An Efficient Framework Integrating Machine Learning and Deep Learning Techniques for Predicting Air Quality Index Category in Dhaka City

Air pollution poses a critical challenge in densely populated urban areas like Dhaka, with profound implications for public health and environmental stability. This research presents a comprehensive framework designed to predict the Air Quality Index (AQI) category with high accuracy. The approach employs robust pre-processing steps, e.g., handling of missing data using Synthetic Minority Oversampling Technique (SMOTE) to mitigate class imbalance and integrates both machine learning and deep neural network classifiers. A diverse set of models, i.e., Random Forest, XGBoost, LightGBM, Multilayer Perceptron, and TabNet are individually assessed in an AQI dataset of Dhaka city. Subsequently, the individual models are combined using ensemble strategies such as Voting and Stacking classifiers. Experimental results demonstrate that the ensemble techniques, i.e., Voting and Stacking classifiers, consistently deliver nearperfect performance, achieving accuracies $\sim 0.9999$ and macro F1scores exceeding 0.99. To enhance interpretability, SHAP (SHapley Additive exPlanations) analysis is performed, offering transparent insights into the model behaviour and aiding policylevel decision-making.