Tracking diurnal variation of NO2 at high spatial resolution in China using a time-constrained machine learning model

• A temporally conservative machine learning model (T-XGB) is designed. • T-XGB shows exemplary NO 2 estimation capability and temporal stability. • Daily cyclic of NO 2 at fine spatial resolution in China is explored. • The NO 2 population exposure risk exhibits a distinct diurnal variation. The spatially continuous dynamic monitoring of near-surface NO 2 concentrations on sub-daily scales would serve to enhance awareness of the current state of air pollution, which is crucial to improving regional air quality. Satellites, like OMI and TROPOMI, are capable of observing atmospheric NO 2 column concentrations on a global scale. However, the fixed transit times of the satellites and severe data deficiencies restricted their applicability for revealing patterns of change in NO 2 on sub-daily scales. This study proposes a time-constrained XGBoost model (T-XGB) to convert multi-source information to daily cumulative near-surface NO 2 concentrations. Furthermore, a temporally conservative downscaling framework is developed to facilitate seamless monitoring of near-surface NO 2 at the 0.03°/3-hour scale in China. Evaluated with in-situ NO 2 measurements, the results have demonstrated the robust and excellent performance of the T-XGB (R 2 : 0.920–0.948; MAE: 2.89–3.67 µg/m 3 /h), as well as the accuracy of the temporally conserved downscaling technique (R 2 > 0.973). The 3-hour near-surface NO 2 was consistent with the TROPOMI observations at the corresponding moments and it exhibited a detailed gradient variation signature. In China, near-surface NO 2 exhibited a single-peak diurnal variation, with an initial increase followed by a subsequent decrease. The maximum concentration was observed between 8p.m. and 11p.m. in local time. The assessment of NO 2 pollution exposure can yield disparate results when evaluated at varying time scales. Sub-daily monitoring of NO 2 provides a more detailed and nuanced understanding of the pollutant, making it a more applicable and flexible tool for use in subsequent studies.