Our analysis builds on the recent development of unprecedented public access to air pollution model output and air quality observations in near-real time. We compile an air quality dataset of hourly surface observations for a total of 5756 sites (4778 for NO2 and 4463 for O3) in 46 countries for the time period 1 January 2018 to 1 July 2020, as summarized in Fig. 1 and Table 1. More detailed maps of the spatial distribution of observation sites over China, Europe, and North America are given in Figs. A1–A3. The vast majority of the observations were obtained from the OpenAQ platform and the air quality data portal of the European Environment Agency (EEA). Both platforms provide harmonized air quality observations in near-real time, greatly facilitating the analysis of otherwise disparate data sources. For the EEA observations, we use the validated data (E1a) for the years 2018–2019 and revert to the real-time data (E2a) for 2020. For Japan, we obtained hourly surface observations for a total of 225 sites in Hokkaido, Osaka, and Tokyo from the Atmospheric Environmental Regional Observation System (AEROS) (MOE, 2020). To improve data coverage in under-sampled regions, we also included observations from the cities of Rio de Janeiro (Brazil), Quito (Ecuador), and Melbourne (Australia). All cities offer continuous, hourly observations of NO2 and O3 over the full analysis period, thus offering an excellent snapshot of air quality at these locations. We include all sites with at least 365 d of observations between 1 January 2018 and 31 December 2019 and an overall data coverage of 75 % or more since the first day of availability. Only days with at least 12 h of valid data are included in the analysis. The final NO2 and O3 dataset comprise 8.9×107 and 8.2×107 hourly observations, respectively.

Meteorological and atmospheric chemistry information at each of the air quality observation sites is obtained from the NASA Goddard Earth Observing System Composition Forecast (GEOS-CF) model (Keller et al., 2020). GEOS-CF integrates the GEOS-Chem atmospheric chemistry model (v12-01) into the GEOS Earth System Model (Long et al., 2015; Hu et al., 2018) and provides global hourly analyses of atmospheric composition at 25×25 km2 spatial resolution, available in near-real time at https://gmao.gsfc.nasa.gov/weather_prediction/GEOS-CF/data_access/, last access: 5 July 2020 (Knowland et al., 2020). Anthropogenic emissions are prescribed using monthly Hemispheric Transport of Air Pollution (HTAP) bottom-up emissions (Janssens-Maenhout et al., 2015), with imposed weekly and diurnal scale factors as described in Keller et al. (2020). The same anthropogenic base emissions are used for the years 2018–2020. Therefore, GEOS-CF does not account for any anthropogenic emission changes since 2018, notably any anthropogenic emission reductions related to COVID-19 restrictions. However, it does capture the variability in natural emissions such as wildfires (based on the Quick Fire Emissions Dataset, QFED) (Darmenov and Da Silva, 2015) or lightning and biogenic emissions (Keller et al., 2014). While the meteorology and stratospheric ozone in GEOS-CF are fully constrained by pre-computed analysis fields produced by other GEOS systems (Lucchesi, 2015; Wargan et al., 2015), no trace gas observations are directly assimilated into the current version of GEOS-CF. It thus provides a “business-as-usual” estimate of NO2 and O3 that can be used as a baseline for input into the meteorological normalization process.

To support interpretation and guide visualizations, we include approximate national lockdown dates in all figures. The start and end dates for these are from https://en.wikipedia.org/wiki/COVID-19_pandemic_lockdowns, last access: 1 July 2020 (as of 1 July 2020) or based on local knowledge, with the full list of start and end dates given in Table A2. It should be noted that in many countries lockdown policy varied regionally and that many locations enacted “soft” stay-at-home orders before the official lockdowns. Human behavior is therefore expected to have changed considerably in many locations before the official lockdowns went into force.

