Diagnosing air quality changes in the UK during the COVID-19 lockdown using TROPOMI and GEOS-Chem

The dramatic and sudden reduction in anthropogenic activity due to lockdown measures in the UK in response to the COVID-19 outbreak has resulted in a concerted effort to estimate local and regional changes in air quality, though changes in underlying emissions remain uncertain. Here we combine satellite observations of tropospheric NO2 from TROPOspheric Monitoring Instrument and the Goddard Earth Observing System (GEOS)-Chem 3D chemical transport model to estimate that NO x emissions declined nationwide by ∼20% during the lockdown (23 March to 31 May 2020). Regionally, these range from 22% to 23% in the western portion of the country to 29% in the southeast and Manchester, and >40% in London. We apply a uniform 20% lockdown period emission reduction to GEOS-Chem anthropogenic emissions over the UK to determine that decline in lockdown emissions led to a national decline in PM2.5 of 1.1 μg m−3, ranging from 0.6 μg m−3 in Scotland to 2 μg m−3 in the southwest. The decline in emissions in cities (>40%) is greater than the national average and causes an increase in ozone of ∼2 ppbv in London and Manchester. The change in ozone and PM2.5 concentrations due to emission reductions alone is about half the total change from 2019 to 2020. This emphasizes the need to account for emissions and other factors, in particular meteorology, in future air pollution abatement strategies and regulatory action.


Introduction
The first cases of COVID-19 were reported in December 2019 in Wuhan, China, the initial epicentre of the outbreak, and rapidly spread to the US and Europe. By mid-March, all European countries had confirmed cases, leading to the implementation of travel restrictions and local, regional, or national lockdown measures. The UK government imposed strict lockdown measures on 23 March 2020 in response to model simulations that showed that the rate of spread of the virus would overwhelm the healthcare system [1].
The movement restrictions imposed by the UK and other governments led to an unprecedented decline in road traffic and slowdown or shutdown of intensive industries. The associated reduction in air pollutant emissions has led to a flurry of research to determine the short-term response in air quality and atmospheric composition. A contemporary record of these is maintained by the Health Effects Institute (https://tinyurl.com/y3xyty6x). The general pattern that emerges from these studies comparing air pollutant concentrations before and after lockdown measures is widespread regional decline in nitrogen oxide (NO x ≡ NO + NO 2 ) concentrations. Also observed is the anticipated increase in ozone due to reduced titration of ozone by NO x in cities where this titration is prevalent [2][3][4][5][6][7][8][9][10][11][12]. Studies that use models to interpret air quality observations have also diagnosed the contribution of meteorology to the observed changes in atmospheric composition and air quality [13], estimated the influence of emission changes on the oxidizing capacity of the atmosphere [5], and quantified regional decline in emissions due to lockdown measures [14]. Across China, for example, NO 2 concentrations declined by 40%-60% [2][3][4], whereas PM 2.5 (particles with aerodynamic diameter < 2.5 µm) increased in heavily populated and very polluted northern China. This was due to a combination of stagnant conditions, unusually high humidity promoting heterogeneous chemistry, sustained industrial emissions, and an increase in the oxidizing capacity of the atmosphere [5]. In India, an increase in fire activity in 2020 masked the anticipated decline in NO x due to lockdown measures in cities close to these fires [15].
In the UK, preliminary reports on changes in air quality were commissioned by the national Department for the Environment, Food and Rural Affairs. These mostly compared surface network observations in 2020 before and during the lockdown. Reported changes in air pollutants included an average 21% decline in NO 2 from roadside reference network monitors [16,17], though this percent decline includes ordinary seasonal decrease in NO x due to changes in photochemistry during the shift from winter to spring [18]. A surprising change in air quality in London was an observed increase in PM 2.5 , despite the decline in NO x , that was attributed to longrange transport of PM 2.5 -laden air from mainland Europe [16]. The impact of the lockdown on air quality throughout the UK was later diagnosed by Lee et al [6] using surface data from 126 in situ measurement sites distributed across the UK. They found from comparison of observations averaged over 23 March to 31 May that surface NO 2 was 42% less and surface ozone 11% more in 2020 than the mean of the 5 preceding years. They speculated that routine exceedances of NO 2 air quality limits in London would halve in 2020 relative to 2019 as a result of the lockdown. They also identified a marked increase in ozone exceedances during the lockdown, but suggested that the influence of this on the rest of 2020 depends on emissions and meteorological conditions in summer when surface ozone concentrations in the UK peak [6].
