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

The Air Quality Index (AQI) is a widely used indicator for assessing ambient pollution levels and associated health risks. However, AQI classification using data-driven approaches is often challenged by severe class imbalance, where extreme pollution categories such as Poor and Very Poor are sparsely represented. In this study, machine learning models were developed for multi-class AQI classification using a dataset comprising major air pollutants (PM₂.₅, PM₁₀, NO, NO₂, NOx, SO₂, CO, O₃, and volatile organic compounds) along with key meteorological variables, including relative humidity, wind speed, wind direction, solar radiation, and vertical wind speed. To improve model robustness, data preprocessing techniques such as handling missing values, distribution normalisation, and cyclical encoding of wind direction were applied. To address the pronounced class imbalance, a cluster-aware synthetic oversampling (CASO) framework was implemented by integrating random oversampling, Edited Nearest Neighbours (ENN), KMeans-SMOTE, and final class equalisation. Five machine learning models-Logistic Regression, Decision Tree, Random Forest, XGBoost, and LightGBM-were evaluated using accuracy, balanced accuracy, and macro-averaged F1-score. The experimental results demonstrate that ensemble gradient-boosting models, particularly LightGBM and XGBoost, consistently outperform other classifiers, achieving the highest test balanced accuracy values (≥0.96) after resampling. The findings confirm that combining cluster-aware synthetic resampling with ensemble learning significantly enhances AQI classification performance under severe data imbalance, offering a reliable and interpretable framework for air quality assessment.