| dc.description.abstract | The Soil and Water Assessment Tool (SWAT) and the Artificial Neural Network (ANN) have been widely used as rainfall-runoff models since the 1990s. The former is a more complex, physically-based model that went through decades of continuous development, while the latter is a simpler data-driven model that focuses on establishing the nonlinear relationship between predictors and targets without considering the physical aspects of hydrological systems. Although both SWAT and ANN are broadly accepted as capable of making successful streamflow estimations, their performance capability had not been adequately compared under various conditions in the past. This dissertation seeks to create watershed level rainfall-runoff models using SWAT and ANN across a range of settings and evaluate their performance capability.
In ANN rainfall-runoff modeling, the three-layered feed-forward neural network is regularly used. Routinely, several neural networks are trained before a model selection process selects the network with the best predictive capability. In study I, two common model selection approaches, including the in-sample approach that is based on Akaike’s information criterion (AIC) and Bayesian information criterion (BIC), and the out-of-sample approach that uses blocked cross-validation (BlockedCV), were compared. The results suggested that the BlockedCV is preferable for selecting the rainfall-runoff model with the best predictive capability.
Study II directly compared the SWAT and ANN models’ streamflow predictive performance in two small watersheds in the karstified region of San Antonio, Texas. The paired watershed approach was employed, with one study watershed being highly urbanized and the other primarily covered with evergreen forest and shrub. In addition, the study used the correction factor approach to adjust the goodness-of-fit indicators to incorporate measurement and model uncertainty in the rainfall-runoff modeling process. The results showed that ANN slightly outperformed SWAT in the urban watershed and performed significantly better in the rural watershed. Therefore, suggesting that ANN is a better real-time simulator of streamflow.
Additionally, as gridded precipitation datasets are gaining popularity as a convenient alternative for hydrological modeling during recent decades, Study III evaluated three gridded precipitation datasets, the Tropical Rainfall Measuring Mission (TRMM), the Climate Forecast System Reanalysis (CFSR), and the Parameter-elevation Relationships on Independent Slopes Model (PRISM), against the conventional gauge rainfall observations, and further assessed their capability of driving hydrological simulations in SWAT and ANN. The results of Study III showed that SWAT and ANN simulation outcomes varied in an identical pattern when different precipitation data were applied. Moreover, the PRISM and TRMM driven models were found to have preferable streamflow prediction results than the CFSR and gauge driven models, with the PRISM data produced the best hydrological simulation outcome. | |
