1. Introduction
Nitrogen dioxide (NO
2) is a common air pollutant that can harm the environment and human health. Anthropogenic NO
2 emissions were primarily reported from the transport, industry, power, heating, and residential sectors [
1,
2]. Exposure to NO
2 has been linked to many respiratory diseases [
3]. Chronic exposure to NO
2 pollution led to a higher mortality rate during the COVID-19 pandemic [
4] and these pollutants affected virus transmission [
5]. Triggered by the pandemic, air pollution and traffic were identified as two of the top six environmental challenges facing cities [
6]. Unprecedented strong policy measures, such as nonpharmaceutical interventions, were taken worldwide, especially in urban areas, restricting human activities and industrial production. The strict restrictions directly affected the emissions of various air-polluting gases, and in turn, such changes in global and regional emissions may reflect the effects of different policies. In this context, a consensus has been proposed that this “pause button” created a unique window to assess anthropogenic effects on atmospheric composition and its relationship with human activities compared to “business as usual”. More importantly, it calls for extensive actions that can make full use of this “anomaly” opportunity to boost our cities’ shift to sustainability [
6], as sustainable development increasingly depends on the successful management of urban growth. NO
2 has been considered a good indicator of changes in local pollutant sources, as it has a shorter lifetime (within 1 day) compared with other air pollutants [
7]. Therefore, in this study, the sensitivity of NO
2 to policy responses during the COVID-19 pandemic can be captured or tracked by restriction measure changes on a daily scale. It provides an observational metric for examining the effects of diverse policy regulations on air pollution, reducing the complexity due to the hysteresis of some policy effects. It is important for understanding which policies (linked to different human activities) are effective in reducing emissions.
At the early stage of the epidemic, the observation-based concentrations of NO
2 pollutants were consistently reported to be below the levels of previous years, as a result of the sudden lockdowns and suspended human activities. Data from both satellite observations and ground-based monitoring stations found that global NO
2 concentrations were lower in 2020 than in 2019, especially in urban areas [
7,
8]. In Wuhan, China, the first reported outbreak city, the NO
2 in 2020 during lockdown was 42% lower than the average level of the previous three years [
9]. The reduction in transport sector emissions is reported as the main cause of global NO
2 anomalies [
8]. London experienced a 50% decrease in NO
2 during the lockdown, with the highest NO
2 emission reductions observed during morning rush hours due to the radical changes in routine life and commuting time [
10]. Observations from other mega-cities, such as New York [
11], Tokyo [
12], Moscow [
7], and Toronto [
13], also showed a decrease in NO
2 concentrations during lockdowns. Nevertheless, some subsequent studies noted the possibility of exaggerating the positive effects of COVID-19 policies and pointed out that the improvement in air quality was not due to COVID-19 interventions but seasonal factors, as the lockdown period coincided with the onset of the rainy season in tropical regions, such as in Nigeria [
14] and Indonesia [
15]. This observation arises from the fact that most studies reported improvements in air quality based on comparisons between lockdowns and pre-lockdown periods or by taking the previous 3- to 5-year average level as the baseline. Some review papers investigated the nexus between the pandemic and the environment, showing that meteorological conditions affected NO
2 concentrations [
15,
16].
With continuous discussions on this popular topic and the accumulation of available multisource datasets, recent studies have begun to use the sensitivity changes of NO
2 to infer, correlate, and retrieve various phenomena, consequences, or environmental indicators related to it. In this context, ground-based instruments, alone or combined with satellite observations (e.g., from the Ozone-Monitoring Instrument (OMI) on-board Aqua satellite and the Tropospheric Monitoring Instrument (TROPOMI) on-board Sentinel-5P), provide fundamental information about anthropogenic impacts. The stringency of policy indicators has been used to investigate certain impacts on the environment. Long-term temporal comparisons of tropospheric NO
2 vertical columns during the lockdown with the counterfactual baseline concentrations have been investigated with different extensions and specific perspectives. For example, Misra et al. [
17] retrieved the NO
2 concentration changes due to lockdown from OMI and TROPOMI and linked the changes to power plants over urban areas in northern India while considering the seasonal components of NO
2 and long-term trends over the same period in 2015–2019. Xing et al. [
5] used machine learning with multisource data to prove that reduced economic activity, inferred from NO
2 reductions (restrictions-induced), drove the slowdown in the number of COVID-19 infections in most regions [
5]. Highlighting that pollution has no borders, Li et al. [
18] clustered global continent grid cells, which were based on historical NO
2 pollution levels from OMI satellites for which seasonal trends had been removed, by subtracting 10-year daily mean values to investigate the impact of policy stringencies on different clusters (regions). There are significant differences between regions due to differences in measures taken, the duration of lockdowns, and the intensity of measures implemented by the government, which allows much potential for in-depth explorations.
