Explainable artificial intelligence for the interpretation of ensemble learning performance in algal bloom...

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Chlorophyll-a (Chl-a) concentration is a critical indicator for monitoring algal blooms.
• Accurate Chl-a estimation is vital for effective water quality management.

Purpose of the Study:

• To estimate Chl-a concentrations using the XGBoost machine learning model.
• To evaluate model performance across different data frequencies and identify key predictive variables.
• To enhance the practical applicability of machine learning models for on-site water quality management.

Main Methods:

• Employed the XGBoost machine learning model with 23 water quality and meteorological variables.
• Evaluated model performance using RMSE, RSR, and Nash-Sutcliffe efficiency across nine datasets with varying time frequencies (1 hour to 1 month).
• Utilized Shapley value (SHAP) analysis, an explainable AI method, to understand variable importance and impact.

Main Results:

• XGBoost model performance was stable for high-frequency data (1-24 hours) with RSR between 0.61-0.65.
• Model performance significantly declined with weekly and monthly data intervals.
• SHAP analysis revealed that stable model performance was achieved with five or fewer key variables, including pH, dissolved oxygen, and turbidity.

Conclusions:

• Machine learning models, particularly XGBoost, can effectively estimate Chl-a concentrations.
• Explainable AI (XAI) is crucial for understanding model behavior and identifying essential predictive variables.
• The study demonstrates the feasibility of using machine learning with real-time sensor data for practical water quality management.