Abstract
Excessive bacteria levels are the leading cause of impairment in U.S. water bodies. This dissertation looked at the use of watershed-scale computer models to predict in-stream bacteria concentrations. The study site was the Little River Experimental Watershed (LREW) in Tifton, GA, and fecal coliform fate and transport models were built for four of the LREW catchments over the period Jan 1996 - Dec 2002. A multi-model approach was used in the study to examine the current capacity of industry-standard models and to avoid conclusions unique to a particular catchment. Three models were examined: HSPF, SWAT, and a new model based on the principles of hydrograph separation called the Characteristic Concentration (CC) model. Sensitive hydrology and water-quality parameters were identified in HSPF and SWAT and a response-surface iterative scheme was used to calibrate the sensitive parameters of the models, while a simpler calibration method was used for the 2-parameter CC model. Model performance, both hydrology and water-quality, was evaluated by the Nash-Sutcliffe statistic. The research was conducted in three substudies: an examination of the three models' performance in modeling bacteria concentrations in all four catchments, an examination in HSPF and SWAT to determine the relationship between model performance and the hydrologic state of the watershed, and an examination of model combination possibilities to further analyze model performance and provide methods for combining model output to maximize model results beyond what the individual models could achieve. The research revealed that while the hydrology components of HSPF and SWAT could be considered strong the water-quality components were not as strong. The model parameters describing in-stream bacteria processes were consistently more sensitive than parameters describing terrestrial processes of bacteria, a result that was reinforced when considering the hydrologic state of the watersheds. A Latin Hypercube analysis revealed that parameter uncertainty is significant in the models, but that structural uncertainty resulting from the model process equations is the dominant source of uncertainty in model predictions. The model combination methods were able to provide an improved set of model predictions and showed that in some cases the use of a single calibrated model may still not be the best representation of a watershed. In all cases the CC model performed comparably or better than HSPF and SWAT, and provided a new model framework for analyzing the environmental fate and transport of bacteria. The CC model is worth using in future modeling studies and may be particularly useful in model combination applications since it is comparatively much simpler to use and less data-intensive than either HSPF or SWAT.