Spatiotemporal estimates of anthropogenic NO_x emissions across China during 2015–2022 using a deep learning model

As one of the significant air pollutants, nitrogen oxides (NO_x = NO + NO₂) not only pose a great threat to human health, but also contribute to the formation of secondary pollutants such as ozone and nitrate particles. Due to substantial uncertainties in bottom-up emission inventories, simulated concentrations of air pollutants using GEOS-Chem model often largely biased from those of ground-level observations. To address this issue, we developed a new deep learning model to simulate the inverse process of the GEOS-Chem model. This framework was applied to improve the total anthropogenic NO_x emission intensity over a five-year period (2015–2019), and then to predict anthropogenic NO_x emission intensity for 2020–2022. The deep learning model showed higher R² value (0.94) based on 10-fold cross-validation, indicating that the model effectively captured spatial features and patterns. Then, NO_x emission intensity was calibrated based on the new framework using high-resolution NO₂ concentration dataset instead of the simulated NO₂ levels derived from GEOS-Chem model. Overall, the top-down inversion result was in agreement with the bottom-up emission inventory at the spatial scale. The top-down emission fluxes were lower in high-emission regions such as central China, Beijing, Shanghai, the Pearl River Delta, and the central part of Liaoning, while posterior estimates were higher in regions with lower prior emission intensity. The posterior NO_x emission intensity suggested that some major regions such as Beijing-Tianjin-Hebei (BTH) (-56.4 %) and Pearl River Delta (PRD) (-52.5 %) experienced dramatic NO_x emission intensity decreases from 2015 to 2022, whereas some remote regions such as Tibetan Plateau remained relatively stable. This research contributes to the timely tracking of changes in pollutant emission and aids in the formulation of more effective and relevant pollution prevention and control policies.