Toward a combined Bayesian frameworks to quantify parameter uncertainty in a large mountainous catchment with high spatial variability

Although hydrological models play an essential role in managing water resources, quantifying different sources of uncertainties is a challenging task. In this study, the application of two parameter uncertainty quantification methods and their performances for predicting runoff was investigated. Sequential Uncertainty Fitting version 2 (SUFI-2) and DiffeRential Evolution Adaptive Metropolis (DREAM-ZS) algorithms were employed to explore the output uncertainty of Soil and Water Assessment Tool (SWAT) at a multisite flow gauging station. In order to optimize the model and quantify the parameter uncertainty, S1 and S2 strategies, which belong to the SUFI-2 and DREAM-ZS algorithms, were defined. The prior ranges of the S1 were adopted from SWAT manual, and the prior ranges of the S2 were selected using a compromising approach between the prior and posterior ranges extracted from S1. P-factor, d-factor, Nash-Sutcliffe coefficient (NS), the dimensionless variant of average deviation amplitude (S), and the average relative deviation amplitude (T), as performance criteria, were assessed. The NS, S, and T for total uncertainty ranged 0.60-0.71, 0.46-0.51, and 0.94-1.01 under S1 strategy and 0.64-0.78, 0.07-0.22, and 0.39-0.64 under S2, respectively. In parameter uncertainty analysis, S and T indices ranged from 1.51 to 1.88 and 2.20 to 2.60, correspondingly. The results showed that the DREAM-ZS algorithm improved model calibration efficiency and led to more realistic values of the parameters for runoff simulation in SWAT model. However, the S2 strategy, which implicitly takes advantage of both formal and informal Bayesian approaches simultaneously, will be able to outperform the S1 for reducing the prediction uncertainties.