Estimating Ground-Level Carbon Monoxide Concentrations Using Machine Learning Techniques: The Metropolitan City of Milan Case Study

Abstract. This work presents a structured data-driven framework for estimating ground-level carbon monoxide (CO) concentrations in the Metropolitan City of Milan (MCM) by integrating Sentinel-5P satellite observations, Copernicus Atmosphere Monitoring Service reanalysis data, and ERA5 meteorological variables with advanced machine learning techniques. The methodology employs unified data preprocessing, systematic feature engineering (e.g., boundary layer height-adjusted CO, lagged meteorological variables), Bayesian optimization for hyperparameter tuning, SHAP-based feature selection, model ensembling, and robust statistical validation. Eight regression models, including a custom Dense Attention Network (DAN), were evaluated across multiple temporal aggregation windows (4–24 hours before 15:00 GMT+1) to identify optimal configurations for CO estimation. Using data from January 2019 to November 2024, the framework identified the 21:00–15:00 GMT+1 window as most effective for capturing atmospheric dynamics such as nighttime accumulation, morning emission peaks, and daytime dilution. The DAN achieved the best performance, with a mean normalized root mean squared error of 0.4879 ± 0.0252 on the test set, outperforming ensemble and traditional regression models, offering a scalable, interpretable, and cost-effective approach to urban CO monitoring in data-scarce environments with potential adaptation to other pollutants and regions.