Comparative evaluation of machine learning models for groundwater quality assessment

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Groundwater contamination by pesticides and nitrate poses a significant risk to water quality, especially in agricultural regions.
• Predicting contamination levels is challenging due to sparse data and complex, non-linear relationships between variables.

Purpose of the Study:

• To evaluate the efficacy of artificial neural networks (ANN), support vector machines (SVM), and extreme gradient boosting (XGB) in predicting groundwater contamination.
• To compare the classification and regression performance of these models under different class scenarios (2, 3, and 4 classes).
• To assess the impact of class imbalance and test mitigation techniques on model performance.

Main Methods:

• Utilized a dataset of 303 wells across 12 Midwestern US states, incorporating hydrogeologic, water quality, and land use features.
• Applied ANN, SVM, and XGB models for both classification and regression tasks.
• Investigated class imbalance using oversampling, weighting, and combined oversampling and weighting techniques.
• Employed accuracy, F1 score, and Matthews Correlation Coefficient (MCC) for performance evaluation.
• Used game-theoretic Shapley values for feature importance analysis and model interpretability.

Main Results:

• All three machine learning models demonstrated strong predictive capabilities for groundwater contamination.
• Class imbalance mitigation techniques, particularly combined oversampling and weighting, significantly improved model performance.
• Feature importance analysis identified key hydrogeologic, water quality, and land use variables influencing contamination levels.

Conclusions:

• Machine learning models are effective tools for predicting pesticide and nitrate contamination in groundwater, even with sparse and non-linear data.
• Addressing class imbalance is crucial for accurate predictive modeling in environmental studies.
• Feature importance analysis provides valuable insights for targeted water quality management strategies.