In an effort to control the spread of COVID-19, the Chinese government implemented a range of restrictions on movement. These led to reductions in industrial and other work-related and personal activities starting 23 January 2020 in Wuhan, Hubei province, and then extending to other cities and regions in the days that followed. On 8 April 2020, Wuhan was the last city to reopen after a complete lockdown that prevented most people from leaving their homes. These measures have been linked to changes in air quality. A network of surface monitoring stations in northern China observed 35 % decreases in PM2.5 and 60 % decreases in NO2 concentrations during 29 January through 29 February, as compared to the preceding 3 weeks; CO and SO2 also declined (Shi and Brasseur, 2020). In and around Wuhan, decreases in NO2 and PM2.5 were similar to regional changes, but there was a slight increase in SO2 concentrations (Shi and Brasseur, 2020). Observations by the Tropospheric Monitoring Instrument (TROPOMI) showed large decreases in tropospheric NO2 column densities over Chinese cities, on the order of 40 % for 11 February to 24 March 2020 compared to the same period in 2019, ranging from roughly 25 % for cities not affected by lockdown to 60 % for Wuhan and Xi'an (Bauwens et al., 2020). Prospective simulations suggested that meteorology may limit the effect of reduced emissions on PM2.5 concentrations, with Chinese cities experiencing less than 20 % reductions (Wang et al., 2020).

The goal of our study was to consider these changes against pollution trends in China using NASA Earth Observing System data by combining several products to give a holistic view covering several emission sectors that are responsible for the observed changes. Over the last 2 to 3 decades, air pollution in China appears to have followed the pattern described by the environmental Kuznets curve (Selden and Song, 1994). This framework describes a relationship in which economic growth is initially accompanied by an increase in air pollution, when poverty remains widespread. But as growth continues, air pollution is expected to level off and decline as a consequence of changes in social awareness of environmental degradation and the economic, political, and technological capacity to limit it (Sarkodie and Strezov, 2019; Selden and Song, 1994).

Bottom-up and top-down assessments of air pollutant emissions and concentrations suggest that China has followed this pattern during the era of satellite monitoring of atmospheric composition, with concentrations of SO2, NO2, CO, and aerosol optical depth (AOD) mostly exhibiting marked and steady declines over the last decade. In the case of NO2, multi-instrument analyses, which extend the observational record beyond the lifetime of a single instrument, depict a consistent regional picture of NO2 trends in China since 1996 (Geddes et al., 2016; Georgoulias et al., 2019; Wang and Wang, 2020; Xu et al., 2020). Column totals show an increasing trend during the first part of the satellite record, but this trend is reversed sometime between 2010 and 2014 (Georgoulias et al., 2019; Krotkov et al., 2016; Lin et al., 2019; Xu et al., 2020; Si et al., 2019; Shah et al., 2020). The trend reversal has been attributed to a combination of emission control measures (Zheng et al., 2018a) and variations in economic growth (Krotkov et al., 2016).

Bottom-up estimates suggest that SO2 emissions peaked earlier, with declines starting around 2005, primarily as a result of power and industrial pollution control measures as well as the elimination of small industrial boilers (Sun et al., 2018; Zheng et al., 2018b). An earlier peak in SO2 emissions is consistent with observations by multiple satellite instruments, which revealed declines in SO2 column densities since 2005 (Fioletov et al., 2016; Krotkov et al., 2016; Wang and Wang, 2020; Zhang et al., 2017; Si et al., 2019).

AOD retrievals from the Along Track Scanning Radiometer instruments show a steady increase over southeastern China from 1995 to 2005 (Sogacheva et al., 2020) and a decline since 2005 in the MODIS AOD (He et al., 2019). The AOD peak has been argued to match the ∼ 2011 peak in NO2 (Zheng et al., 2018b; Xie et al., 2019), to match the ∼ 2005 peak of SO2, or to have occurred at some point in between (Ma et al., 2016), with more rapid decreases in AOD after 2011 (Lin et al., 2018). The recent decrease in AOD is also seen in Visible Infrared Imaging Radiometer Suite (VIIRS) retrievals (Sogacheva et al., 2020). Most mitigation of direct PM2.5 emissions since 2010 was by industry, with residential emissions also decreasing substantially (Zheng et al., 2018b). The decline in SO2 emissions also exerted an important influence, with the sulfate concentration of PM2.5 decreasing substantially between 2013 and 2017 (Shao et al., 2018), reflecting the negative trend in SO2 emissions.