Changes in air quality by comparison of the lockdown period to previous weeks or years will be impacted by interannual variability due to extreme weather or shifts in transport patterns, seasonality in photochemistry affecting formation and persistence of air pollution, and ongoing decline in emissions due to air quality policies. As a result, it is challenging to disentangle the contribution of changes in emissions and other factors to changes in air quality with observations alone. A more suitable approach is to use an explicit and detailed atmospheric chemical transport model (CTM) [10,13,14,19] or to account for confounding effects of weather using appropriate statistical methods [7,11,13,20]. The UK network of reference monitors, while relatively extensive compared to most countries, has large monitoring gaps, so may not be appropriate for diagnosing national and regional changes in air quality.
Here we use satellite observations of tropospheric column NO 2 from the TROPOspheric Monitoring Instrument (TROPOMI) to quantify national, regional and large city (London, Manchester) changes in NO 2 in the UK due to the lockdown. Analysis of the satellite observations is supported by observations of surface concentrations of NO 2 from the air quality networks in London and Manchester and interpreted with targeted Goddard Earth Observing System (GEOS)-Chem CTM simulations. The model is used to account for non-linearities in atmospheric chemistry driving changes in air quality [21], relate changes in the tropospheric column to changes in emissions at the surface, discern the role of meteorology and emission changes, and quantify the effect of lockdown measures on regional and national surface concentrations of the criteria pollutants ozone and PM 2.5 that adversely affect our health and the environment. 2 We use tropospheric NO 2 column densities from TROPOMI onboard the Sentinel-5 Precursor (S5P) satellite [22,23]. S5P was launched on 13 October 2017 and has an overpass time of 13:30 local solar time (LST). TROPOMI achieves daily global coverage with a swath width of 2600 km and a ground pixel resolution of 7.2 km × 3.5 km at nadir (along track × across track) until 5 August 2019, refined thereafter to 5.6 km × 3.5 km [24]. We use TROPOMI NO 2 over the UK from the offline Level 2 product (version 01-03-02) from the S5P Pre-Operations Data Hub (https://s5phub.copernicus.eu/dhus/; last accessed 20 July 2020) for 1 January to 30 June 2019 and 2020. TROPOMI data are filtered to remove poor quality retrievals with a quality flag ('qa_value' in the data file) >0.5 and cloudy scenes identified with a TROPOMI cloud radiance fraction ⩾0.5, as suggested by Eskes et al [25] for comparison of TROPOMI to models. TROPOMI validation studies generally find that TROPOMI overestimates NO 2 at rural and remote locations and underestimates NO 2 over polluted scenes from comparison to ground-based total and tropospheric column measurements [26][27][28][29][30].

Surface air quality measurements
Surface hourly in situ observations that we use to assess temporal variability of TROPOMI NO 2 in Manchester and London and the ability of the model to reproduce changes in surface concentrations of ozone and PM 2.5 across the UK are from the Automatic Urban and Rural Network (AURN). All data are downloaded from the UK-AIR data portal (https://uk-air.defra. gov.uk/networks/network-info?view=aurn; accessed 11 August 2020 for NO 2 and 25 September 2020 for ozone and PM 2.5 ). We use NO 2 data from eight sites in London and three sites in Manchester. Citywide average NO 2 is estimated for each city from observations obtained during the satellite overpass (13:00-14:00 LST). Ozone data are available from 71 monitors in both 2019 and 2020 and 80 monitors in 2019 and 81 in 2020 for PM 2.5 . These are used to calculate 24-hour averages at each site to compare to coincident GEOS-Chem data. The model also includes detailed gasand aerosol-phase chemistry and physical loss processes (wet and dry deposition). Emission inventories are updated in this work to include anthropogenic emissions over the UK and mainland Europe from the gridded European Monitoring and Evaluation Programme inventory for 2016 scaled to the simulation year of interest using reported rates of annual emission reductions [31]. The model also includes natural emissions from vegetation, soils, seabirds, lightning and volcanoes.

