The performance of the Soil and Water Assessment Tool (SWAT) and artificial neural network (ANN) models in simulating hydrologic response was assessed in an agricultural watershed in southeastern Pennsylvania. All of the performance evaluation measures including Nash-Sutcliffe coefficient of efficiency (E) and coefficient of determination (R^sup 2^) suggest that the ANN monthly predictions were closer to the observed flows than the monthly predictions from the SWAT model. More specifically, monthly streamflow E and R^sup 2^ were 0.54 and 0.57, respectively, for the SWAT model calibration period, and 0.71 and 0.75, respectively, for the ANN model training period. For the validation period, these values were -0.17 and 0.34 for the SWAT and 0.43 and 0.45 for the ANN model. SWAT model performance was affected by snowmelt events during winter months and by the model's inability to adequately simulate base flows. Even though this and other studies using ANN models suggest that these models provide a viable alternative approach for hydrologic and water quality modeling, ANN models in their current form are not spatially distributed watershed modeling systems. However, considering the promising performance of the simple ANN model, this study suggests that the ANN approach warrants further development to explicitly address the spatial distribution of hydrologic/water quality processes within watersheds.

Watershed hydrology is of central importance to the structure and function of stream ecosystems. Stochastic by nature, streamflow varies over time in response to precipitation and is inherently subject to episodic extremes of high and low flows. Streamflow also varies among watersheds due to complex physiographic, landscape, and disturbance characteristics. Stream organisms are generally well adapted to a range of streamflow conditions but can be strongly affected by extreme flow events. Extreme high flow events can reduce benthic algae abundance (Biggs and Close, 1989) and disrupt benthic macroinvertebrate communities through scouring and bed instability. Low flows can severely stress fish communities and, in extreme cases, may cause changes in local community composition and density. Given the great importance of hydrology to stream ecosystems, it follows that any understanding of ecosystem structure and function in these systems must explicitly account for the effects of flow variability (Poff et al., 1997; Richter et al., 1997). Further, together with sediment loadings from watersheds, streamflow is also a key variable influencing the morphology of stream channels, which in turn determines the physical habitat -the foundation of any stream ecosystem (Leopold et al., 1964).

Given the great importance of hydrology to stream ecosystems, accurate prediction of streamflow is of utmost importance. However, accurate predictions of rainfall-runoff and consequent streamflows from a regional scale watershed are extremely difficult because of tremendous spatial and temporal variability of watershed characteristics and weather patterns and an incomplete understanding of complex underlying physical processes. The traditional approach to hydrologie modeling is to link several detailed mechanistic, quasi-mechanistic, or empirical submodels of system processes (e.g., evapotranspiration, infiltration, percolation, flow routing) in an effort to build a process-based model of the complete hydrologie system such as Hydrologie Simulation ProgramFORTRAN (HSPF) (Bicknell et al., 2001) and SWAT (Neitsch et al., 2001a). Because of the enormous spatial scale of the system and the complex, poorly understood processes and their interactions, alternatives to traditional process-based modeling approaches must be developed. Further, process-based models are usually difficult to implement because of their high complexity and massive input data requirements, much of which are unavailable for significant portions of any regional scale watershed. Such models also require considerable technical expertise to implement and use and therefore are not suitable for most watershed managers. Moreover, inadequate scientific understanding of underlying processes implies that structural (or knowledge) uncertainty will be substantial in such models, imposing severe constraints on efforts to reduce prediction uncertainty.

In response to these concerns and the demonstrated success of ANNs applied to complex problems in a variety of disciplines (e.g., physics, biomedical engineering, electrical engineering, computer science, robotics, and image processing), several researchers in the hydrologie community are attempting to use ANN as an alternative modeling tool for streamflow predictions. Maier and Dandy (2000) provide a review of neural network models for predicting and forecasting various water resources variables. For example, Karunanithi et al. (1994), Minns and Hall (1996), Braddock et al. (1998), Dawson and Wilby (1998), Anmala et al. (2000), Gupta et al. (2000), Salas et al. (2000), and Zhang and Govindaraju (2000, 2003) used various forms of ANNs for rainfall-runoff modeling. A number of researchers also compared performance of ANNs with other empirical approaches. Anmala et al. (2000) compared the performance of various ANN architectures with empirical approaches such as complex time series and multivariate time series in three Kansas watersheds and found ANNs and recurrent neural networks (RNNs) to perform as well or better. Hsu et al. (1995) found ANN performance to be better than a conceptual Sacramento Soil Moisture Accounting (SAC-SMA) model and an autoregressive moving average with exogenous variables (ARMAX) time series approach in a subbasin of the Mississippi River basin for daily streamflow predictions. Fernando and Jayawardena (1998) found a radial basis function network (a type of ANN) to work better than the ARMAX approach. Tokar and Johnson (1999) found ANNs to be superior to simple conceptual models in the Little Patuxent River basin, Maryland.