Support vector machine (SVM) model development for prediction of fecal coliform of Upper Green River Watershed, Kentucky, USA

The classification and prediction of water quality parameters (WQPs) such as Fecal Coliform in river waters are crucial for developing a Decision Support System or Tool for water quality protection or water resource management. Using Support Vector Machine (SVM) classification and regression, a predictive modeling attempt is made for the Upper Green River Watershed, Kentucky, the U.S.A. The Linear, Polynomial, and Radial Basis Function (RBF) Kernels are used for classification and regression. A sensitivity analysis is performed for SVM models with the help of variants of Gamma and C values to obtain the best predictions of fecal coliform. Further, Least Squares Support Vector Machine (LS-SVM) is also employed to strengthen the accuracy of forecasts of individual input parameters. The results of SVM are compared with Artificial Neural Networks (ANN) for the same watershed. It is found that while the ANN models perform better than linear, polynomial SVM models, the SVM RBF regression models stream water quality (as good as or) slightly better than ANN models for the same inputs. This study obtains coefficients of determination of 0.91, 0.87, and 0.90 using the SVM RBF model in training, testing, and overall, respectively. These coefficients are 0.82, 0.90, and 0.85 using feed-forward ANNs for fecal coliform in training, testing, and overall. The results of LS-SVM indicate that the climate parameters are more crucial for water quality modeling than land use parameters.