A cluster-aware synthetic resampling and machine learning framework for multi-class Air Quality Index classification.

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Air Quality Index (AQI) is crucial for assessing pollution and health risks.
• Data-driven AQI classification faces challenges due to imbalanced datasets, with extreme pollution levels underrepresented.
• Existing methods struggle with accurately classifying sparse, severe air pollution categories.

Purpose of the Study:

• To develop robust machine learning models for multi-class AQI classification.
• To address severe class imbalance in AQI datasets using advanced resampling techniques.
• To evaluate the effectiveness of ensemble learning models in improving AQI classification accuracy.

Main Methods:

• Utilized a comprehensive dataset including major air pollutants (PM2.5, PM10, etc.) and meteorological variables.
• Applied data preprocessing: missing value imputation, distribution normalization, and cyclical encoding for wind direction.
• Implemented a cluster-aware synthetic oversampling (CASO) framework integrating random oversampling, ENN, KMeans-SMOTE, and class equalization.

Main Results:

• Ensemble gradient-boosting models (LightGBM, XGBoost) demonstrated superior performance.
• Achieved high test balanced accuracy (≥0.96) after applying the CASO framework.
• The proposed resampling technique significantly enhanced classification accuracy for imbalanced AQI data.

Conclusions:

• Combining cluster-aware synthetic resampling with ensemble learning substantially improves AQI classification under severe data imbalance.
• The developed framework offers a reliable and interpretable approach for accurate air quality assessment.
• This study provides a robust methodology for handling imbalanced environmental datasets in machine learning applications.