Nonstationary Frequency Analysis of Annual Maximum Flow Series: Climate Change Versus Land Use/Land Cover Change

Abstract Accurate estimation of flood peak discharges is essential for effective flood risk management. Traditional flood management approaches that assume stationarity may be inadequate due to the dynamic influences of climate change (CC) and land use/land cover (LULC) changes, necessitating nonstationary flood frequency analysis. This study investigates the challenges associated with selecting suitable probability distributions and covariates for nonstationary frequency analysis of annual maximum flow series (AMFS). Six probability distributions (i.e., normal, log-normal, logistic, gamma, Gumbel, and Generalized Extreme Value (GEV)) and seven covariates (i.e., time, reservoir index (RI), annual maximum precipitation, annual average temperature, annual total precipitation, population, and the Normalized Difference Vegetation Index (NDVI)) are evaluated for the Silifke region in Turkey. Based on Akaike Information Criterion (AIC) scores, the GEV distribution performed better than others. However, the 500-year return period flood (Q 500 ) estimated using the best-performing nonstationary GEV model was an outlier, deviating by nearly 19 standard deviations away from the maximum observed discharge. This indicates that reliance solely on AIC scores may result in misleading conclusions. Considering both AIC scores and Q 500 estimates, Gumbel is identified as the best distribution for modeling AMFS in the study area. Results also show that LULC-related covariates more strongly influence nonstationarity than CC-related ones. Among these, RI proves to be the most representative covariate. Notably, NDVI, employed as a covariate in nonstationary frequency analysis for the first time in this study, appears frequently among the most effective models, underscoring its potential relevance in future research.