Decision-tree-model identification of nitrate pollution activities in groundwater: A combination of a dual isotope approach and chemical ions

To develop management practices for agricultural crops to protect against NO3− contamination in groundwater, dominant pollution activities require reliable classification. In this study, we (1) classified potential NO3− pollution activities via an unsupervised learning algorithm based on δ15N- and δ18O-NO3− and physico-chemical properties of groundwater at 55 sampling locations; and (2) determined which water quality parameters could be used to identify the sources of NO3− contamination via a decision tree model. When a combination of δ15N-, δ18O-NO3− and physico-chemical properties of groundwater was used as an input for the k-means clustering algorithm, it allowed for a reliable clustering of the 55 sampling locations into 4 corresponding agricultural activities: well irrigated agriculture (28 sampling locations), sewage irrigated agriculture (16 sampling locations), a combination of sewage irrigated agriculture, farm and industry (5 sampling locations) and a combination of well irrigated agriculture and farm (6 sampling locations). A decision tree model with 97.5% classification success was developed based on SO42− and Cl− variables. The NO3− and the δ15N- and δ18O-NO3− variables demonstrated limitation in developing a decision tree model as multiple N sources and fractionation processes both resulted in difficulties of discriminating NO3− concentrations and isotopic values. Although only the SO42− and Cl− were selected as important discriminating variables, concentration data alone could not identify the specific NO3− sources responsible for groundwater contamination. This is a result of comprehensive analysis. To further reduce NO3− contamination, an integrated approach should be set-up by combining N and O isotopes of NO3− with land-uses and physico-chemical properties, especially in areas with complex agricultural activities.