Nitrogen dioxide (NO2; together with NO, a constituent of NOx = NO + NO2) is a very well-established cause of poor air quality in the most urbanized and industrialized areas of the world. NO2 is harmful for living organisms by long-term atmospheric concentration exposure. It also plays a major role in urban ozone formation and secondary aerosols, which are also harmful for living organisms at high levels in the lower atmosphere (Lelieveld et al., 2015; IPCC, 2014). According to the European Environment Agency (EEA, 2020a) the main European anthropogenic NOx sources are road transport (39 %), energy production and distribution (16 %); commercial, residential and households (14 %); energy use in industry (12 %); agriculture (8 %); non-road transport (8 %); and industrial processes and product use (3 %). With an atmospheric lifetime typically below 1 d, NOx is relatively short-lived and is mainly controlled by photochemical reactions. The majority of NOx therefore does not get transported far downwind from its sources (Seinfeld and Pandis, 2006). Thus, near-surface NOx concentrations are high over cities and densely populated areas and low otherwise. Besides emissions, the variability in NOx is strongly driven by meteorological conditions, especially atmospheric transport, vertical mixing, and solar radiation, affecting the level of accumulation close to the emission sources (Arya, 1999). For example, increased wind speed and a higher planetary boundary layer height will increase the dispersion of NOx from the emission sources. It is this short lifetime, which is partly modulated by atmospheric conditions such as temperature and radiation combined with localized emission sources, that makes NO2 an excellent proxy for detecting emission reductions, from both surface and satellite measurements.

The worldwide outbreak of the coronavirus disease (COVID-19), which arose in late 2019 in China and spread around the world in early 2020, led many countries to take action to slow down the infection growth rate of the virus. The so-called lockdowns severely restricted or banned movements of people, closed most public places. and limited journeys to essential work commutes. Some measures started in China in late 2019, with stricter lockdowns in January 2020. In Europe, lockdown measures were implemented on various dates during February and March 2020. These lockdowns drastically reduced traffic and also activity levels in most industries (Guevara et al., 2021; Le Quéré et al., 2020). These sectors represent a large share of NOx emissions (51 % according to EEA, 2020a). Studying NO2 concentration changes during the lockdown is therefore very important to assess the impact of such activity-level reductions on a population's exposure to pollution. The COVID-19 lockdown is a unique opportunity to assess the impact of future pollution reduction measures, in particular, the impact of drastic reductions on the road transport sector using combustion energy.

The lockdowns were expected to have large effects on urban NO2 air pollution levels in conjunction with other modulating factors (i.e. weather conditions). The first quarter of 2020 had specific and highly variable meteorological conditions. Storm Ciara crossed over Europe in the second week of February, followed by Storm Dennis, which crossed Europe a week later. Both extratropical storms generated strong winds over the northern half of Europe (above 45∘ N) from 9 until 18 February 2020. Strong winds, yet milder than during storms Ciara and Dennis, were also generated by storms Karine and Myriam over the Iberian Peninsula, the southern part of France, and the northern part of Italy in the first week of March. Moreover, February and March 2020 displayed stronger positive temperature anomalies over Europe in comparison with February and March 2019 (https://surfobs.climate.copernicus.eu/stateoftheclimate, last access: January 2021). Such weather anomalies, however, did not persist during the second quarter of 2020. Accounting for the effect of such meteorological variations is very important to assess accurately the effect of COVID-19-related mobility restrictions on air pollution. Different approaches can be used to assess the pollution changes based on different types of data, such as satellite observations, surface site observations, and air quality models.

Several studies used the recently launched (October 2017) Tropospheric Monitoring Instrument (TROPOMI; Veefkind et al., 2012) onboard the Copernicus Sentinel-5 Precursor (S5P) satellite to highlight the NO2 reductions caused by the COVID-19 lockdowns. The substantial interannual variability in meteorological conditions together with the young age of the instrument prevented the estimation of a representative climatological baseline to which NO2 levels observed during the lockdown period could be compared. As a result, satellite-based studies using TROPOMI comparing before- and after-lockdown periods (e.g. Wang et al., 2020) or comparing the lockdown period with its 2019 equivalent (e.g. Bauwens et al., 2020; Nakada and Urban, 2020; Zambrano-Monserrate et al., 2020) have given little to no weight to the synoptic meteorological conditions and how they could potentially flaw the emission change estimates.

