Assessment of non-linear models based on regional frequency analysis for estimation of hydrological quantiles at ungauged sites in South Korea

The study region is comprised of 33 river networks, including hydrometric stations in South Korea. The estimation of low-flow quantiles for ungauged sites has become increasingly important in the hydrological management of water resources and environmental preservation. Although hydrological processes exhibit nonlinear behaviors and complexity, low-flow quantile estimates in South Korea are typically derived using linear models, such as the drainage area ratio. This study explores the application of nonlinear models, particularly the ensemble artificial neural network (EANN) and generalized additive model (GAM), to determine whether the model is more effective in calculating low-flow quantiles. Canonical correlation analysis is employed to enhance the accuracy and performance of regional frequency analysis. Thirty-three river networks in South Korea are used to estimate 2-year and 5-year low-flow quantiles for ungauged sites. Further, the jackknife resampling technique is used to validate the results of the proposed models based on statistical indices, including the Nash–Sutcliffe efficiency index, relative root mean squared error, and relative mean bias. The results indicate that the GAM outperforms the EANN model in estimating the low-flow quantiles for river networks in South Korea. This suggests that selecting an appropriate model can improve the accuracy and performance of regional estimates, offering significant insights into water resources and environmental management while accounting for hydrological nonlinearity. By enhancing the low-flow quantile estimates for ungauged sites, we can better manage water resources and aquatic ecosystems, ultimately promoting healthier streams.