The GEOS-Chem CTM
The model is simulated in 2019 as a reference year in terms of anthropogenic emissions and in 2020 using five emission scenarios. These include business-as-usual (BAU) using normal annual decline in anthropogenic emissions due to air quality policies, and lockdown emission reduction scenarios that, in addition to the normal annual decline in emissions, include anthropogenic emission reductions of NO x , SO 2 , non-methane volatile organic compounds, and primary PM 2.5 in the UK and mainland Europe. We refer to these throughout as GC_BAU for the BAU scenario, GC_x for x% emission reduction, where x is 15, 30, 45, and 60. In the GC_x simulations, emissions reductions are applied to mainland Europe on 15 March and the UK on 23 March to represent the earlier lockdown measures across mainland Europe. Sectors that are reduced include industry, transport, and other residential and commercial activity. The instantaneous reduction in emissions in the model is approximate, as some countries gradually transitioned from local to national lockdown measures (such as in Italy). There was also a range in the timing of national lockdowns (17 March for France to 23 March for Germany) and severity of measures. The Netherlands and Greece, for example, did not impose a national lockdown [32]. The model is sampled from 1 February to 30 June for the reference year (2019) and the GC_BAU emission scenario. The GC_x simulatios are sampled from 15 March to 30 June 2020. Figure 1 shows TROPOMI monthly mean tropospheric NO 2 columns over the UK for 2019 and 2020. The decline in TROPOMI tropospheric column NO 2 from January to June is expected, due to seasonal variability in photochemistry leading to a longer lifetime and greater abundance of NO x in winter than in spring and summer. The location of cities with large traffic emissions of NO x , such as London in the southeast and Manchester in the northwest, are particularly apparent in winter 2019.

Changes in column densities of NO 2 in the UK
TROPOMI NO 2 over the UK is in general lower in 2020 than 2019, even for the months preceding the lockdown. Lower concentrations of NO 2 in January and February for 2020 compared to the mean of the five preceding years has also been observed for surface NO 2 and attributed to unusually high wind speeds and frequent storms in 2020 [6]. In April, the peak of the lockdown period, the relative decrease from 2019 to 2020 is 36 ± 13% (median ± standard error) for the SE, 31 ± 11% for the SW, 19 ± 11% for the NE, 38 ± 11% for the NW, 31 ± 11% for SC, 40 ± 12% for London, and 31 ± 14% for Manchester. In May, when there was a gradual transition to more lenient lockdown measures in the UK, there is generally a smaller reduction in TROPOMI NO 2 in 2020 relative to 2019 (29 ± 15% for the NW, 17 ± 11% for SC, and 31 ± 17% for London). The SE (36 ± 9%) and Manchester (34 ± 14%) show greater decline in May than April. The decline in the SW in May is not significant (20 ± 19%) and there is a non-significant increase in the NE of 9 ± 18%. Figure 2 assesses the skill of TROPOMI at reproducing day-to-day variability in surface NO 2 in London and Manchester which have been used in the study by Lee et al [6]. City-average daily mean surface observations are compared to TROPOMI daily mean tropospheric columns for TROPOMI pixel centres within the city domains shown in figure 1. The range in surface NO 2 concentrations is similar for the two cities (5-75 µg m −3 ), declining from winter to spring, as is expected from seasonal changes in photochemistry. There is also large day-to-day variability related to weather and weekend-weekday variability in traffic activity [34,35], though midday TROPOMI and surface observations will be less sensitive to this effect than observations during rush hour [36]. TROPOMI Individual daily TROPOMI measurements are gridded to 0.04 • × 0.04 • (∼4 km) using the gridding technique described in Pope et al [33]. In the lower right panel, black boxes indicate the sampling extent of the five regions (SC for Scotland, NW for northwest England and Wales, NE for northeast England, SW for southwest England and Wales, and SE for southeast England) and blue boxes the sampling extent of London and Manchester. NO 2 are sparse in winter due to persistent clouds. A relatively shallow planetary boundary layer in winter also leads to reduced sensitivity to surface NO 2 and degrades temporal consistency between the column and surface [37]. The surface in situ observations are also more prone to positive interference in winter from thermal decomposition of reservoir compounds to NO 2 [38]. Despite these issues, the temporal correlation between the surface and satellite observations is reasonably strong. For the full record shown in figure 2, the Pearson's correlation coefficient, R, is 0.78 for London and 0.51 for Manchester using a 75% spatial coverage threshold for TROPOMI. Similarly, during just the lockdown period and again using a 75% spatial coverage threshold, R is 0.71 for London and 0.70 for Manchester. The increase in R for Manchester during the lockdown period is because the pre-lockdown period has fewer coincident observations. The small decrease in R for London may be due to less dynamic variability in NO 2 during the lockdown period.