Figure 4 shows the weekly mean observations of NO2 concentration, the GEOS-CF estimate, and the BCM prediction based on the machine learning predictor trained on 2018–2019 for the five cities of Wuhan (China), Taipei (Taiwan), Milan (Italy), New York (USA), and Rio de Janeiro (Brazil) from January 2018 through June 2020. We choose these five cities for illustration as they represent a diverse level of socio-economic development and due to the cities' variable responses to the COVID-19 pandemic. These five cities are also illustrative of the varying quality of the uncorrected GEOS-CF predictions compared to the observations. For example, as shown by the dashed grey lines vs. the solid black lines in Fig. 4, the uncorrected model predictions are in good agreement with observations in Rio de Janeiro but underestimate the observed NO2 concentrations in Taipei and Milan while overestimating concentrations over New York. These differences are a combination of the observation–model scale mismatch (25×25 km2 vs. point observation) and model errors, such as the simulated spatiotemporal distribution of NOx emissions or the modeling of the local boundary layer. The model–observation mismatch is particularly pronounced for Wuhan, where the model does not capture the observed seasonal cycle, pointing to errors in the imposed seasonal cycle of NOx emissions in the model.

In contrast to the uncorrected model predictions, the BCM closely follows the observations for years 2018 and 2019 (dashed black lines in Fig. 4). The grey region in Fig. 4 shows the start and end of the implementation of COVID-19 containment measures. Once containment is implemented, observed concentrations start to diverge from the BCM prediction for Wuhan, Milan, and New York (Fig. 4). For Wuhan, we find a reduction in NO2 of 54 (48–59) % relative to the expected BCM value for February and March 2020, and average decreases of 30 %–40 % are found over Milan (24 %–43 %) and New York (20 %–34 %) starting in mid-March and lasting through April (Fig. 4; Tables A3–A5). For cities where restrictions have been mainly removed (Wuhan, Milan) concentrations rise back towards the BCM value, although the concentrations in neither city have been fully restored to what might be expected based on the business-as-usual GEOS-CF simulation.

Looking more broadly at cities around the globe, 53 of the 64 specifically analyzed cities feature NO2 reductions of between 20 %–50 % (Figs. A6–A8 and Tables A3–A5). Most locations issued social distancing recommendations prior to the legal lockdowns and observed NO2 declines often precede the official lockdown date by 7–14 d (e.g., Brussels, London, Boston, Phoenix, and Washington, D.C.).

For Taipei and Rio de Janeiro, the observations and the BCM show little difference (Fig. 4), consistent with the less stringent quarantine measures in these places. Other cities with only short-term NO2 reductions of less than 25 % include Atlanta (USA), Prague (Czech Republic), and Melbourne (Australia), again fitting with the comparatively relaxed containment measures in these places (Figs. A6–A8). In contrast, Tokyo (Japan) and Stockholm (Sweden), which also implemented a less aggressive COVID-19 response, exhibit NO2 reductions comparable to those of cities with official lockdowns (>20 %), suggesting that economic and human activities were similarly subdued in those cities.

Substantial differences exist between cities in South America, with Rio de Janeiro and Santiago de Chile showing little change over the analyzed period, whereas Quito (Ecuador) and Medellín (Colombia) experienced a greater than 50 % reduction in NO2 after the initiation of strict restrictions measures in mid-March (Fig. A8 and Table A5). Concentrations in Medellín rebounded sharply in April and May, while concentrations in Quito remained 55 (52–58) % below business as usual throughout May and only started to return back to normal in June.

To evaluate the large-scale impact of COVID-19 restrictions on air quality, we aggregate the individual observation–model comparisons by country. We note that our estimates for some countries (e.g., Brazil, Colombia) are based on a single city and likely not representative of the whole country. On a country level, we find the sharpest and earliest drop in NO2 over China, where observed concentrations fell, on average, 55 (51–59) % below their expected value in early February when restrictions were implemented (Fig. 5). Concentrations remained at this level until late February, at which point they started to increase until restrictions were significantly relaxed in early April. Our analysis suggests that Chinese NO2 concentrations have recovered to within 5 (1–9) % of the business-as-usual values since then. For 2019 (dashed line in Fig. 5) the BCM shows a reduction in NO2 concentrations around Chinese New Year (5 February 2019), and it is likely that some reduction around the equivalent 2020 period (25 January 2020) would have occurred anyway. However, the 2020 reductions are significantly larger and more prolonged than in 2019. Similar to China, India shows large reductions in NO2 concentration (58 (49–67) %) coinciding with the implementation of restrictions in mid-March (Fig. 5); however, NO2 concentrations have not yet recovered by the end of June, reflecting the prolonged duration of lockdown measures. Other areas of Asia, such as Hong Kong and Taipei, implemented smaller restrictions than China or India and they show significantly smaller decreases (less than 20 %).

