Abstract
This dissertation utilizes data from four sub-watersheds in the Little River Experimental Watershed, GA to develop models to improve forecast predictions related to the management of surface-water pollution due to non-point source runoff. Non-point source pollution is the primary cause of US surface-water quality impairment and a main transport mechanism for pathogens and other pollutants into receiving surface water bodies (US EPA 2008). In response to pollution reduction and watershed remediation mandates under the Federal Clean Water Act (1972)-particularly the Total Maximum Daily Load (TMDL) program-the role of water quality modeling in effectively rehabilitating impaired waters has taken on greater importance. Consequently, the significance of this study is that it is the first of its kind to incorporate a multi-model approach to address limitations in using single water quality models. In this regard, it builds on water quality engineering research by presenting methods to estimate contaminant concentrations and reduce uncertainty in overall model predictions in impaired water-bodies. Methodologically, the key point of departure in this dissertation is centered on the fact that water quality modeling is the cornerstone of TMDL analyses but the associated prediction uncertainty affects their adequacy in providing reliable contaminant loadings estimates in an impaired water body. As such, utilizing hydrological and water-quality process equations embedded in the two most widely used watershed-scale models, the Soil and Water Assessment Tool (SWAT) and Hydrological Simulation Program-Fortran (HSPF), and observed data from the sub-watersheds mentioned above, the dissertation addresses this limitation by combining results from the two competing models to reduce uncertainty and enhance accuracy of predictions. The study was conducted in two phases. First, HSPF and SWAT-two extensively-used, scientifically-rigorous, US EPA-approved watershed-scale codes-were used to build models of the four study catchments. The models were individually calibrated and shown (based on Nash-Sutcliffe Efficiency (NSE) ratios) to produce reliable simulations of the hydrologic and water quality conditions in the watershed. The second phase of the analysis involved using a multi-model approach to combine model forecasts. Model combination, introduced by Bates and Granger in 1969, has emerged as a viable analytical technique (Claesken and Hjort, 2008; Ajami et al., 2006) and widely-used across disciplines to improve model-forecasting results (Kim et al., 2006; Shamseldin et al., 1997; Granger, 2001; Clemens, 1989; Thompson, 1976; Newbold and Granger, 1974; Dickinson, 1973). After calibration, the model predictions were combined for each catchment using three different methods: the Weighted Average Method (WAM), the Nash-Sutcliffe Efficiency Maximization Method (NSE-max) and an Artificial Neural Network Method (ANN). Comparison of the results of the multi-model formulation with original individual model results showed improved estimates with all three combination methods. The improvement in model accuracy (based on NSE ratios) varied from modest to significant in both hydrologic and water quality variables. These improvements were attributed to a reduction in model structural uncertainty resulting from the ability to capture aspects of some of the more complex watershed interactions from exogenous information provided by the contributing models. It should be noted here, however, that as model availability increases, if additional models (beyond those utilized here) are used with this approach, care should be taken to ensure the credibility of each individual model for simulating the watershed scale processes under review. Limitations of this study include possible bias introduced by the use of deterministic models to estimate probabilistic contaminant distributions, limitations in available data, and the use of a seven-year study period that did not account for possible impacts of shorter periods of extreme hydrologic conditions on the individual model performances and model combination weightings. Recommendations for future research include (a) improving watershed-scale codes to better describe the probability distribution functions characteristic of contaminant distributions and data collection on wildlife species and populations; and investigating the fate and transport processes of pathogenic indicator bacteria deposited in forested areas and the impact of extreme hydrologic conditions on model performance and weighting. Overall, the findings from this dissertation suggest that water quality modeling incorporating a multi-model approach has the potential to significantly improve predictions compared to the predictions obtained when only one model is used. Clearly, the findings reported here have significant implications in improving TMDL analyses and remediation plans by presenting an approach that exploits the strengths of two of the most complete and well-accepted watershed-scale water quality models in the United States. Moreover, the findings of this dissertation auger well for the future of TMDL management in that it provides a more robust and cost effective basis for policy makers to decide on effective management strategies that incorporate acceptable risk, allowable loading and land use.