NOx Emission Changes Over China During the COVID‐19 Epidemic Inferred From Surface NO2 Observations

Abstract The COVID‐19 epidemic has substantially limited human activities and affected anthropogenic emissions. In this work, daily NOx emissions are inferred using a regional data assimilation system and hourly surface NO2 measurement over China. The results show that because of the coronavirus outbreak, NOx emissions across the whole mainland China dropped sharply after 31 January, began to rise slightly in certain areas after 10 February, and gradually recover across the country after 20 February. Compared with the emissions before the outbreak, NOx emissions fell by more than 60% and ~30% in many large cities and most small to medium cities, respectively. Overall, NOx emissions were reduced by 36% over China, which were mainly contributed by transportation. Evaluations show that the inverted changes over eastern China are credible, whereas those in western China might be underestimated. These findings are of great significance for exploring the reduction potential of NOx emissions in China.

WRF version 4.0 and CMAQ version 4.7.1 are applied to simulate the meteorological fields and atmospheric compositions (Byun & Schere, 2006;Skamarock & Klemp, 2008;Yu et al., 2012). WRF simulations cover the whole of East Asia (169 west-east and 129 south-north cells) with a 36-km horizontal grid spacing ( Figure S2). The vertical grid on sigma-pressure coordinates was extended to 100 hPa with 51 layers.
The underlying surface of urban and built-up land is replaced by the latest MODIS land cover retrieval to adapt to the rapid expansion of urban development. The CMAQ model is run with the same domain but with three gird cells removed from each side of the WRF domain. There are 15 layers in the CMAQ vertical coordinate, which are compressed from 51 WRF layers. Meteorological initial and lateral boundary conditions come from the 6-h intervals, 1° × 1° Final (FNL) Operational Global Analysis data of the National Center for Environmental Prediction (NCEP). Chemical initial and lateral boundary conditions come from the last cycle and background profiles.
The Carbon Bond 05 chemical mechanism (CB05) is chosen as the gas-phase chemical mechanism (Guenther et al., 2012). Detailed physical and chemical options are listed in Table S1. NCEP's GSI 3DVAR DA system is used to optimize chemical initial condition. To find a best estimate of analysis in the sense of minimum analysis error variance, the GSI 3DVAR algorithm considers two sources of initial information: observations at irregularly spaced points and a gridded background field. The analysis can be determined by the minimizing a scalar objective function J(x) given by where is the ensemble-estimated background error covariance matrix, K is the Kalman gain matrix, and N is the ensemble size. Combined with observational vector y, the ensemble mean of the analysis state are updated.
The quality control method of CO and SO2 employed is similar to that of NO2. CO and SO2 values larger than 10 mg m -3 and 400 µg m −3 respectively are classified as unrealistic and rejected. Additionally, time-continuity is checked to eliminate the values larger/smaller (Ta + 0.15y(t)) than the data at adjacent times, where y(t) represents the observations, and Ta is set to 2 and 80 for CO and SO2, respectively. Observations within each city and the same grid are averaged to improve the representativeness and reduced the observation error correlations. For error settings, except that the of measurement error is set to 0.02 and 1.0 for CO and SO2, respectively, other parameters are the same as those of NO2.

Text S2.
The performance of meteorological simulation is critical for emission inversion, because meteorological processes notably affect pollutants' transport, mixing and chemical reactions, and determine the estimation of the flow-dependent background error covariance. Generally, both higher temperature and lower relative humidity lead to a faster photolysis of NO2, and a stronger wind corresponds to a better diffusion and transport of air pollutants, resulting in lower NO2 concentrations in the atmosphere.
Because all the biases between the simulated and observed concentrations are assumed to be attributed to the emissions during the inversion, the lower concentrations caused by the errors in the simulated meteorological fields may lead to overestimation of the emissions. To quantitatively evaluate the accuracy of the simulated meteorological fields, the mean bias (BIAS), root mean square error (RMSE), and correlation coefficient (CORR) are calculated against the surface meteorological observations obtained from the National Climate Data Center (NCDC) integrated surface database (http://www.ncdc.noaa.gov/oa/ncdc.html). The spatial distribution of 350 meteorological sites is shown in Figure S2. The simulations were conducted from 11 January to 29 February 2020. Figure S3 shows the spatial distribution of the mean bias of the WRF simulations and their changes before and during the outbreak. The   16.4% and 25.6%, respectively. Although the difference of the total emissions across the country between prior and posterior emissions is small, the difference at the regional/local scale is quite large ( Figure S5d).  (Table S3).

Text S4.
The city human mobility index represents the scale of the population moving into the city. The administrative boundary is adopted as the migration boundary of the city.
Therefore, the index could reflect the human activity level in a specific city to some extent. These migration data were obtained from the Baidu Map (http://qianxi.baidu.com/), which uses big data technology to analyze "location-based services" data, and dynamically and real-time displays the source and destination of population migration from the regional and time dimensions.
Text S5.            Table S3. Statistics comparing the NO2 concentrations from the simulations with prior and posterior emissions against assimilated and independent observations, respectively. Statistics are calculated against daily regional averaged observations (50 pairs). The number next to the region name indicates the number of cities.