Assessing the environmental determinants of micropollutant contamination in streams using explainable machine...

Area of Science:

• Environmental Science
• Water Quality
• Data Science

Background:

• Micropollutants like pesticides and pharmaceuticals in aquatic systems pose ecological risks.
• Existing machine learning (ML) methods struggle with complex environmental data and understanding variable interactions.
• There's a need for advanced analytical tools to identify micropollutant sources and drivers.

Purpose of the Study:

• To integrate explainable ML and network analysis for identifying micropollutant sources in a large watershed.
• To determine the key factors influencing micropollutant concentrations.
• To understand contamination pathways under different environmental conditions.

Main Methods:

• Applied and compared four ML algorithms: support vector machine, random forest, extreme gradient boosting (XGB), and autoencoder-XGB.
• Utilized Shapley Additive Explanations (SHAP) for model interpretability.
• Employed network analysis to reveal contamination pathways and event-specific characteristics.
• Used synthetic minority oversampling technique to handle data imbalance.

Main Results:

• The XGB model achieved high accuracy (87%-99%) in predicting micropollutant levels, especially high concentrations.
• SHAP analysis identified temperature and rainfall as significant factors affecting micropollutant levels.
• Agricultural activities were linked to pesticide pollution, while urban activities correlated with pharmaceutical contamination.
• Network analysis revealed distinct discharge pathways during dry and wet events.

Conclusions:

• Explainable ML and network analysis effectively identify micropollutant sources and influencing factors in mixed land-use watersheds.
• Understanding contamination pathways aids in developing targeted control measures and informed water quality policies.
• This integrated approach enhances the management of aquatic ecosystems facing micropollutant threats.