Abstract
Flood frequency analysis (FFA) is essential for hydrological risk assessment and infrastructure planning. However, traditional methods often assume stationarity, a premise increasingly challenged by climate change and human activities. This study explores nonstationary FFA in the Kosi River Basin using three approaches: Maximum Likelihood estimation, two-stage regression modeling, and Generalized Additive Models for Location, Scale, and Shape (GAMLSS). Daily flow records were used to extract annual maximum discharges, which were fitted to time-varying Generalized Extreme Value (GEV) models. Results show clear nonstationarity, with rising flood quantiles, especially for the 50- and 100-year return periods. While the Maximum Likelihood and Two-Stage approaches captured linear trends, GAMLSS revealed nonlinear dynamics. Model comparisons via Akaike Information Criterion (AIC) indicated no single method was best overall; multimodel averaging weighted by AIC provided more reliable quantile estimates. Bootstrap resampling confirmed increasing uncertainty with longer return periods and consistently highlighted growing flood risks. Stationary models tended to overestimate current design floods by about 35–40%, risking over-design if stationarity is wrongly assumed. This research illustrates that combining multimodal averaging with bootstrap uncertainty offers a robust framework for nonstationary flood frequency analysis, aiding climate-resilient water management and flood risk planning.