The peak in concentrations of CO, which has an atmospheric lifetime ranging from weeks to months, is less easily identified. Some studies suggest that trends have been negative potentially throughout the 21st century (Han et al., 2018; Strode et al., 2016; Wang et al., 2018; Yumimoto et al., 2014; Zheng et al., 2018a), but others suggest that emissions and/or column densities were increasing or flat during at least the first decade of the century (Sun et al., 2018; Zhao et al., 2013, 2012). The negative trend has been attributed largely to reductions in emissions from industrial activity, as well as from residential and transportation sectors (Zheng et al., 2018a, b).

In addition to these long-term trends, a number of air pollutants also exhibit strong seasonal variation in China. Anthropogenic emissions of CO, SO2, and PM2.5 are highest in winter, reflecting large variation in emissions from the residential sector and, in the case of CO, increased emissions associated with cold-start processes in the transportation sector (Li et al., 2017). Outflow of CO and AOD has a spring maximum, resulting from transport of pollution, dust, and boreal biomass burning emissions (Han et al., 2018; Luan and Jaegle, 2013).

Changes in pollution over China have also come from short-term interventions. To improve air quality for the 2008 summer Olympics – a time when emissions in China were high and still increasing – the Chinese government imposed a series of strict emission control measures from July through 21 September 2008, which were qualitatively similar to the emission reductions expected to have accompanied the COVID-19 lockdown (UNEP, 2009). As a result, NO2 concentrations over Beijing were estimated to have declined by between 40 % and 60 % based on satellite observations, with substantial but smaller reductions in surrounding cities often on the order of 20 % to 30 % compared to previous years (Mijling et al., 2009; Witte et al., 2009). Regional reductions of SO2 and CO during the months of the games were estimated to be 13 % and 19 %, respectively (Witte et al., 2009). These results are broadly consistent with on-road observations (Wang et al., 2009) but larger than some surface observations comparing concentrations before and after the emission control measures were implemented (Wang et al., 2010).

The COVID-related lockdowns provide a similar natural experiment to the 2008 Beijing Olympics but on the other side of the Kuznets curve. The fact that the lockdowns occurred during years of decreasing air pollution needs to be taken into account in attributing changes in atmospheric composition to COVID-19 lockdowns, independent of the long-term trend. Following Chen et al.'s (2020) analysis of air quality improvements on mortality which controlled for changes in air quality since 2016, in this study we determine whether changes in 2020 in satellite retrievals of CO, SO2, NO2, and AOD departed significantly from the expected declines associated with the long-term decreases in concentrations resulting from pollution controls and technological change.

We used daily level-3 (L3) retrievals from four different instruments on three different NASA Earth Observing System satellites. The Atmospheric Infrared Sounder (AIRS) instrument aboard NASA's Aqua satellite is a 2300-channel infrared grating spectrometer in a sun-synchronous orbit with northward Equator crossing time of 13:30. AIRS carbon monoxide (CO) profiles are retrieved with horizontal resolution of 45 km at nadir, in a swath width of 30 fields of view or about 1600 km. The retrieval uses a cloud-clearing methodology providing CO with sensitivity that peaks around 500 hPa, with ∼ 0.8–1.2 degrees of freedom of signal for 50 %–70 % of scenes. More sampling and higher information content is obtained in clear scenes (Warner et al., 2013). We used the daily version 6 (AIRS3STD.006) product.

