Assessing recreational water quality with machine learning and SHAP-based interpretability

This paper aims to explore how machine learning (ML) can be used to forecast water quality tailored for recreational use by identifying contamination levels in water. The project investigates the physical and chemical properties of water that can be used to forecast bacterial contamination, with the goal of identifying the predominant water features influencing bacterial contamination forecast. Multiple classification models were developed and evaluated including Random Forest, Logistic Regression, k-Nearest Neighbors (KNN), and XGBoost. I additionally utilized SHAP (SHapley Additive ex Planations) and SHAP analysis to provide a detailed explanation of the model’s performance. Based on the Fecal Coliform levels, the impact on the model’s performance was the strongest, with both the Conductivity and Biodegradable Oxygen Demand (BOD) coming after and being of the same level. This aligns the findings with the general notions of environmental science. The limitations of the research include regional basis and applicability of the model to recreational water use. The developments of the research can help incorporate larger datasets with coverage of larger regions and try to predict drinking water safety to help improve the model’s accessibility.