In contrast, Schiermeier (2020) mentioned the “weather factor” early on in the COVID-19 crisis, which can strongly affect the pollution levels. And studies such as Le et al. (2020) showed 2019 and 2020 TROPOMI NO2 comparisons but acknowledged the impact of weather anomalies on pollution levels. It is only very recently that a weather-normalization technique has been applied to estimate NO2 changes due to the COVID-19 restrictions across cities in the US based on TROPOMI (Goldberg et al., 2020). Yet, such analyses place insufficient importance and provide insufficient clarity about the fact that satellite data used in such analyses are conditioned by the cloud coverage, revisit frequency, and quality flag of the satellite observations. Ignoring or not acknowledging such information can also lead to flawed satellite-based estimates and provide misleading information (https://atmosphere.copernicus.eu/flawed- estimates-effects-lockdown-measures-air-quality-derived-satellite-observations, last access: January 2021).

Several studies have investigated lockdown impacts using surface measurement sites. For example, Wang and Su (2020) showed that lower emissions from motor vehicles and secondary industries were most likely responsible for the observed decreases in NO2 concentrations in China during January–March 2020. Collivignarelli et al. (2020) showed using surface station measurements that major NO2 reductions occurred in Milan, a city that showed a rapid increase in cases early in the European COVID-19 crisis (February 2020) and was one of the first cities to be put into lockdown in Europe. Past studies such as Carslaw and Taylor (2009) showed the usefulness and the importance of weather-normalization techniques for air pollution applications based on surface observations, such as the local air traffic activity impact on NO2 predictions. This was followed more recently by Grange et al. (2018) and Grange and Carslaw (2019), who used machine learning techniques to perform weather normalization for analysing trends and detecting the impact of policy measures on air quality. Built on this previous work, several studies made use of machine learning to estimate the impact of the COVID-19-related mobility restrictions on air pollution levels, taking into account the confounding effect of the meteorological variability. Using machine learning (ML) models fed with ERA5 reanalysis meteorological data, Petetin et al. (2020) highlighted a strong reduction in surface NO2 concentrations across most Spanish urban areas during the first weeks of lockdown. Similarly, Keller et al. (2021) assessed the NO2 pollution changes using worldwide surface measurements showing country-dependent variations in reductions.

Finally, air quality modelling systems offer a valuable tool for representing the evolution of pollutants in the atmosphere according to changes in emissions, physical processes, and weather variability. The Copernicus Atmosphere Monitoring Service (CAMS) produces daily European air quality forecasts and analyses using an ensemble of 11 models, ensuring unique reliability and quality (Marécal et al., 2015). Using emission scaling factors to account for lockdown measures, such an ensemble of models can be used to estimate lockdown reductions in NO2 pollution (amongst other pollutants) and account for the weather variability at the same time (Colette et al., 2020; Guevara et al., 2021).

This paper aims to provide a comprehensive and comparative assessment of the impact of the first European COVID-19 lockdown on NO2 pollution levels over major European urban areas using satellite observations, surface in situ observations, and air quality models. We firstly illustrate how misleading it can be to ignore the influence of the weather variability when assessing the lockdown-induced changes in NO2 with TROPOMI. Then, in order to quantify these changes, we use ML-based weather-normalization methods for estimating the “business-as-usual” (BAU) NO2 pollution levels that would have been observed without any lockdown measures, based on both TROPOMI NO2 tropospheric columns (Sect. 2) and surface in situ observations (Sect. 3). NO2 changes are then investigated with the CAMS regional ensemble (Sect. 4). We compare and discuss the three different approaches in Sect. 5 followed by conclusions in Sect. 6.

In Table 4 and Fig. 9 we summarize the results of this study. Table 4 shows the average reduction in all the median estimates together with the inter-urban-area variability over Europe. Figure 9 shows the distribution of the NO2 changes estimated for the lockdown period per urban area. This figure provides estimates equivalent to box plots where the median and the interquartile range are displayed. For clarity, we chose to display only urban areas that have more than 1 million inhabitants. The values of each estimate for all urban areas considered in this study are given in Table B1 in Appendix B.

