Improving fecal bacteria estimation using machine learning and explainable AI in four major rivers, South Korea

Area of Science:

• Environmental Science
• Water Quality Management
• Computational Hydrology

Background:

• Fecal coliform contamination poses a significant public health risk in South Korea's major rivers.
• Accurate estimation of fecal coliform levels is challenging due to complex environmental variables and data limitations.
• Existing models struggle with both prediction accuracy and interpretability of influencing factors.

Purpose of the Study:

• To enhance the accuracy and interpretability of fecal coliform contamination prediction in major South Korean rivers.
• To identify and understand the key environmental and water quality variables influencing fecal coliform levels.
• To apply advanced machine learning and Explainable Artificial Intelligence for improved water quality assessment.

Main Methods:

• Employed Random Forest (RF), Extreme Gradient Boosting (XGBoost), Deep Neural Network (DNN), and Convolutional Neural Network (CNN) models.
• Utilized Shapley Additive Explanations (SHAP) for model interpretability and variable contribution analysis.
• Analyzed a comprehensive dataset (2014-2022) of 16 water quality parameters and meteorological data from four major rivers.

Main Results:

• XGBoost and CNN models demonstrated improved accuracy in fecal coliform estimation.
• XGBoost achieved the optimal result with a validation Nash-Sutcliffe efficiency of 0.597 in the Han River.
• SHAP analysis provided clear insights into the significant factors driving fecal coliform concentrations across different river systems.

Conclusions:

• XGBoost offers superior estimation accuracy and robust explanations for variable contributions in fecal coliform prediction.
• The study enhances understanding of the complex relationships between water quality parameters and fecal coliform contamination.
• Findings support improved water quality management strategies for South Korea's major rivers.