Most of the existing studies documented the short-term positive effects of containment policy measures on declining air pollution; however, the long-term effects have not been fully considered, especially for other confounding factors, such as seasonality, climate conditions, air quality improvement trends, and the local context. It may exaggerate the effects of COVID-19 restrictions. Furthermore, few evidence-based methodologies could identify the effective policy, which has a direct impact on the NO2 anomaly and is linked to specified human activities. In other words, very limited efforts have been made to differentiate and quantify the effects of different policies. This challenge may be a barrier for policy-makers or create indecision when effective actions are considered.
To fill the research gaps discussed above, in this study, we proposed that observations combine with predictions. We employed historical ground-station data combined with Sentinel-5P data to illustrate the sensitivities of NO2 concentrations to different policy measures of COVID-19. Particularly, we estimated the portion of NO2 anomalies caused by COVID-19 restriction policies by removing the confounding factors. The spatial and temporal variations of NO2 were visualized for interpretations based on remote sensing. Ultimately, this study applied spatial stratified heterogeneity statistics in time-series analysis: By linking the daily-scale variations of containment measures to NO2 anomalies, we quantified and ranked the contributions of different policies. Such analysis is expected to generate a more reliable and accurate evaluation of the effects of different restriction policies on human activities that reflect NO2 anomalies. It is essential to evaluate how policies can further guide human activities as cities transition to environmental sustainability during their post-epidemic recoveries. Furthermore, this evaluation may help shed light on some uncertainties regarding future global and regional climate responses to polluting gas emissions.
4. Results and Discussion
4.1. Seasonality and Trend
From the six-time series shown in
Figure 3, NO
2 concentrations show a certain seasonal distribution pattern (annual periodicity), which is generally at a low level from March to September each year and then has high peaks in winter, especially in November and December. The meteorological factors show obvious seasonal characteristics, especially in the annual periodic changes.
From the X11 time series decomposition in
Figure 4, the original NO
2 curves on the top consist of multiplying the three components below, that is, trend, seasonal, and irregular components. It can be seen from the figure that after removing the effects of seasonal and irregular components, NO
2 showed a downward trend. Moreover, the decomposition shows obvious seasonal periodic features where there is a nearly robust seasonal component, with no big change every year. Seasonality and gradual downward trends were detected, which should not be ignored in terms of analyzing the effects of policy measures on NO
2 emissions during COVID-19. The irregular component shows unexpected changes, which may reflect different policies at different times during the COVID-19 pandemic. The residual item is found by subtracting the seasonal item and the trend item from the original time series data. These residual items distributed at each time node reflect disturbances due to certain events or policy measures. The implementation of the strict lockdown policy can be traced to the first half of 2020, from March to July. In addition, a number of significant air quality control measures that were implemented in September 2017 could also be traced on the timeline, for example, the closing of several coal-fired power plants and phasing out of more than 2 million high-emission vehicles. The long-term downward trend may be related to these structural emission reduction policies, which have been implemented in recent years in Beijing.
4.2. Prediction versus Observation
From the time-series prediction model, we obtained the expected NO
2 value (prediction results) for the “business-as-usual” condition, as shown in
Figure 5. In this case, the time-series prediction covered the seasonality and a downward trend as the above section indicates; comparing the observed NO
2 in 2020 and this prediction result gives the NO
2 anomaly due to COVID-19 policy interventions. Compared with the prediction line, the observed NO
2 in 2020 decreased by 6.03% on average for the whole year, exceeding the expected level. In particular, from March to June 2020, the observed NO
2 level was significantly lower than the predicted line, with the largest decrease of 17.04% recorded in March. This period coincided with the implementation of strict lockdown policies; as the pandemic situation improved in the second half of the year, policies were dynamically adjusted, and economic activity gradually recovered. A rebound peak appeared in October and November, showing an even higher NO
2 emission than the expected (non-COVID-19) level, though it was still below the 5-year average baseline.
Compared with the 5-year average level, NO
2 observed in 2020 was significantly lower, with an average decrease of 14.10% and the largest decrease in March (23.92%). However, it should be noted that since 2018, it has been lower than the baseline. The rebound peak would be hard to detect if only comparing with previous average data and not removing the other factors (seasonality and the downward trend). We also compared the results with the 2019 data. This showed that the expected NO
2 levels in 2020 would be expected to decrease by 2.68%, while the actual NO
2 levels in 2020 decreased by 8.60%.
Table 1 presents a more detailed comparison between observed and expected values.
4.3. Interpretations from Remote Sensing
The differences in the spatial distribution of NO
2 were detected using remote sensing data. Using spatial raster calculations between 2019 and 2020, the spatial NO
2 changes show where decreases and increases have taken place (
Figure 6). The findings indicated that the implementation of the policy had the greatest impact on economically developed and densely populated urban areas, as the urban areas maintained a consecutive and significant reduction of NO
2 from February to July 2020, particularly in the northern urban areas of Beijing. On the other hand, suburban areas with lower population density and less economic activity experienced less change.
