Interpretation of ensemble learning to predict water quality using explainable artificial intelligence.

Area of Science:

• Environmental Science
• Water Quality Management
• Machine Learning Applications

Background:

• Algal blooms pose significant risks to freshwater quality and drinking water safety.
• Chlorophyll-a (Chl-a) concentration is a key indicator for estimating algal presence.
• Accurate prediction of Chl-a is essential for effective water resource management.

Purpose of the Study:

• To develop and evaluate an XGBoost machine learning model for predicting Chl-a concentration.
• To assess the impact of input variable selection on model performance using Explainable Artificial Intelligence (XAI).
• To identify the most effective method for prioritizing input variables for improved prediction accuracy.

Main Methods:

• An XGBoost ensemble model was trained using eighteen input variables.
• Input variable selection was guided by Shapley value (SHAP), feature importance (FI), and variance inflation factor (VIF).
• Model performance was evaluated using metrics such as RMSE, RMSE-observation standard deviation ratio, and Nash-Sutcliffe efficiency.

Main Results:

• The XGBoost model demonstrated the most stable and accurate performance when input variables were prioritized using SHAP analysis.
• SHAP analysis provided consistent and interpretable insights into the relative importance of input variables.
• XAI techniques, including SHAP plots, enhanced the understanding of the XGBoost model's predictive behavior.

Conclusions:

• SHAP-based variable selection is a highly effective strategy for optimizing machine learning models in water quality prediction.
• Implementing SHAP analysis can lead to more cost-effective on-site monitoring by focusing on critical input variables.
• This study highlights the successful application of XAI in improving the interpretability and reliability of water quality prediction models.