The Ozone Monitoring Instrument (OMI) aboard NASA's Aura satellite was launched in July 2004 and has a local Equator crossing time of roughly 13:45. OMI is a nadir-viewing spectrometer, which measures solar backscatter in the UV–visible range (Krotkov et al., 2017). We used NASA's L3 tropospheric NO2 column density standard product v3 (OMNO2d_003) and the OMI principal component analysis planetary boundary layer (PBL) SO2 product (OMSO2e_003), which grid retrievals to 0.25∘ resolution (Krotkov et al., 2017; Li et al., 2013). Both products are cloud-screened; only pixels that are at least 70 % cloud-free are included in the NO2 product, and those that are at least 80 % cloud-free are included in the SO2 product. The NO2 product relies on air mass factors (AMFs) calculated with the assistance of an atmospheric chemical transport model and are sensitive to model representations of emission, chemistry, and transport data. Instead of AMFs, the SO2 product uses spectrally dependent SO2 Jacobians but can be interpreted as having a fixed AMF that is representative of summertime conditions. We applied basic transient SO2 plume filtering, excluding retrievals with SO2 > 15 DU (Wang and Wang, 2020).

Because our trend analysis uses a seasonal mean as the response variable, we assume that random errors cancel out, leaving only systematic errors, which do not contribute to uncertainty in the trend analysis. Systematic errors in the OMI NO2 product have an uncertainty of 20 % (McLinden et al., 2014) and are associated with AMFs and tropospheric vertical column contents. The OMI NO2 products use an implicit aerosol correction to account for the optical effects of aerosols, but retrievals can be biased when aerosol loading is extreme (Castellanos et al., 2015). Under these conditions, the OMI NO2 retrieval is biased low by roughly 20 % to 40 % (Chimot et al., 2016). Note that any aerosol-related error would have the potential effect of underestimating the magnitude of decreases in NO2 column densities when comparing 2020 to previous years. Additional bias in the NO2 product may be introduced due to the reliance on nearly cloud-free pixels, in which greater sunlight may induce higher photochemical rates. For example, the current NO2 product is biased roughly 30 % low over the Canadian oil sands (McLinden et al., 2014). The level-2 OMI-NO2 product has been validated against in situ and surface-based observations showing good agreement (Lamsal et al., 2014). The use of fixed Jacobians in the SO2 product introduces systematic errors of 50 % to 100 % for cloud-free observations (Krotkov et al., 2016).

Starting in 2007, the quality of level 1B radiance data for some OMI viewing directions has been affected, known as the row anomaly. The L3 products used here exclude all pixels affected by the row anomaly from each observation, but the locations of the row anomaly pixels were dynamic between 2007 and 2011, which could affect any comparisons including those years. Since 2011, the pixels affected by the row anomaly problem are the same, so comparisons for data only since 2011 are not affected by changes in the row anomaly.

Moderate Resolution Imaging Spectroradiometer (MODIS) sensors observe the Earth from polar orbit, from the Terra satellite since 2000 and from Aqua since mid-2002. In this study we use MODIS-derived AOD at 550 nm obtained by merging Dark Target and Deep Blue retrievals (Sayer et al., 2014). Specifically, we use the Deep_Blue_Aerosol_Optical_Depth_550_Land_Mean field over land and the over ocean AOD_550_Dark_Target_Deep_Blue_Combined_Mean the from Collection 6.1 L3 Gridded products MYD08 and MOD08 (Hubanks et al., 2019), though very few retrievals over ocean are included in our analysis. L3 values are computed on 1∘ × 1∘ spatial grid from L2 AOD products with resolution of 10 km × 10 km. Over land 66 % of MODIS-retrieved Dark Target AOD values were shown to be ± 0.05 ± 0.15*AOD AErosol RObotic NETwork (AERONET)-observed values, with high correlation (R = 0.9) (Levy et al., 2010). Around 78 % of the Deep Blue retrievals are within the expected error range of ± 0.05± 0.20*AOD (Sayer et al., 2013). MODIS AOD data have been extensively used by the modeling and remote sensing scientific communities and inter-compared with a wide range of satellite AOD products (see Schutgens et al., 2020, and references therein).

