Stacked Ensemble Learning for Air Quality Category Prediction with Interpretable Insights

Air pollution is nowadays a major global issue that requires precise and particularly reliable systems for predicting air quality to reduce the health hazards in humans. This research work suggests a stacked learning method that identifies the Air Quality Index (AQI) by using Ensemble learning with Random Forest (RF), Extra Trees (ET), Extreme Gradient Boosting (XGB), as base models, and Logistic Regression (LR) as a Meta model. The data is categorized into four classes as Good, Moderate, Unhealthy for sensitive, and Unhealthy. A Stacking ensemble strategy is used, where the outputs of RF, ET, and XGB classifiers are passed to an LR meta-learner that learns to combine them for improved predictive accuracy. Performance evaluation metrics, including accuracy, precision, recall, and f1-score, show that the ensemble was outperforms standalone models, achieving superior classification accuracy. To enhance model transparency and interpretability, Explainable AI techniques, including SHAP and LIME, are used to analyze feature contributions. Furthermore, the feature importance analysis identifies key pollutant sources (e.g., PM10, O2, and SO2) that have an impact on the AQI predictions. The simulation result shows that the proposed framework achieved an accuracy of 93.15% indicating robust performance in classifying AQI categories. This framework is an accurate tool for environmental policymakers and agencies to assess and predict the quality of air and enables timely intervention. The future research could explore deep learning-based ensembles as well as live-time prediction of AQI with wider application.