Date of Graduation


Document Type



Today, nonpoint source (NPS) pollution remains as the leading cause of water quality problems in the United States (USEPA, 1996). The most common NPS pollutants are sediment (i.e. total suspended solids, TSS) and nutrients (i.e. total nitrogen, TN). These wash into water bodies from agricultural land, construction sites, and other areas of disturbance. To develop the virtual non-point source assessment system (VNPSAS) model, monitoring gauges were selected for data extraction from West Virginia and also the western part of Virginia. A geographic information system (GIS) was used as a productive environment for data extraction and spatial analysis. The information for each watershed was extracted from digital databases based on the coordinates of an imaginary grid over the watershed boundary. Different data sets were extracted based on the type of the pollutant and number of the cells in the grid (one, two, and four-cell). The VNPSAS model was developed based on artificial neural network technology. A three-layer backpropagation neural network was used as the main engine for the VNPSAS model. Four sets of training data from each dataset were used to test the VNPSAS model (cross-validation runs). It was found that the VNPSAS model provides relatively accurate estimates of NPS pollution concentration. The average R2, between actual and virtual measurements of TSS for one, two, and four-cell grids were 0.939, 0.910 and 0.408 for verification datasets, and 0.954, 0.993, and 0.888 for the training datasets, respectively. The average R2, between actual and virtual measurements of TN for one and two-cell grids were 0.845 and 0.244 for verification datasets, and 0.929 and 0.875 for the training datasets, respectively. The high values of the R2 for most of the cases demonstrated the high performance of the VNPSAS for NPS pollution assessment. This research showed that by increasing the number of cells in the grid, the accuracy of the predictions declines. It was demonstrated that this behavior is a function of data availability and not the inability of the model to predict the NPS pollution. This research is one of the first attempts made to model the NPS pollution assessment using GIS and artificial intelligence.