Enhancing groundwater vulnerability assessment: Comparative study of three machine learning models and five...

Area of Science:

• Environmental Science
• Hydrogeology
• Machine Learning Applications

Background:

• Traditional groundwater vulnerability assessments often use binary classification, limiting nuanced understanding.
• Nitrate contamination poses a significant threat to groundwater resources globally.
• Machine learning offers advanced tools for complex environmental modeling.

Purpose of the Study:

• To develop and compare multi-class machine learning models for groundwater vulnerability assessment against nitrate contamination.
• To investigate the impact of the number of classes (three vs. five) on model performance.
• To identify the optimal machine learning model and classification scheme for accurate vulnerability mapping.

Main Methods:

• Employed three machine learning models: Random Forest, Extreme Gradient Boosting, and CART.
• Utilized parameters from the DRASTIC method, augmented with a Landuse parameter.
• Developed models using both three-class and five-class classification schemes.
• Evaluated models using metrics including Accuracy, Kappa, PPV, NPV, and AUC-ROC.

Main Results:

• The Random Forest model with a three-class classification achieved the highest performance, with an Area Under the Curve (AUC) of 0.95.
• Model evaluation metrics indicated significant differences in performance based on classification scheme and model choice.
• The study demonstrated the importance of the data classification process and the number of classes in machine learning predictions for groundwater vulnerability.

Conclusions:

• The Random Forest model, utilizing a three-class classification, is recommended for effective groundwater vulnerability assessment against nitrate contamination.
• The selection of an appropriate number of classes is crucial for enhancing the predictive accuracy of machine learning models in environmental studies.
• Integrating Geographic Information System (GIS) with advanced machine learning provides valuable strategies for groundwater resource management and mitigation of contamination.