We analyzed these retrievals over two large regions (Fig. 1). Central east China was comprised of Shaanxi, Hubei, Anhui, Jiangsu, Shanxi, Henan, Hebei, Shandong, Beijing, and Tianjin provinces. Southern China was comprised of Guizhou, Guangxi, Hunan, Jiangxi, Guangdong, Fujian, and Zhejiang provinces. Daily mean quantities were calculated across all valid retrievals falling within the provinces comprising the regions. For the OMI NO2 columns, individual retrievals were weighted by the L3 “weight” field, which is proportional to the fraction of the grid cell with higher-quality retrievals, identified as those have less than 30 % cloud fraction and not affected by the row anomaly problem. We also calculated the daily value from the median of all retrievals to understand whether individual high values (mainly SO2) had any effect on the significance of trends or differences between 2020 and different background periods. Monthly averages were calculated from the daily regional averages, with each day weighted in the monthly average by the number of valid retrievals so as to not overrepresent days with little satellite coverage or significant cloud cover. The monthly data were used to visually identify COVID-19-related changes against background seasonality and trends since 2005.

We examined the difference in the distribution of daily data during the 23 January to 8 April 2020 lockdown period to the same period during previous years since 2005. We compared 2020 to 2019 to different background periods and to the expected value for 2020 estimated from trends over different background periods. Given the uneven nature of changes in atmospheric composition over different parts of China identified in previous studies, background periods were defined for each possible starting year between 2005 and 2018, with each ending in 2019. Retrieved quantities in 2020 were compared to the background means over each period and to the value expected for 2020 estimated from the linear trend over each period. We tested the significance of these differences using bootstrap resampling (Efron and Gong, 1983) with a resampling size of 2000.

We also considered how the analysis depended on how the lockdown period was defined. Emissions and pollution can decrease during the Chinese New Year holidays (Chen et al., 2020), which started as early as 23 January in 2012 and as late as 19 February in 2015, complicating COVID-19-related analyses of atmospheric composition over China (Bauwens et al., 2020; Chen et al., 2020). The timing and extent of lockdowns also varied between provinces and we assume that “slowdowns” could have happened before or after stricter, official lockdowns – for example, ground and air transportation remaining below lockdown levels nationally at least through 14 April 2020 (International Energy Agency, 2020). Excluding the holiday period from all years is a straightforward approach to excluding any New Year holiday effects but will exclude simultaneous lockdown effects during the initial, and presumably most strict, stages of the lockdown. Rather than specifying different combinations of New Year holiday period and provincial-level lockdown timing, we used 23 January–8 April as our baseline period (which will include all holiday periods since 2005) but examined the sensitivity of the statistics to the length of the lockdown period, namely a longer lockdown period beginning 1 week earlier and 1 week later, and a shorter lockdown period for February only. In interpreting the data, we put more confidence in 2020 differences that were insensitive to these choices.

The degree to which the COVID-19 lockdowns in China resulted in changes in atmospheric composition depended strongly on the background period and whether existing trends were taken into account. For AIRS CO over central east China, the 2020 mean was 3 %–12 % lower compared to different background periods. Relative to mean CO concentrations during periods beginning between 2005 and 2016, there were significant decreases in CO but CO in 2020 was not consistently different from what would be expected from trends calculated over this period. These longer-term declines in CO concentrations do appear to flatten out in recent years; assuming that the flat CO during 2017–2019 would have persisted, we estimate a 3 %–4 % reduction in CO in 2020 relative to that period. For MODIS AOD, the 2020 mean was between 14 % and 30 % less than different background averages but not significantly different from what would be expected for trends beginning between 2008 and 2014. As with CO trends, the negative AOD trends in the region also appear to flatten in recent years. Relative to the flat AOD over 2016–2019, 2020 AOD was 14 %–17 % lower than the background mean; as with CO, this range would be the more meaningful estimate of changes in 2020 if we assume that this flattening were to persist. The 2020 SO2 was significantly lower than background averages calculated over most periods, ranging from 83 % less than over 2005–2019 to 30 % less than over 2016–2019. Compared to the 2012–2019 period when there were no significant SO2 increases, 2020 SO2 was 200 % greater than what would be expected based on a trend starting in 2021, only 8 % greater when the median of daily retrievals was used, and not significantly different from expected relative to the trends beginning later than 2012. SO2 concentrations were relatively flat from 2016–2019; when using 2016 as the first year of the trend, SO2 was significantly higher than the expected value when calculated from the mean of the daily SO2 retrievals but not significantly different when calculated from the median. We note also that analyses of SO2 and NO2 that include years prior to 2012 may be affected by changes in observation sample size due to changes in the OMI row anomaly.