For Europe and the United States, we find widespread NO2 reductions averaging 22 (19–25) % in March and 33 (30–36) % in April (Fig. 5). In some countries, recovery is evident as lockdown restrictions are removed or lessened (e.g., Greece, Romania), but in 29 out of 36 countries concentrations remain 20 % or more below the business-as-usual scenario throughout May and June.

We follow the same methods for developing a business-as-usual counterfactual for O3 as we did for NO2 in Sect. 3.1. Any change in local O3 concentration arising from COVID-19 restrictions is set against a large seasonal increase in (background) concentrations in the Northern Hemisphere springtime (Fig. 6). Due to the longer atmospheric lifetime of O3 compared to NO2, the local O3 signal is expected to be comparatively small. This makes attributing changes in O3 concentration more challenging than for NO2. Our analysis shows an O3 increase of up to 50 % for some periods in cities with large NO2 reductions (e.g., Wuhan, Milan, Quito; Figs. 3 and A9–A11), but there is much less convincing evidence for a systematic O3 response across cities or on a regional level (Fig. 7). For example, our analysis shows little O3 difference in Beijing and Madrid during lockdown despite NO2 declines comparable to Wuhan or Milan (Figs. A9–A11). O3 enhancements of up to 20 % are found over Europe (e.g., Belgium, Luxembourg, Serbia), with a peak in early April, approximately 2 weeks after lockdown started (Fig. 7).

The analysis of O3 is complicated by its nonlinear chemical response to NOx emissions. In the presence of sunlight, O3 is produced chemically from the oxidation of volatile organic compounds in the presence of NOx (Seinfeld and Pandis, 2016). Therefore, a decline in NOx emissions could decrease O3 production and thus suppress O3 concentrations. On the other hand, the process of NOx titration, in which freshly emitted NO rapidly reacts with O3 to form NO2, acts as a sink for O3 (Seinfeld and Pandis, 2016). Odd oxygen (Ox) is conserved when O3 reacts with NO and thus offers a tool for separating these competing processes. Figure 8 presents the global mean diurnal cycle for O3 and Ox for the 5-month period since 1 February 2020 for both the observations and the BCM model, based on the individual hourly predictions at each observation site aggregated by local hour. The analysis of O3 and Ox is based on the same set of observation sites where both NO2 and O3 observations are available (see Fig. 1). Compared to the BCM model, there has been an increase in the concentration of nighttime O3 (00:00–05:00 LT, local time, Fig. 8a) by 1 ppbv (ppbv = nmol mol-1) compared to the BCM, whereas Ox shows a decrease of 1 ppbv (Fig. 8b). While these changes are small in magnitude, they represent a multi-month aggregate over 3485 observation sites that are statistically significant at the 1 % confidence interval. It should be noted that the biases of the machine learning models show little diurnal variability (Figs. A12–13), suggesting that this result is not caused by poor model performance during specific times of the day.

Our results indicate that during the night reduced NO emissions led to a reduction in O3 titration, allowing O3 concentrations to increase. During the afternoon, we find that O3 concentrations are lower by 1 ppbv (Fig. 8a), while observed Ox concentrations are lower than the baseline model by almost 2 ppbv at 14:00 LT (Fig. 8b). We attribute the lower Ox to reduced net Ox production due to the lower NOx concentration, but as titration is also reduced, daytime O3 concentrations change little. Overall changes to mean O3 concentrations are small, but there is a flattening of the diurnal cycle.