The three types of weather-normalized estimates agree on identifying stronger reductions where more severe lockdown measures were implemented. As shown in Sect. 2, satellite-based estimates show a relationship between NO2 tropospheric-column reductions and the extent and generalization of restrictive measures in each country. A similar relationship is observed for surface sites and model estimates (Sects. 3 and 4). The largest NO2 reduction estimates of around 50 % to 60 % for both surface and tropospheric columns are found in Spanish, Italian, and French urban areas. In countries that implemented softer lockdown measures, urban areas show smaller reductions, e.g. Germany, Netherlands, Poland, and Sweden. Although significant discrepancies exist between the satellite-, surface-, and model-based estimates in urban areas such as Naples (Italy), Sofia (Bulgaria), and Katowice (Poland), the three methods provide an overall consistent picture. It is remarkable to note that this result contributes to establishing the usefulness of satellite-based estimates for urban air quality and not only for atmospheric pollution in general. Having a range of three different types of estimates helps to provide estimates of pollution changes across Europe with a certain level of certainty. When all the estimates agree, it is more likely that the values of reduction due to the lockdown implementations are reliable. Conversely, if the different types of estimates show discrepancies, less confidence should be given to the reduction estimates. In Fig. 8, Madrid, Turin, and Milan, to mention a few urban areas, show consistency between the different types of estimates, expressing more certainty in the results. In other locations such as Sofia, Athens, and Budapest, strong discrepancies indicate that the estimates could be uncertain. Average scores in Table 4 show that surface station observations provide stronger reduction estimates and that satellite-based estimates provide weaker reduction estimates. Model estimates are mostly in between and show much less spread within a given urban area (bars in Fig. 9) and less variation between urban areas (standard deviation in Table 4). The origin of such differences can vary and is detailed below.

Machine learning estimates that are observation-based (satellite and surface stations) show more spread compared to the model estimates. In Fig. 9 the interquartile ranges for the observation-based ML estimates are much larger than for the model estimates. Such large ranges show that there is a strong spread in the ML-based estimates that is not seen in the model-based estimates. Model estimates are based on country-dependent emission reduction or scaling factors that are constant over time. The variability is induced by the changes in atmospheric conditions but not by changes in the emissions. The estimates from the ML approach can represent the transition into the lockdown where emissions gradually decreased. This contributes to the increased spread seen in the ML estimates. Scores from ML estimates (see Tables 1 and 3) also show significant RMSE that can add noise to the time series and add to the resulting spread of the distributions. A stronger spread in TROPOMI estimates is likely due to the small training set used. Disentangling the noise and the actual variability would need to be done carefully in future work.

All the different estimates presented in this study are consistent in their spatial scale, using 0.1∘ × 0.1∘ TROPOMI-averaged pixels that match the CAMS forecasts and surface stations within a 0.1∘ range from the city centre. Some of the smaller urban areas considered in this study likely have a footprint that is smaller than 0.1∘, meaning that high pollution levels from the urban area are mixed with low pollution background levels. This could cause the pollution changes in the gridded estimates to be weaker than expected in certain urban areas (e.g. Katowice, Budapest, Glasgow, etc.). Also, even if the urban and suburban background stations are selected, the in situ surface observations sample the pollution levels within a 0.1∘ × 0.1∘ pixel given their location. This sampling might not be exactly representative of the average pollution footprint within the same pixel. This average is the information given by the models or the satellites. These representativeness issues contribute to creating discrepancies between the type of estimates and hence generate uncertainty. The differences seen in Fig. 9 between surface station estimates and gridded estimates (models and satellites) point out such possible representativeness issues. Representativeness is a difficult and important topic and deserves further research as it would require careful examination of the stations' locations in specific urban areas and also using higher-resolution modelling than 10 km.