Interpretation of observed changes in NO 2 with GEOS-Chem
We use GEOS-Chem to determine the contribution of anthropogenic emissions changes associated with lockdown measures to the observed changes in TROPOMI NO 2 from 2019 to 2020. We do this by identifying the emission scenario in 2020 that yields results consistent with regression of GEOS-Chem versus TROPOMI tropospheric NO 2 columns in 2019. This approach is conceptually similar to the standard approach of estimating emissions by applying the modelled ratios of tropospheric column NO 2 to surface emissions of NO x to TROPOMI tropospheric NO 2 columns to derive NO x emissions [39] for two different time periods and estimating the relative change in satellite-derived emissions. Our regression approach eliminates unnecessary steps in the calculation and also minimizes influence of systematic biases in the model and observations. In all comparisons of the model to TROPOMI we apply the TROPOMI averaging kernels to the model to ensure the comparison is independent of the TRO-POMI a priori [25,40,41]. City and regional averages for the domains shown in figure 1 are obtained using the Level 2 pixels and coincident GEOS-Chem grids. The model is sampled during the satellite overpass (12:00-15:00 LST), as is standard. The difference in modelled fields obtained for a narrower sampling range (13:00-14:00 LST) is negligible. Figures 3 and 4 compare temporal variability in daily mean TROPOMI and GEOS-Chem regional (SE, figure 3) and city (London, figure 4) tropospheric columns. Only the results for the GC_BAU, GC_30, and GC_60 scenarios are shown, as the response of the model to emission changes is approximately linear. Comparisons for the other regions and Manchester for all scenarios are in supplementary figures S1-S5 (available online at stacks.iop.org/ERL/16/054031/mmedia). We use 2019 to quantify how the model compares to the satellite observations during a normal year. This accounts for systematic biases in the modelled tropospheric NO 2 column that may result from model representation of emissions, chemistry, and loss pathways, and also systematic biases in TROPOMI [26][27][28][29][30]. To do this, we regress TROPOMI columns against colocated GEOS-Chem columns for 2019. We find that the model underestimates variance in tropospheric column NO 2 . Regression slopes for 2019 are less than unity in all regions (0.70-0.94) and cities (0.70 for London, 0.84 for Manchester). This may be due to a model underestimate in free tropospheric NO 2 that will be pronounced in comparison to TROPOMI due to its enhanced vertical sensitivity to NO 2 in the free troposphere [42,43]. We would then expect that the 2020 model scenario that most closely matches air pollutant emissions during the lockdown is the one that yields a slope similar to that obtained with the 2019 data.