As shown in the lower panels in Fig. 8, both factors – enhanced nighttime O3 and reduced daytime Ox – are more pronounced at locations where preexisting NO2 concentrations are high (>15 ppbv). This suggests that the observed O3 deviations from the BCM are indeed coupled to NOx reductions due to COVID-19 restrictions, given that those are most pronounced at polluted sites.

The NO2 analysis presented in Sect. 3.1 implies a stark reduction in NOx emissions. However, due to the impact of atmospheric chemistry, changes in NO2 concentrations do not reflect the same relative change in NOx emissions. Because of this, the NO2 / NOx ratio and the NOx lifetime, both of which depend on seasonality and the local chemical environment, need to be taken into account when inferring NOx emissions from NO2 concentrations (Lamsal et al., 2011; Shah et al., 2020). To estimate the relationship between changes in NOx emission and changes in NO2 concentrations, we conducted a sensitivity simulation for the time period 1 December 2019 to 8 June 2020 using the GEOS-CF model with perturbed anthropogenic emissions. The perturbation simulation uses anthropogenic NOx emissions scaled based on adjustment factors derived from NO2 tropospheric columns observed by the NASA OMI instrument (Boersma et al., 2011). Daily scale factors were computed by normalizing coarse-resolution (2×2.5∘) 14 d NO2 tropospheric column moving averages by the corresponding moving average for year 2018 (the emissions base year in GEOS-CF; Sect. 2.2). Forest fire signals were filtered out based on QFED emissions and no scaling was applied over water. This results in anthropogenic emission adjustment factors of 0.3 to 1.4 (Fig. A14), comparable to the magnitude obtained from the observation–BCM comparisons at cities globally (Fig. 5) and capturing the range of expected NOx emission changes. However, it should be noted that the scale factors do not necessarily coincide in space and time with the ones derived from observations and the BCM, and they do not include any adjustment for the NO2 / NOx ratio.

Figure 9a shows the response of NO2 surface concentration to a change in NOx emissions derived from the comparison of the sensitivity experiment against the GEOS-CF reference simulation. Our results indicate that NO2 concentrations drop, on average, by 80 % of the fractional decrease in anthropogenic NOx emission, with a further diminishing effect for emission reductions greater than 50 %. This reflects both the buffering effect of atmospheric chemistry and the presence of natural background NO2. The here-derived average sensitivity of 0.8 between a change in surface NO2 to a change in NOx emissions is comparable to the value of 0.86 (1/1.16) obtained by Lamsal et al. (2011) for the relationship between NOx emissions and tropospheric column NO2 observations.

To infer the reduction in anthropogenic NOx emissions due to COVID-19 containment measures during the first 6 months of 2020, we use the best linear fit between the simulated NOx / NO2 sensitivity (dashed purple line in Fig. 9a). To do so, we calculate the monthly percentage emission change at each observation site based on the NO2 anomalies derived in Sect. 3.1 and the corresponding best fit NOx / NO2 sensitivity (Fig. 9a). This is a simplification, as the local NOx / NO2 sensitivity ratio is highly dependent on the local environment. To account for this uncertainty, we assign an absolute error of 15 % to our NOx / NO2 sensitivity, as derived from the spread in the NOx / NO2 ratio in the sensitivity simulation (Fig. 9a). We then aggregate these estimates to a country level by weighting them based on average NO2 concentrations per location, thus giving higher weight to locations with more nearby NOx emission sources. It should be noted that for some countries our estimates are based upon a small number of observation sites that might not be representative for the country as a whole. This is particularly true for India and Brazil, where less than 10 observation sites are available. While the smaller observation sample size is reflected in the wider uncertainty associated with these emission estimates compared to countries with a much denser monitoring network (e.g., China or Europe), the applied extrapolation method might incur errors that are not reflected in the stated uncertainty ranges.

