Diagnosing Ozone-NOx-VOCs-Aerosols Sensitivity to Uncover Urban-nonurban Discrepancies in Shandong, China using Transformer-based High-resolution Air Pollution Estimations
Abstract. Narrowing surface ozone disparities between urban and nonurban areas escalate health risks in densely populated urban zones. A comprehensive understanding of the impact of ozone photochemistry processes on this transition remains constrained by our knowledge of aerosol effects and the spatial availability of surface monitoring. Here we developed a novel deep learning framework, which could perceive spatiotemporal dynamics from adjacent grids by multidimensional self-attention operation, integrating multi-sources data to estimate daily 500 m surface ozone, nitrogen dioxide (NO2) and fine particulate matter (PM2.5) concentrations. Subsequently, three distinct ozone formation regimes linked with its precursors, aerosols, and meteorology were delineated through an interpretable machine learning method. The evaluations of the framework exhibited average out-of-sample cross-validation coefficient of determination of 0.96, 0.92 and 0.95 for ozone, NO2 and PM2.5, respectively. In 2020, urban ozone levels in Shandong surpassed those in nonurban due to a more pronounced decrease in ozone in the latter where PM2.5 is the dominant anthropogenic driver. The ozone sensitivity to volatile organic compounds (VOCs), the dominant regime in urban areas, was observed to shift towards a NOx-limited when extended to rural areas. A third ‘aerosol-inhibited’ regime was identified in the Jiaodong Peninsula, where the uptake of hydroperoxyl radicals onto aerosols suppressed ozone production under low NOx levels during summertime. The reduction of PM2.5 would increase the sensitivity of ozone to VOCs, necessitating more stringent VOC emission abatement for urban ozone mitigation. Our case study demonstrates the critical need for advanced modeling approaches providing finer spatially resolved estimations.
Surface Ozone, NO2, and PM2.5 Concentrations Estimated by the Deep Learning model (Air Transformer) based on Satellite data. https://doi.org/10.5281/zenodo.10071408
Model code and software
Viewed (geographical distribution)