Uncertainty Quantification of a Distributed Hydrology Model Using Bayesian Framework Across a Heterogeneous Watershed

Abstract:

Coastal Plain watersheds have high levels of spatial heterogeneity due to rainfall variability in space and time, dominate shallow water table, nonlinearity of river hydraulics properties, wide floodplains and dense vegetation. Bayesian algorithms can be used to capture key hydrological dynamics across the heterogeneous watershed system. This study examined two Bayesian frameworks to quantify parameter uncertainty during 2003-2005 period in the distributed hydrologic model (i.e. Soil and Water Assessment Tool (SWAT)) using Sequential Uncertainty Fitting (SUFI-2) and Differential Evolution Adaptive Metropolis (DREAM) algorithms in the Black River watershed, in the southeastern (SE) United States. Both algorithms were calibrated using 19 absolute parameter ranges of the SWAT model and streamflow predictive uncertainty showed well agreement to physical variation and system dynamics across the dry to moderately wet calibration period. In this study, the calibrated p-factor computed using SUFI-2 and DREAM algorithms respectively bracketed 63% and 69% of the predictive uncertainty, underlying the importance of parameter uncertainty in the watershed under study. In addition, SWAT parameters exhibited significant seasonal variation in dry and wet hydrological conditions and both Bayesian algorithms demonstrated that groundwater parameters, and soil and land use properties contributed more uncertainty to model. To reduce the calibration load, a hypothesis was further tested about whether convergence of DREAM algorithm can be achieved quicker by incorporating the best parameter ranges of the SUFI-2 model. The results revealed that the convergence criterion of R <= 1.2 in DREAM was met after about 21,000 simulation runs using SUFI-2 parameter ranges while absolute parameter ranges met the convergence criterion after approximately 77,000 model simulations. In principle the methodology proposed here led to some improvement in parameter indentification, diminished the dimensionality of the parameter space and reduced burn-in period, indicating that DREAM can be utilized to identify optimal parameter values in addition to quantifying parameter and predictive uncertainty.