OMI NO2 in 2020 over central east China was consistently lower than the background average and expected value from the trends. There was a 17 % decrease in 2020 relative to the value expected from a trend calculated over 2011–2019 but a 30 %–33 % decrease relative to the different background means since 2016 when the NO2 was relatively flat. Again assuming that this flattening were to persist, this latter range may be the more meaningful baseline for the 2020 decrease. For reference, Bauwens et al. (2020) reported a ∼ 40 % drop in OMI NO2 from 2019 to 2020 over cities affected by the lockdown using the QA4ECV retrieval (Boersma et al., 2018) and a ∼ 51 % drop in NO2 over the eight cities (Beijing, Jinan, Nanjing, Qingdao, Tianjin, Wuhan, Xi'an, and Zhengzhou) falling within our central east China region. Our analysis cannot be compared directly because we include non-urban areas and define the lockdown period differently, but we do note that NO2 during the same period in 2019 appeared to be anomalously high relative to the previous few years, which would make the decreases in 2020 appear more significant.

The modest decreases in CO and AOD over central east China were unexpected; given its high population density and level of industrial activity, lockdowns may have been anticipated to lead to larger decreases. In the case of MODIS AOD, these modest decreases were possibly due to contributions from other sources unaffected by COVID-19-related lockdowns – limitations in the MODIS AOD retrieval under cloudy conditions, climatological variability from other sources such as mineral dust, and meteorology favorable to secondary aerosol formation which could have offset lower emissions (Wang et al., 2020). The 2020 increase in SO2 is more difficult to interpret because of the discrepancies between daily values calculated from the mean or median of individual retrievals but is broadly consistent with surface observations that find no significant change in in situ surface SO2 over Wuhan in the daily mean and a slight increase in daytime SO2 possibly associated with increased residential heating and cooking (Shi and Brasseur, 2020).

Over southern China, retrieved 2020 SO2 was significantly lower than the background average only for periods beginning between 2005 and 2011. Significant departures from expected trends were uneven when using the mean value of daily retrievals and absent when using the median value. As with central east China, we conclude that no significant changes could be robustly detected in 2020 SO2. NO2 in 2020 was between 23 % and 32 % less than the background average for different periods. Here, the flattening in NO2 beginning in 2013 is easier to identify than over central east China because of the much higher NO2 during the 3 years prior; 2020 NO2 was 23 %–27 % less than different background means between 2013 and 2019. The more significant reductions in NO2 in central east China compared to the south is presumably due the former's greater population and industrialization and consequently higher pollution levels. This is consistent with Chen et al.'s (2020) detection of a larger 2020 decrease in surface NO2 in Wuhan compared to Shanghai. Retrieved CO in 2020 was between 4 % and 8 % greater than background averages beginning in 2014 but between 11 % and 19 % higher than what would be expected given the decreasing trends over any period. AOD in 2020 was lower than background averages calculated starting with years earlier than 2012 but higher or not significantly different from expected for trends calculated starting in years after 2012.

The focus of this analysis is on whether satellite retrievals of atmospheric composition over 2020 departed significantly from different background periods and expected values for 2020 when daily variability and trends are accounted for, but it is useful at a preliminary stage to speculate as to how different emission changes could have contributed to (1) why NO2 was robustly lower in 2020 over central east China compared to CO and AOD and (2) why CO and perhaps AOD were higher over southern China compared to what would be expected from recent trends.

