From engines to orbits: Unveiling the hidden CO and NO2 footprint of vehicle exhaust on satellite retrievals through machine-learning

Air pollution in rapidly urbanizing megacities is difficult to assess because ground-based monitoring networks are sparse and fragmented. This study proposes a transferable framework that integrates roadside LiDAR traffic sensing, adaptive traffic reconstruction (implemented via the Adaptive Smoothing Method), and Sentinel-5P satellite retrievals through ensemble machine-learning. Applied to Jakarta, the framework indicates that carbon monoxide (CO), as a direct primary pollutant, can be predicted from traffic emissions with very high accuracy (cross-validated R² up to 0.99), whereas nitrogen dioxide (NO₂), shaped by secondary atmospheric processes, exhibits more moderate predictability (cross-validated R² ≈ 0.73). For CO, motorcycles consistently emerge as the dominant source, while for NO₂ their influence remains substantial within traffic-related contributions, alongside passenger vehicles and heavy-duty traffic. These distinctions underscore the diverse roles of vehicle classes in shaping satellite-detectable urban air pollution. By bridging ground-level activity with orbital detection, the study presents a scalable approach to emission attribution in data-scarce environments. Beyond the Jakarta case, the framework offers a pathway for strengthening air quality assessment and supporting targeted pollution control strategies in cities worldwide.