The integration of machine learning models into smart city infrastructures enables real-time pollution monitoring. This study demonstrates approaches for data collection, preprocessing, model deployment, and adaptive learning. Gradient Boosting achieved R² of 0.93, MAE 2.9 μg/m³, RMSE 4.2 μg/m³; DBSCAN identified pollution hotspots (silhouette 0.71, Davies-Bouldin 0.35). Deep Q-Network optimized control strategies with cumulative reward 1400, convergence 90%. An ensemble of Random Forest and Gradient Boosting reached R² 0.94. These models enhance pollution monitoring and management, improving urban air quality and public health. © 2025 Elsevier B.V., All rights reserved.