Top-down estimates of anthropogenic NO_x emissions over China through a new zone-stratified RF machine learning model

This study designs a new zone-stratified Random Forest (RF) machine learning model to estimate monthly anthropogenic NO_x emissions over China at the spatial resolution of 0.25° × 0.25°. Zone-stratified RF divides China into different emission zones to estimate anthropogenic NO_x emissions separately, thereby showing the improvement over the unstratified RF that is applied for the whole study domain and resulted in apparent overestimation at high emission regions and underestimation at low emission regions. In the training, Multi-resolution Emission Inventory for China (MEIC) anthropogenic NO_x emission inventory is used as a target, while features include Tropospheric Monitoring Instrument (TROPOMI) tropospheric NO₂ Vertical Column Density (VCD), fifth generation ECMWF reanalysis meteorological data, Infrared Imaging Radiometer Suite nighttime light data, and WorldPop Global Project Population data. Once trained for 2019, the zone-stratified RF model is further applied in other years for independent evaluation. Total anthropogenic NO_x emission estimates over China in 2020 from zone-stratified RF is 20.6 Tg, comparable to 19.7 Tg from MEIC, and the difference between them is less than 0.1 Tg for 27 provinces. Additionally, the nationwide mean difference is 0.04 Tg, corresponding to an average error of 8.6%. The pronounced urban-rural emission gradient in MEIC is accurately reproduced by zone-stratified RF, but not by unstratified RF. Furthermore, over Taiwan province where MEIC data is not available for the training, the trained zone-stratified RF model estimate of anthropogenic NO_x emissions exhibit good agreement of spatial pattern with Emissions Database for Global Atmospheric Research (EDGAR). Temporally, zone-stratified RF also identifies emission reduction during COVID-19 lockdown in Taiwan. Overall, zone-stratified RF is effective for rapid update and projection of NO_x emissions.