To understand why NO2 differences over central east China were more significant than other quantities, Table 3 shows the emissions by sector for a representative set of constituents from the Community Emissions Data System (CEDS) (Hoesly et al., 2018) over China for 2014, the most recent year available. Other bottom-up emission inventories will vary in absolute emission amounts and their sector contributions, particularly for more recent periods, but CEDS is the standard available emission dataset available globally as a baseline for the next Intergovernmental Panel on Climate Change (IPCC) assessment, in anticipation of assessing 2020 COVID-19-related changes to atmospheric composition in other regions, and for modeling studies involving a transboundary transport component. Across all species, energy production, industrial activity, transportation, residential/commercial/other (RCO), and waste disposal constitute the bulk of the emissions. Based on activity data for the first quarter of 2020, energy demand across China declined by 7 % compared to 2019, and transportation sector activity declined by 50 % to 75 % in regions with lockdowns in place (International Energy Agency, 2020). These sectors are direct or indirect sources of numerous pollutants, including SO2 (the precursor of sulfate aerosol), NOx, CO, and primary anthropogenic aerosols classified broadly as organic carbon (OC) and black carbon (BC). If we apply the 7 % reduction in energy production and midpoint 62.5 % reduction to transportation from the International Energy Agency (IEA) and assume a 20 % reduction in industrial emissions, 5 % reduction in waste emissions, and no change in RCO (with commercial decreases offset by residential increases), this yields a 10 % reduction in BC, 5 % reduction in OC, 14 % reduction in SO2, 14 % reduction in CO, and 21 % reduction in NO2. The larger reduction in NO2 relative to other emissions could partly explain why OMI NO2 column density changes over central east China were stronger than in the other retrievals.

Following Si et al.'s (2019) consideration of biomass burning as a pollution source in China alongside anthropogenic sources, we considered transboundary smoke transport as a possible reason for the higher 2020 CO over southern China, guided by higher CO over the upper Mekong region in 2020 compared to 2019 (Fig. 2a) and the predominant westerly flow during this time of year (Reid et al., 2013). Table 4 compares 23 January–8 April AIRS CO over southern China to CO emission estimates from biomass burning from the Global Fire Assimilation System (GFAS) (Kaiser et al., 2012) over the upper Mekong region (17 to 25∘ N, 95 to 105∘ E) including parts of eastern Myanmar, northern Thailand, and northern Laos. From 2005 to 2020, variation in GFAS CO over this region explained a moderate (32 %) amount of variability in AIRS CO over southern China, suggesting it is a non-negligible contributor to variation in CO concentration and a contributor to higher CO in 2020. This illustrates that, at a minimum, sources such as biomass burning smoke and dust that are less affected by COVID-19-related measures will complicate attribution studies. To that end, modeling studies following Wang et al. (2020) will be required to isolate emissions, meteorological and chemical drivers of changes in atmospheric composition, and their effects at a process level. With proper instrument-equivalent comparisons, modeling studies will also help to identify the extent to which the lack of significant changes are due to retrieval limitations, namely low sensitivity near the surface where differences would presumably be more pronounced, particularly given remote emission sources such as dust, biomass burning smoke, and volcanic SO2, which will arrive at higher altitudes.

The key implication of our study is that interpreting differences in 2020 retrievals of atmospheric composition depends strongly on how the background period is defined and whether trends over these periods are accounted for. Not taking these into account could lead to misattribution of changes in air quality to COVID-19 lockdowns. At a minimum, whether differences in 2020 are significant depends on the choice of background period, which is somewhat subjective. Leading up to 2020, there was an apparent flattening of decreasing trends beginning earlier in the decade across the retrievals; the considerable variability in the data made identifying this flattening easier in some cases than in others. We are more confident, for example, in our estimate of a 23 %–27 % decrease in 2020 NO2 over southern China relative to a flat background period than the 30 %–33 % decrease over central east China, where the recent variability was greater and the flattening less apparent. Revisiting this type of analysis in the years when regional economies have fully recovered post COVID-19 will help to distinguish between further decreases and flat trends and will lend themselves to using non-linear models in estimating the trends. We have approached the issue by comparing data for 2020 to what would have been expected given recent trends and by applying a single lockdown period to two large regions, with additional analyses to gauge the sensitivity of the 2020 differences to these choices. Other studies over China or elsewhere will inevitably use other approaches that more explicitly account for seasonality and meteorology, and which relate changes in pollution over smaller areas (e.g., single provinces or states) to region-specific lockdown measures and timing at a process level. Regardless of the approach, however, it is important to consider recent trends and variability. In places where pollution has decreased, not accounting for recent context could result in over-attribution of changes in pollution to COVID-19. In places where pollution has increased, such as parts of South Asia, this could result in under-attribution.