Quantifying predictive uncertainty for a mountain-watershed model

Watershed models require calibration before they are utilized as a decision-making tool. This paper describes a rigorous sensitivity analysis, automated parameter estimation and evaluation of prediction uncertainty for a Watershed Analysis Risk Management Framework (WARMF) model of the Turkey Creek Watershed. Sensitivity analysis was conducted using UCODE calibration and uncertainty-analysis software. Simulated stream flow is strongly sensitive to 7 of the 20 parameters evaluated: hydraulic conductivity, field capacity, total porosity, precipitation weighting factor, evaporation magnitude, evaporation skewness and snow melting rates; and parameter sensitivity is dependent on site-specific climate and soil conditions. Simulated stream flow matched observed stream flow fairly well with an R(2) value of 0.85, Nash-Sutcliffe coefficient of efficiency (NSE) value of 0.75 and Root Mean Squared Error (RMSE) of 0.23 m(3)/s. The calibrated model was used to predict changes in stream flow that would result from changes in land use, including development of forested areas in parts of the watershed to commercial and residential areas. As expected, new development resulted in increased peak flows and reduced low flows. Uncertainty associated with all model parameters, including those not estimated by calibration by enhancing the parameter variance/covariance matrix, was considered when evaluating prediction uncertainties. Seventy percent of the time, predicted flows had uncertainties less than 20% with more of the uncertainty during low flow conditions. (C) 2009 Elsevier B.V. All rights reserved.