6f76733c-507e-42bb-bab9-825179deeec1

An assessment of changing flood hazard in the river basins of Turkey is currently lacking. This study evaluates the drivers of flood flows using distributional regression models. The generalized additive models for location, scale, and shape (GAMLSS), a flexible framework well-suited to probabilistic nonstationary modeling, is applied to annual instantaneous maximum flows (AIMF) from 12 gauging stations located in the upper, middle and lower parts of the Euphrates basin, the most important river basin in the Middle East. First, the Pettitt test was used to detect abrupt changes in the AIMF data series, and the Mann–Kendall test to evaluate temporal trends. Significant change points for two stations and significant decreasing temporal changes for seven stations were observed. Three strictly positive, continuous distributions (namely, Gamma, Lognormal, and Weibull) were then fitted to the AIMF data series and the distribution parameters were specified as functions of both time and physically based covariates (precipitation, temperature, and climate oscillation indices). Among the candidate distributions, the Lognormal was found to be the most appropriate for describing flood flows across the basin. Annual precipitation, seasonal temperature and annual oscillation indices were the most relevant covariates. Among the climate indices, the East Atlantic Oscillation was most relevant in the lower part of the basin, the Southern and North Atlantic Oscillations in the middle, and the East Atlantic-West Russia pattern in the upper part. The models were then employed to estimate and compare historical stationary flood quantiles with nonstationary quantiles (identified with physically based covariates) under various return periods (10, 20, 50, 100, and 200 years). Results indicate that flood processes can be better described by employing distributional regression models with physical covariates and this study can serve as a reference for regional water resources management in the basin and possibly for other river basins.