Classification of Nitrate Polluting Activities through Clustering of Isotope Mixing Model Outputs

Apportionment of nitrate (NO) sources in surface water and classification of monitoring locations according to NO polluting activities may help implementation of water quality control measures. In this study, we (i) evaluated a Bayesian isotopic mixing model (stable isotope analysis in R [SIAR]) for NO source apportionment using 2 yr of δN-NO and δO-NO data from 29 locations within river basins in Flanders (Belgium) and five expert-defined NO polluting activities, (ii) used the NO source contributions as input to an unsupervised learning algorithm (k-means clustering) to reclassify sampling locations into NO polluting activities, and (iii) assessed if a decision tree model of physicochemical data could retrieve the isotope-based and expert-defined classifications. Based on the SIAR and δB results, manure/sewage was identified as a major NO source, whereas soil N, fertilizer NO, and NH in fertilizer and rain were intermediate sources and NO in precipitation was a minor source. The k-means clustering algorithm allowed classification of NO polluting activities that corresponded well to the expert-defined classifications. A decision tree model of physicochemical parameters allowed us to correctly classify 50 to 100% of the sampling locations as compared with the k-means clustering approach. We suggest that NO polluting activities can be identified via clustering of NO source contributions from samples representing an entire river basin. Classification of future monitoring locations into these classes could use decision tree models based on physicochemical data. The latter approach holds a substantial degree of uncertainty but provides more inherent information for dedicated abatement strategies than monitoring of NO concentrations alone.