Confronting input, parameter, structural, and measurement uncertainty in multi-site multiple-response watershed modeling using Bayesian inferences

2012 Fall.Includes bibliographical references.Simulation modeling is arguably one of the most powerful scientific tools available to address questions, assess alternatives, and support decision making for environmental management. Watershed models are used to describe and understand hydrologic and water quality responses of land and water systems under prevailing and projected conditions. Since the promulgation of the Clean Water Act of 1972 in the United States, models are increasingly used to evaluate potential impacts of mitigation strategies and support policy instruments for pollution control such as the Total Maximum Daily Load (TMDL) program. Generation, fate, and transport of water and contaminants within watershed systems comprise a highly complex network of interactions. It is difficult, if not impossible, to capture all important processes within a modeling framework. Although critical natural processes and management actions can be resolved at varying spatial and temporal scales, simulation models will always remain an approximation of the real system. As a result, the use of models with limited knowledge of the system and model structure is fraught with uncertainty. Wresting environmental decisions from model applications must consider factors that could conspire against credible model outcomes. The main goal of this study is to develop a novel Bayesian-based computational framework for characterization and incorporation of uncertainties from forcing inputs, model parameters, model structures, and measured responses in the parameter estimation process for multisite multiple-response watershed modeling. Specifically, the following objectives are defined: (i) to evaluate the effectiveness and efficiency of different computational strategies in sampling the model parameter space; (ii) to examine the role of measured responses at various locations in the stream network as well as intra-watershed processes in enhancing the model performance credibility; (iii) to facilitate combining predictions from competing model structures; and (iv) to develop a statistically rigorous procedure for incorporation of errors from input, parameter, structural and measurement sources in the parameter estimation process. The proposed framework was applied for simulating streamflow and total nitrogen at multiple locations within a 248 square kilometer watershed in the Midwestern United States using the Soil and Water Assessment Tool (SWAT). Results underlined the importance of simultaneous treatment of all sources of uncertainty for parameter estimation. In particular, it became evident that incorporation of input uncertainties was critical for determination of model structure for runoff generation and also representation of intra-watershed processes such as denitrification rate and dominant pathways for transport of nitrate within the system. The computational framework developed in this study can be implemented to establish credibility for modeling watershed processes. More importantly, the framework can reveal how collection of data from different responses at different locations within a watershed system of interest would enhance the predictive capability of watershed models by reducing input, parametric, structural, and measurement uncertainties