Satellite overpass times (13:30 LST) and the presence of clouds in the measurement pixel can potentially influence the reduction estimates from the TROPOMI data. We considered 1.5 months to compute the satellite reduction estimates. Overall, the sample size (valid S5P overpasses) in Fig. 9 ranges between 14 (Sevilla) and 37 (The Hague). Also in Fig. 9, surface sites and model estimates are displayed for hourly and S5P-sampled estimates. Smaller or larger samples cannot really explain discrepancies between all the different estimates. Results, however, can be affected when the sample size becomes statistically very small and if shorter time periods (e.g. 1 or 2 weeks) are considered for satellite reduction estimates. Very small samples over the 6-week period were not considered in this study to avoid this effect. The sampling effect also shows greater changes in the surface station estimates than in the model estimates. As mentioned above and seen in Fig. 9, the surface station estimates provide more variability that accounts for hourly variations. The model estimates have fixed emission scaling factors for the entire lockdown period. The surface station estimates show more sensitivity to the time sampling than the model estimates. On average (see Table 4), the S5P overpass sampling changes the estimates by around -6 % for surface station estimates and only by -1.5 % for model estimates. This suggests that the lockdown-induced reduction estimates depend upon the time of the day, i.e. those times when the road transport activity peaks.

Finally, the reduction estimates for tropospheric NO2 columns displayed in Fig. 9 are generally not as strong as the NO2 surface estimates (observations and model). Some exceptions can be seen in certain Spanish (e.g. Barcelona, Madrid) and Italian (e.g. Milan, Turin) urban areas, where column estimates are close to the surface estimates, but overall column reductions are weaker. With all urban areas considered, the satellite estimates show around 23 % reduction on average, which is 10 % to 20 % less than the model and surface station estimates (see Table 4). This can be expected as NO2 surface site measurements do not directly translate to the TROPOMI NO2 tropospheric column, which is the integrated NO2 content from the surface to about 200 hPa altitude. Due to the short lifetime of NO2 (around 12 h), only small lockdown-induced changes to the free-tropospheric NO2 contents are expected. Changes are mainly expected near the surface and within the PBL. Therefore, the different nature of the vertical sampling is likely to contribute to the differences between the relative reduction estimates from tropospheric columns versus surface concentrations. Further work will be needed to quantitatively link the tropospheric-column and surface-level variations, including sampling the model estimates using an observation operator commonly used in data assimilation and inverse-modelling systems. This important work will be carried out in a further study.

In this paper, we first show the importance of accounting for weather variability in satellite-based estimates of NO2 changes due to the COVID-19 lockdown. While focusing on Europe and using the TROPOMI instrument, we show that the satellite estimates based on direct comparisons between different time periods without accounting for weather variability can be flawed and should not be used for this kind of assessment. To account for weather variability in satellite estimates, we use a recently developed methodology based on the gradient boosting machine learning technique. This methodology has proven to be efficient with surface sites to estimate lockdown-induced changes over Spain (Petetin et al., 2020). We extended those surface estimates over Europe to compare with the satellite estimates. Finally, we included estimates of NO2 changes using the 11-model CAMS regional ensemble, using emission reduction factors representative of the lockdown period. By providing and comparing the three different methodologies, we provided a comprehensive and complementary assessment of NO2 pollution level changes during the COVID-19 European lockdown. These assessments of pollution changes, when activity levels of key emitting sectors are significantly reduced (i.e. road transport and industry) in lockdown conditions, also provide crucial information to accurately quantify the benefits of the potential implementation of air quality policies for these emission sectors.

Main results show a consistent tendency of stronger reduction in NO2 where more stringent lockdown measures were implemented. On average, the three types of estimates show a reduction of 23 %, 43 %, and 30 % for satellite, surface stations, and model estimates, respectively. Differences are explained by the different nature of the methods used, i.e. observation-based versus model-based, horizontal and vertical sampling, variability representation and time sampling. By providing an array of different methods, we provide an indication of how reliable the pollution reduction estimates are for the various urban areas considered in this study. Accurately quantifying the pollution changes is also important for the impact of these pollution reductions on the COVID-19 pandemic itself. Several studies have investigated the correlation between the high level of COVID-19 mortality and atmospheric pollution (e.g. Contincini et al., 2020; Ogen, 2020; Achebak et al., 2020). Feedbacks are then to be expected between the effects of short-term air pollution exposure on COVID-19 mortality and lockdown measures. Beyond the quantification of the impact of COVID-19-related restrictions on pollutant concentrations, the observation-based weather-normalization methodology used in this study is of general interest for assessing the impact of any type of emission changes (e.g. regulation and policy) on air quality (Grange et al., 2018; Grange and Carslaw, 2019) in the future.