Towards green transportation: Predictive modeling of intersection congestion using machine learning for sustainable urban traffic management

Intersection congestion — primarily caused by frequent vehicle stops — leads to elevated fuel consumption and increased tailpipe emissions (CO, NO 2 , SO 2 , O 3 , PM 10 , PM 2.5 ), with well-documented adverse effects on public health. To enable smarter and more sustainable traffic operations, we propose a machine learning framework for classifying congestion levels at signalized intersections. The study is conducted using the CN+ dataset from Bremen, Germany. The target variable is constructed based on a capacity-driven volume-to-capacity (v/c) ratio using 10-minute traffic aggregates. Input features include traffic composition, approach direction, and temporal variables, with optional integration of meteorological and pollution data. To enhance model interpretability and reduce dimensionality, we introduce a novel feature selection method — Dual Importance Intersection Feature Selection (DIFS) — which combines Random Forest (RF) embedded importances with Chi-square statistics. Class imbalance is addressed through fold-internal application of the Synthetic Minority Over-sampling Technique with Edited Nearest Neighbors (SMOTE–ENN). All models are trained within a unified pipeline and evaluated via 5-fold stratified cross-validation (CV). The F1-score is adopted as the primary evaluation metric, while the Quadratic Weighted Kappa (QWK) is used to measure ordinal classification performance. Experimental results demonstrate that gradient-boosted tree models dominate in performance: Categorical Boosting (CatBoost) achieves an F1-score of 0.9937 and QWK of 0.9971, followed by Light Gradient Boosting Machine (LightGBM) (0.9723/0.9654) and eXtreme Gradient Boosting (XGBoost) (0.9652/0.9658). The optimized RF model achieves 0.9270/0.9316, while a compact Artificial Neural Network (ANN) yields lower performance (0.8563/0.8440). Final validation on a strictly unseen 10% hold-out set confirms the generalization ability of CatBoost, achieving an F1-score of 0.9957, QWK of 0.9957, and overall accuracy of 0.9956. These findings suggest that combining capacity-based congestion labeling, robust feature selection via DIFS, and ensemble learning offers a high-performance, deployment-ready solution for real-time, emission-aware urban traffic management systems.