Interpretable tree-based ensemble model for predicting beach water quality.

Area of Science:

• Environmental Science
• Machine Learning
• Public Health

Background:

• Microbial fecal contamination in beach water poses health risks.
• Existing predictive models are often 'black boxes,' lacking transparency.
• Interpretable models are needed for accountability and understanding decisions.

Purpose of the Study:

• To develop and evaluate interpretable tree-based machine learning models for predicting beach water quality.
• To assess the performance of models like LightGBM and XGBoost using SHAP for explainability.
• To identify key environmental predictors for microbial contamination in freshwater lakes.

Main Methods:

• Evaluated five tree-based models: classification tree, random forest, CatBoost, XGBoost, and LightGBM.
• Employed the SHAP (SHapley Additive exPlanations) method for model interpretability.
• Tested models using Escherichia coli (E. coli) concentration data from Lake Erie beaches.

Main Results:

• LightGBM and XGBoost demonstrated the highest precision and recall scores.
• Lake turbidity was identified as the most significant predictor across all sites.
• Wave height and rainfall data were crucial for accurate model development, with notable turbidity-day-of-year interactions.

Conclusions:

• The combination of LightGBM and SHAP offers a promising approach for interpretable microbial water quality prediction.
• Accurate local environmental data enhances the reliability of predictive models for beach water safety.
• Interpretable models can improve public health risk communication and decision-making.