Comparison of GLUE Weighting Methods for Uncertainty Estimation in Hydrologic Modeling

ABSTRACT The generalized likelihood uncertainty estimation (GLUE) framework has widely been used to quantify uncertainty in hydrologic modeling. Subjective factors such as likelihood function and threshold values affect the performance of this framework. This study investigated the impacts of likelihood measure with and without nonconventional weighting methods on the uncertainty results of the Genie Rural (GR) model applied at a catchment in the northeastern area of Iran. We used two likelihood measures including the Nash‐Sutcliffe efficiency (L NSE ) and flow duration curve (L FDC ), and three weighting methods including the conventional GLUE, rank reciprocal (RR) and multi‐objective optimization called the non‐dominated sorting genetic algorithm II (NSGAII), methods. Moreover, the impact of threshold values on GLUE results was evaluated. The same behavioral simulations selected using both likelihood measure ranged from 4% to 14.5% for the highest and lowest thresholds, respectively. For the likelihood measures, the groundwater exchange coefficient and groundwater exchange threshold parameters were found to have the least uncertainty. For most parameters, as the acceptable sample rate increased, the uncertainty of the parameters increased as well. For the conventional GLUE, RR, and NSGAII weighting methods, the L FDC likelihood measure provided wider prediction uncertainty bounds that bracketed more observations compared to L NSE . For both likelihood measures, RR weighting provided narrower width while NSGAII provided those that are more reliable. To the best of the authors’ knowledge, we used L FDC likelihood measure and NSGAII weighting method within the GLUE framework for the first time in this study to quantify uncertainty of the GR model.