Toward Avoiding the "Right Answer for the Wrong Reason": Probabilistic Appraisal of Competing Model Structures in Distributed Hydrologic Models

The performance of watershed models is often evaluated based on a deterministic approach, i.e. calibrate → validate → predict, where calibration is conducted to obtain a parameter set that provides the best fit between model responses and observations at the outlet of the watershed. Such assessments can be inadequate, and in many cases misleading in the case of distributed hydrologic models, particularly when the model is used to simulate interior hydrologic processes or to assess responses at various locations within the watershed. Probabilistic approaches can address the equifinality and nonuniqueness issues in parameter estimation, input uncertainty, and measurement errors when assessing competing model structures. A probabilistic approach is presented to assess the performance validity of the empirical Curve Number (CN) and physically-based Green and Ampt (G&A) rainfall-runoff methods in the SWAT model. Specifically, the effects of modeling uncertainties on characterization of the hydrologic budgets and streamflow regimes at various spatial scales and upstream land use conditions are investigated. While the models were trained for streamflow estimation only at the watershed outlet, the performances of the models were compared at different stream locations within the watershed. At the watershed outlet, the CN method had a slightly better, but not significant, performance in terms of streamflow error statistics. Similar results were obtained for the predominantly forested and agricultural tributaries. However, in tributaries with higher percentage of developed land, G&A outperformed the CN method in simulating streamflow based on various performance metrics. In general, the 95% prediction intervals from the models with G&A method covered a higher percentage of observed streamflow especially during the high flow events. Using 95% prediction interval for estimated flow duration curves, the CN-based models underestimated high flow events especially in tributaries with highly developed land use while generating higher water yields to streams. The results of this study have important implications for application of distributed hydrologic models in mixed-land use watersheds.