Development of an artificial neural network for hydrologic and water quality modeling of agricultural watersheds

Agriculture is the leading source of nonpoint-source pollution on a national scale. The driving force of nonpoint-source pollution is the rainfall-runoff process, which is the transformation of rainfall to streamflow. This is a complex, nonlinear, time-varying, and spatially distributed process on the watershed scale that is difficult-to effectively model by conventional, deterministic means. Artificial neural networks (ANNs) offer a new approach to forecasting the hydrologic and water quality response of a watershed system. The goal of this work is to develop an ANN model as a long-term forecasting tool for predicting the hydrology and water quality of agricultural watersheds where the physical processes are difficult to model using traditional hydrologic/water quality models. The chosen form of neural network is a flexible mathematical structure, which is capable of identifying complex nonlinear relationships between input and output data sets. In this article, a multi-layer feed forward ANN model was developed and tested using historical daily rainfall, streamflow, and nitrate data from the Vermilion River in Illinois, a watershed with intensive subsurface drainage and historically high nitrate concentrations. The ANN was applied to predict daily streamflow and nitrate load based on rainfall. The results show highly accurate performance of the ANN model (r(2) values > 0.80) in predicting daily streamflow and nitrate loads.