Implementing machine learning methods for in-depth analysis and classification of surface water quality in Central...

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Water quality monitoring is crucial for environmental protection and public health.
• Ecological challenges and sustainable resource management necessitate advanced monitoring solutions.
• Machine learning offers potential for accurate water pollution assessment.

Purpose of the Study:

• To propose and evaluate a machine learning framework for water pollution classification.
• To compare the performance of Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost) models.
• To integrate regulatory compliance validation into the predictive framework.

Main Methods:

• Utilized Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost) for predictive classification.
• Employed Synthetic Minority Over-sampling Technique (SMOTE-Tomek) to address data imbalance.
• Conducted interpretability analyses using SHAP and LIME, and validated against Indonesia's Government Regulation No. 22/2021.

Main Results:

• XGBoost achieved 98.76% accuracy and 97.62% F1-Macro score; SVM achieved 90.25% accuracy and 83.57% F1-Macro score.
• Biological and chemical indicators (fecal coliform, BOD, COD) showed high feature importance.
• Model predictions identified potential "false-safe" classifications violating regulatory thresholds.

Conclusions:

• Machine learning models show high predictive accuracy for water pollution levels.
• Ensuring regulatory compliance is critical, as models may produce "false-safe" results.
• A regulatory-aware layer with rule-based validation and calibration is proposed for enhanced real-world applicability and policy alignment.