Dynamic Source Apportionment of Riverine Nitrate Pollution Using Dual Stable Isotopes and Machine Learning Algorithms

This study integrated a 10-year data set of nitrate isotopes (15N-NO3–/18O-NO3–) and water quality containing precipitation, soil, groundwater, and river water in the Yongan watershed. We developed a window-based dynamic MixSIAR framework that explicitly accounts for temporal variations in the source isotope characteristics. Compared with the fixed-source MixSIAR, the dynamic framework significantly reduced the uncertainty of source apportionment results (the coefficient of variation decreased from 1.21 to 0.55) and effectively identified the primary N pollution sources and their temporal variations across different watershed types. Contributions were 41.1 ± 13.3% from soil in the entire watershed, 38.5 ± 12.4% from soil in a typical agricultural catchment, 58.8 ± 17.9% from wastewater in a typical residential catchment, and 45.9 ± 16.1% and 42.4 ± 17.1% from soil and groundwater in a typical forest catchment, respectively. Building upon the dynamic apportionment results, machine learning models (ML) were further developed by using water quality data to predict source contributions. Among these models, XGBoost, random forest, and support vector machine achieved the highest predictive performance for calculating the contributions of soil (R2 = 0.83, Nash–Sutcliffe efficiency coefficient (NSE) = 0.83, and root-mean-square error (RMSE) = 0.07), groundwater (R2 = 0.76, NSE = 0.73, and RMSE = 0.10), and wastewater (R2 = 0.71, NSE = 0.68, and RMSE = 0.14). As an exploratory study integrating long-term isotope data with ML, this study provides a comprehensive approach for apportioning riverine N sources in watersheds with limited isotope data worldwide.