This study introduces a digital twin paradigm that uses both stationary and mobile sensors and cutting-edge machine learning for urban air quality monitoring. By boosting $R^2$ values from 0.29 to 0.87–0.95, our two-step calibration method increased the accuracy of low-cost PM sensors, showing the possibility of growing monitoring networks without sacrificing measurement accuracy. Significant temporal and spatial variability in PM concentrations was found by mobile sensor deployments, with variations of up to 300% over short distances, predominantly during heavy traffic. During rush hours, peak concentrations were found on multi-lane boulevards and intersections, indicating important exposure concerns usually overlooked by stationary monitoring networks. According to our Graph Neural Network model, which successfully described pollutant dispersion patterns, road dust resuspension predominates in residential areas, while vehicle emissions account for 65% of PM_2.5 along high-traffic corridors. Urban green areas lower PM levels by 30%, yet when the current low-emission zones were first implemented, they had no discernible effect on air quality. Municipal authorities can use this digital twin strategy to acquire practical insights for focused air quality improvements. The method helps make evidence-based traffic management and urban planning judgments by identifying unidentified pollution hotspots and source contributions. The technique offers a scalable option for establishing healthier urban development and marks a substantial leap in environmental monitoring.