The slopes in figures 3 and 4 and S1-S5 for the GC_BAU scenario are 21%-49% larger than the reference year (2019) slopes everywhere except SC and the NE. In SC, GC_BAU is 7% less than 2019 and in the NE GC_BAU is consistent with (<1% difference) that for 2019, suggesting either no appreciable change in emissions during the lockdown or limited regional sensitivity of tropospheric columns to changes in surface emissions. Regional GEOS-Chem NO x emissions in SC in 23 March to 31 May 2019 are relatively low. NO x emissions in SC are 10% of national total emissions. This is similar to the contribution from just the London domain. NO x emissions are also limited to a few large cities such as Edinburgh and Glasgow (figure 1). Tropospheric columns of NO 2 over SC will be more influenced by factors such as long-range transport and meteorology than perturbations in local sources. NO x emissions in the NE are 20% of the national total, but this includes a large contribution from industrial activity and power generation. According to the UK Office of National Statistics, activity from these decreased less dramatically (20% decline in April 2020 relative to April 2019) than road traffic. Road traffic in April 2020 was 35% of typical conditions for April and 50% of typical conditions in May [45,46]. By interpolation of slopes for the GEOS-Chem 2020 emission scenarios, we estimate that emission reductions that would yield a slope that best matches the slope for 2019 include 42% for London, 29% for Manchester and the SE, 23% for the NW, and 22% for the SW. There is no detectable change for the NE and SC.  The bar chart in figure 6 shows the relative change in city and regional median TROPOMI and GEOS-Chem tropospheric NO 2 columns from 2019 to 2020 as an alternate approach to the regression method to identify the scale of emission reductions required to reproduce the observations. The decline in TRO-POMI NO 2 in 2020 relative to 2019 is 41% over London and 32% over Manchester. This is greater than the decline reported by Barré et al [20] for London (30%) and Manchester (27%), though within the margin of error for Manchester. They also used TROPOMI, but for a lockdown period that starts earlier than the UK lockdown (15 March) and so includes more days in March than us when 2019 and 2020 TROPOMI concentrations are relatively similar (figure 1). The regional decline we obtain with TRO-POMI ranges from 6% over the NE to 35% over the NW. The relative change in the model from 2020 to 2019 due to normal emission mitigation measures and meteorology (GC_BAU) ranges from +9% for London to −23% for SC. The emission reduction estimates obtained with the relative difference approach are 46% for London, 18% for Manchester, 15% for the SW, 32% for the SE, 33% for the NW, and BAU for the NE and SC. These are similar to the regression approach for London, the SE, the SW, the NE and SC, but less than the regression approach for Manchester and more than the regression approach for the NW, as the relative difference is more influenced by extreme values than the regression approach.
The size of the errors are relatively large in figure 6, so we use the two-sample t-test to determine whether the differences in TROPOMI and GEOS-Chem  tropospheric NO 2 columns for 2019 and 2020 are statistically significantly different. The t-scores and p-values from this analysis are in the supplementary (table S1). The model emission reduction scenarios for which the differences between GEOS-Chem and TROPOMI are inconsistent with those for 2019 have a t-score significantly different from zero (and pvalue < 0.05). The t-test confirms a 30% reduction for the NW and Manchester and no emissions reduction for the NE and SC, but yields larger reduction for London (GC_60). The results are ambiguous for the SW (GC_15 and GC_30) and the SE (GC_30 and GC_45). Figure 7 shows the complete GEOS-Chem time series of tropospheric column and surface NO 2 for 1 February to 31 May in 2019 and 2020 for the GC_BAU scenario and the emission scenario that best fits the observations over London according to the regression slope approach (GC_45). Time series for the SW, SE, NW and Manchester are in the supplementary ( Figures S6-S9). Modelled columns and surface NO 2 before the lockdown in 2020 are much less than NO 2 in the previous year, as was observed for TROPOMI NO 2 (figure 1). According to the model, this is due to buildup of pollution in 2019 due to stagnant conditions, and less precipitation and colder temperatures in 2019 than 2020. This is consistent with the interpretation by Lee et al [6] that lower pre-lockdown surface concentrations of NO 2 in 2020 than 2019 were caused by dissipation of pollution by unusually high wind speeds and frequent storms in early 2020.
The change in NO 2 concentrations due to a 45% reduction in emissions in figure 7 during the lockdown (GC_45 minus GC_BAU for 23 March to 31 May 2020) is 42% at the surface and for the column, supporting sensitivity of the column to changes in local emissions. There is evidence of influence of air pollution from mainland Europe, as GC_BAU and GC_45 surface and column NO 2 diverge after the mainland Europe lockdown and before the UK lockdown. This influence is apparent in all regions (figures S6-S8).

