Machine learning reveals flowpath-structured distributions of pesticides and pharmaceuticals

Pesticides and pharmaceuticals comprise hundreds of compounds applied in agricultural systems, either directly for pest control or indirectly via irrigation with treated wastewater, posing risks to downstream aquatic ecosystems. Advances in analytical chemistry now allow the simultaneous detection of dozens of micropollutants per sample, yielding extensive datasets. Identifying hydrological controls on compound occurrence from such datasets remains challenging due to strong non-linearities and complex compound behavior.

We investigated two agricultural fields located along the Kishon Stream (Israel), characterized by heavy clay soils and subsurface drainage systems that enable direct sampling of distinct hydrological flowpaths, including subsurface drainage discharge, surface runoff, and shallow groundwater. Sampling focused on first-storm conditions, when compound mobilization is most pronounced. While qualitative differences in compound occurrence among flowpaths were evident, quantitative attribution was hindered by the complexity of compound–flowpath relationships.

To address this, we applied Kernel Canonical Correlation Analysis (KCCA), a machine-learning method that captures non-linear associations through kernel mapping while retaining interpretability in a latent canonical space. KCCA was combined with non-parametric analyses and applied to both the original and transposed datasets. The analysis shows that pesticide and pharmaceutical distributions are strongly structured by hydrological flowpaths. Compound mobility and degradability modulate their occurrence within and across these pathways. We further define a dominant flowpath for individual compounds, identified as the pathway in which a compound attains its maximum representative concentration, providing a concise compound-level descriptor of flowpath association.

These results demonstrate the utility of KCCA for revealing hydrological structure and chemical properties in complex environmental datasets and highlight the importance of flowpath-specific distributions for understanding micropollutant occurrence in agricultural catchments.