To obtain absolute estimates in emission changes, the monthly country-level percentage emission changes are convoluted with bottom-up emissions estimates for 2015 from the Emission Database for Global Atmospheric Research (EDGAR v5.0_AP; Crippa et al., 2018, 2020). The choice of EDGAR v5.0 as the bottom-up reference inventory (over, e.g., the HTAP emissions inventory used in GEOS-CF) was motivated by the fact that its baseline has been updated more recently and the country emission totals – which our analysis is based on – are readily available.

As summarized in Table 2, we calculate that the total reduction in anthropogenic NOx emissions due to COVID-19 containment measures during the first 6 months of 2020 amounted to 3.1 (2.6–3.6) TgN (Fig. 9b and Table 2). This is equivalent to 5.5 (4.7–6.4) % of global annual anthropogenic NOx emissions (Table 2). Our estimate encompasses 46 countries that together account for 67 % of the total emissions (excluding international shipping and aviation). We have no information for significant countries such as Russia, Indonesia, or anywhere in Africa due to the lack of publicly available near-real-time air quality information. China accounts for the largest fraction of the total deduced emission reductions (28 %), followed by India (25 %), the United States (18 %), and Europe (12 %).

While our method does not allow for sector-specific emission attribution, we assume our results to be most representative for changes in traffic emissions (rather than say aircraft emissions) given the location of the observation sites. On average, traffic emissions represent 27 % of total anthropogenic NOx emissions (Crippa et al., 2018), and our derived total NOx emission reduction from January–June 2020 corresponds to 21 (17–24) % of global annual traffic emissions. The share of transportation on total NOx emissions is higher in the US and Europe (approx. 40 %) compared to India and China (20 %–25 %). Taking this into account, the derived ratio of NOx emission reductions to annual traffic emissions is 21 (16–26) % in the US, 25 (20–30) % in Europe, 39 (34–44) % in China, and 62 (55–69) % in India.

The response of O3 to NO2 declines in the wake of the COVID-19 outbreak is complicated by the competing influences of atmospheric chemistry. From February through June 2020, the diurnal observation-BCM comparisons suggest that the reduction in photochemical production was offset by a smaller loss from titration, as described in Sect. 3.2. This resulted in a flattening of the diurnal cycle and an insignificant net change in surface O3 over a diurnal cycle. The competing impacts of reduced NOx emissions on O3 production and loss are dependent on the local chemical and meteorological environment. This is reflected in the variable geographical response of O3 following the implementation of COVID-19 restrictions (Le et al., 2020; Dantas et al., 2020). Moreover, as atmospheric reactivity increases through the Northern Hemisphere spring and summer, the relative importance of photochemical production is expected to increase in the Northern Hemisphere.

To assess the potential seasonal-scale impact of reduced anthropogenic emissions on O3, we conducted two free-running forecast simulations between 8 June and 31 August 2020, initialized from the GEOS-CF simulation and the sensitivity simulation described in Sect. 3.3, respectively. Both simulations use the same biomass burning emissions based on a historical QFED climatology. For the forecast sensitivity experiment, we assume a sustained, time-invariant 20 % reduction in global anthropogenic emissions of NOx, carbon monoxide (CO), and VOCs. We chose to alter not only the anthropogenic emissions of NOx but also other pollutants whose anthropogenic emissions are highly correlated to NOx, as a reduction in NOx emissions without corresponding declines in CO and VOC emissions seems unrealistic.

Figure 10 shows the differences between the reference forecast and the sensitivity simulation over the United States, Europe, and China. Our results indicate that sustained lower anthropogenic emissions lead to a general decrease in surface O3 concentrations of 10 %–20 % over eastern China, Europe, and the western and northeastern US during July and August relative to the business-as-usual reference forecast simulation. However, it is also notable that in some locations the model forecast O3 concentrations increase by an equivalent amount (e.g., Scandinavia, southern central US and Mexico, northern India), reflecting the high nonlinearity of atmospheric chemistry. This highlights the complex interactions between emissions, chemistry, and meteorology and their impact on air pollution on different time scales.