Lockdown impact on surface ozone and PM 2.5
We now use the model to determine the contribution of emission reductions due to lockdown measures to changes in ozone and PM 2.5 . We use results from a model simulation with a uniform 20% reduction in UK anthropogenic emissions, as this is representative of nationwide decline in NO x emissions (figure 5).
We first assess the ability of the model to reproduce AURN ozone and PM 2.5 in 2019 and the observed change in PM 2.5 and ozone in 2020 relative to 2019. AURN PM 2.5 are obtained at standard atmospheric pressure, 20 • C and 50% relatively humidity [47], so GEOS-Chem PM 2.5 is calculated for the same conditions. Additional details of the GEOS-Chem simulation of PM 2.5 components and calculation of total PM 2.5 is in the supporting information. The increase in modelled ozone is more consistent with AURN using a 45% emission reduction scenario (2.2 ± 1.9 ppbv), as the effect of NO x on ozone is local to cities where NO x emission reductions exceed 40%, according to our emission reduction estimates for London using TROPOMI and GEOS-Chem. PM 2.5 is less spatially heterogeneous than ozone, as its physical removal from the atmosphere (wet deposition) can take days to weeks, whereas persistence of ozone in cities is determined by its reaction with short-lived (seconds to minutes) NO. Modelled and observed PM 2.5 are similar in 2019 (13.5 µg m −3  figure S10). The model underestimates the observed 6-7 µg m −3 decline in PM 2.5 in cities, but captures the decline in rural and remote regions (figure S10). The aerosol nitrate component of PM 2.5 is from oxidation of NO x and the decline in emissions of these is constrained in this work with TROPOMI, whereas emissions of other precursors and primary PM 2.5 are not. The spatial distribution of the modelled decline in nitrate and associated ammonium and aerosol water is consistent with decline in AURN PM 2.5 in rural and remote areas (figure S11). The observed change in ozone and PM 2.5 in cities is, in general, greater than in the model using 20% emission reduction (figure S10), as the measurements are influenced by local effects not captured at the spatial resolution of the model and the regional analysis using TROPOMI does not capture local emission reductions in cities. Figure 9 shows the effect of 20% emission reductions alone on ozone and PM 2.5 . The national increase in ozone is small (0.24 ppbv) and ranges from increases in ozone over cities of 0.89 ppbv for Manchester and 1.0 ppbv for London to regional decline in ozone in remote and rural areas. The SE experiences the largest regional increase in ozone of 0.67 ppbv. A simulation with 45% lower emissions, likely more representative of the decline in emissions in cities, leads to an increase in surface ozone of 1.9 ppbv for Manchester and 2.1 ppbv for London. The national decline in PM 2.5 for 20% emission reduction is 1.1 µg m −3 , ranging from a decrease of Figure 9. Response of surface ozone and PM2.5%-20% reduction in anthropogenic emissions associated with lockdown measures. Maps are differences between BAU and a 20% reduction in anthropogenic emissions. Numbers inset give mean changes in ozone for target regions (boxes) and cities (arrows) and in PM2.5 for target regions only. 0.61 µg m −3 in SC to 2.0 µg m −3 in the SW. The largest response in the SW is due in part to decline in primary and secondary sources of PM 2.5 in the SE that would ordinarily be transported by prevailing southeasterlies. The response in simulated ozone and PM 2.5 averaged across the UK to emissions alone (figure 9) is about half the combined response of meteorology and emissions (figure S10).

Conclusion
We combined data from TROPOMI and the GEOS-Chem model to determine that lockdown measures in the UK in response to the rise in COVID-19 cases in March 2020 led to a nationwide decline in NO x emissions of ∼20%. These range from being indistinguishable from BAU in Scotland and northern England to >40% reduction in London. Emission changes only account for half the national mean increase in surface ozone and decline in PM 2.5 in 2020 relative to 2019. The remaining contribution is from meteorological differences between the 2 years, emphasizing the need to account for meteorology and pollution sources in future abatement strategies and in imposing regulatory action.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https://s5phub.copernicus.eu/dhus/#/home.