Predicting water quality in unmonitored watersheds using artificial neural networks

Land use and land cover (LULC) play a central role in fate and transport of water quality (WQ) parameters in watersheds. Developing relationships between LULC and WQ parameters is essential for evaluating the quality of water resources. In this paper, we present an artificial neural network (ANN)-based methodology to predict WQ parameters in watersheds with no prior WQ data. The model relies on LULC percentages, temperature, and stream discharge as inputs. The approach is applied to 18 watersheds in west Georgia, United States, having a LULC gradient and varying in size from 2.96 to 26.59 km2. Out of 18 watersheds, 12 were used for training, 3 for validation, and 3 for testing the ANN model. The WQ parameters tested are total dissolved solids (TDS), total suspended solids (TSS), chlorine (Cl), nitrate (NO3), sulfate (SO4), sodium (Na), potassium (K), total phosphorus (TP), and dissolved organic carbon (DOC). Model performances are evaluated on the basis of a performance rating system whereby performances are categorized as unsatisfactory, satisfactory, good, or very good. Overall, the ANN models developed using the training data performed quite well in the independent test watersheds. Based on the rating system TDS, Cl, NO3, SO4, Na, K, and DOC had a performance of at least "good" in all three test watersheds. The average performance for TSS and TP in the three test watersheds were "good." Overall the model performed better in the pastoral and forested watersheds with an average rating of "very good." The average model performance at the urban watershed was "good." This study showed that if WQ and LULC data are available from multiple watersheds in an area with relatively similar physiographic properties, then one can successfully predict the impact of LULC changes on WQ in any nearby watershed.