Enhancing Air Quality Index Classification Based on Ensemble Machine Learning Techniques

Abstract

The accurate classification of Air Quality Index (AQI) is critical for environmental monitoring and public health protection. In this paper, we utilized a publicly available daily air quality dataset from U.S. counties, comprising six classification categories: Good, Moderate, Unhealthy for Sensitive Groups, Unhealthy, Very Unhealthy, and Hazardous. The dataset underwent preprocessing through missing value imputation and class balancing using the Synthetic Minority Over-sampling Technique (SMOTE). Several machine learning and deep learning models were trained and evaluated on the dataset, including Random Forest (RF), Extra Trees (ET), K-Nearest Neighbors (KNN), Naive Bayes (NB), Logistic Regression (LR), and a Multi-Layer Perceptron (MLP) neural network. The models were assessed using cross-validation accuracy, test set accuracy, macro-averaged recall, F1-Score, and ROC-AUC metrics. Ensemble methods (RRF and ET) and the MLP classifier achieved superior results compared to traditional models. The RF model achieved a test accuracy of 99.3%, while the MLP classifier achieved 99.0% . The stacking ensemble model achieved a test accuracy of 99.99 %, a macro-averaged recall of 87.12 %, and an ROC-AUC of 1.0000, highlighting the strong potential of ensemble learning techniques in enhancing the performance of AQI multi-class classification.