Learning Emissions Signatures: A Machine Learning Approach to Vehicle CO₂ Classification

This study introduces a machine learning framework for classifying vehicle CO₂ emissions using vehicle specifications. A quantile-based method categorized emissions into Low, Medium, and High levels. Decision Tree, Random Forest, and LightGBM classifiers were trained and optimized using the Puma Optimizer (PO), and a Soft Voting Ensemble was created to enhance generalization. Random Forest yielded the highest performance (97% accuracy, 0.97 macro-F1 score), closely followed by the Soft Voting Ensemble (96.82% accuracy, 0.97 macro-F1 score). SHAP analysis revealed combined fuel consumption and engine size as key predictors. This work uniquely combines quantile-based classification, optimized ensemble learning, and explainable AI for interpretable vehicle emission assessment and green transport policy support.