Statistical Modeling Approaches for Nonstationary Flood Frequency Analysis in the Kosi River Basin

Abstract

Flood frequency analysis (FFA) is a critical tool for hydrological risk assessment and infrastructure design, yet traditional approaches often assume stationarity, an assumption increasingly challenged by climate variability and anthropogenic change. This study investigates nonstationary FFA in the Koi River Basin by applying three approaches: Maximum Likelihood (ML) estimation, Two-Stage regression-based modeling, and Generalized Additive Models for Location, Scale, and Shape (GAMLSS). The annual maximum discharge series was extracted from daily flow records and fitted to time-varying Generalized Extreme Value (GEV) models. Results indicate clear evidence of nonstationarity, with upward trends in flood quantiles, particularly for 50- and 100-year return periods. While ML and Two-Stage approaches captured linear changes in flood behaviour, Generalized Additive Models for Location, Scale and Shape (GAMLSS) revealed nonlinear dynamics. Model comparisons based on Akaike Information Criterion (AIC) showed no single method was universally superior, and AIC-weighted multimodel averaging produced more stable quantile estimates. Bootstrap resampling confirmed the widening of uncertainty bands with increasing return periods, but consistently indicated an intensification of flood hazard. Stationary models were found to underestimate present-day design floods by up to 40 %, implying significant risks of under design. The study demonstrates that integrating multimodal averaging with bootstrap uncertainty provides a robust framework for nonstationary FFA, supporting climate-resilient water management and flood risk planning.