Optimizing Air Quality Index Classification Using Multiple Machine Learning Models and Oversampling Techniques

Abstract

Air quality significantly affects public health, environmental stability, and ecosystem balance. Accurate classification of the Air Quality Index (AQI) is critical for effective monitoring and management. Previous studies often relied on a single machine learning algorithm, which limited classification performance, particularly under class imbalance conditions. This study evaluates multiple machine learning algorithms for AQI classification, including Logistic Regression, Decision Tree, K-Nearest Neighbors, Random Forest, Support Vector Machine, and Naïve Bayes. A random oversampling technique was applied to address the imbalance among AQI categories. The dataset consists of secondary data on pollutant concentrations (PM₁₀, SO₂, CO, O₃, NO₂) and AQI categories collected from five monitoring stations between 2010 and 2023. Model performance was assessed using accuracy, precision, recall, and F1-score. Before applying oversampling, the Random Forest model achieved an accuracy of 97.68%. After applying random oversampling, performance improved to 99.60%, with consistently high precision, recall, and F1-scores across classes. Feature importance analysis revealed that ozone (O₃) was the most influential pollutant, contributing 67.14% to model decision-making. The results demonstrate that combining random oversampling with ensemble-based machine learning substantially enhances AQI classification performance. This approach offers a robust and scalable framework for future air quality monitoring and environmental data analysis applications.