From August to December 2020, NO2 in some places increased but declined in other places, which is evident from the remote sensing data but is not reflected in the average aggregated data from the ground monitor. For example, the line graphs of NO2 levels measured with ground-station data show that there was a NO2 rebound anomaly in November 2020, which was higher than the expected results, while the remote sensing data showed where the spatial differences were. This may indicate that some conditional economic recovery activity (coexistence with the pandemic mode) led to the spatial heterogeneity of production activities.
Interpreted from the satellite observation, from September to December, it also showed that NO2 concentrations in the urban center area primarily decreased, while the southern urban area showed an increasing trend. This increase may be related to flight activity following the reopening of Daxing International Airport and the recovery of industrial activity. In addition, winter heating related to burning fossil fuels may also account for the high levels in November and December 2020.
4.4. Quantifying Different Policy Effects on NO2 Anomalies
The model results show that the dominant factors and level of influence of factors can be ranked according to their q values, namely, C2 > C5 > C1 > C3 > C6 > C7 > C8 > C4, as
Table 2 shows. Note that the “dominant” factor here means that when it has a larger q value, it will have a relatively larger impact than other factors.
Specifically, among the measures, C2 “Workplace closing” had the greatest effect, explaining 54.8% of the NO
2 anomaly. This is likely because the C2 policy directly affected people’s commuting activities. The second most influential policy was C5 “Close public transport”, which explained 52.3% of the NO
2 anomaly. This is consistent with many existing studies, which find that the transportation sector affected reductions in NO
2 during the COVID-19 pandemic [
10,
11]. The third dominant factor is C1 “School closing”, where school closure measures explained 46.4% of the NO
2 anomaly. This was similar to the influence of C3 “Cancel public events” (44.8% explained). The fifth explanatory variable was C6 “Stay at home requirements”, which explained 42.1% of the NO
2 anomaly. The top five dominant factors are all related to people’s travel activities.
The sixth explanatory variable was C7 “Restrictions on internal movement”, that is, restricting people’s movements within the city. After the implementation of the green code system and the precise division of high-, medium-, and low-risk areas in Beijing, this measure primarily refers to restricting the flow of people from high-risk areas to other areas. At the same time, people in non-high-risk areas were also advised not to travel to high-risk areas unless necessary. The last two impact factors are C8 “International travel controls”, which restricted international flights, and C4 “Restrictions on gatherings”, which restricted population gatherings of different sizes. These last three policy factors were once considered to be the most effective means of controlling the spread of the virus, but under the “dynamic zero-COVID” policy in China, these policy measures have had the least impact on the reduction of NO2 emissions. This indicates that controlling the routine activities within the city had a higher impact on NO2 changes than controlling inter-city activities.
Significant differences can be seen between the levels of implementation of the various policies, that is, when a policy is not taken (e.g., no measures = 0), not enforced (e.g., recommended = 1), and enforced (e.g., different intensity for 2~4). However, the conditional threshold and limit size of the specific levels of implementation need further exploration. For example, our preliminary results show that the impact of C2, level 2 (require closing (or working from home) for some sectors or categories of workers) is the most important, rather than all closing or other levels. Another example can be seen in the differences in the levels of C4 “Restrictions on gatherings”, where the effect was most pronounced when the population size was limited to less than 100 people. Further research is needed on how to design the effective policy threshold for guiding a new balance between low air pollution and people’s lifestyles, as well as other policymaking objectives.
4.5. Discussion
Taking the 5-year average level as a baseline, our results show a reduction of 14.10% due to COVID-19 policy measures. Compared with existing research, it is significantly less than the 30%~40% reduction suggested in [
9,
11] at the city level, but the value is closer to the national average level by 6% [
5] although the spatial-temporal records show differences between national and regional scales. Beyond the effects of policies on NO
2 anomalies widely discussed in many studies [
7,
8,
18], this study also identified the top three measures that have dominant effects on urban NO
2 levels: C2 “Workplace closure”, C5 “Restricted public transport usage”, and C1 “School closure”, accounting for 54.8%, 52.3%, and 46.4%, respectively. These three dominant policies were linked to restrictions on commuting activities, transportation, and education activities, respectively, which were likely to mitigate traffic congestion pressures in this city. By analyzing the contributions of different behavioral constraints to the NO
2 anomalies, we can speculate and reflect on the problems of urban structure and spatial development imbalances in Beijing, that is, the mismatch of employment and housing (determined by commuting patterns), educational inequality, and the long-term unsolved congestions, as discussed in previous studies [
32,
33]. This convincing evidence influenced the top three corresponding restriction policies; in other words, promoting the transformation of urban spatial structure will effectively alleviate air pollution. Now we have the chance to boost the urban shift to environmental sustainability goals, as cities have become more flexible and open to change than in the past in regard to urban planning and environmental management for post-pandemic recovery.
5. Validation
We evaluated the effectiveness of the model from each indicator shown in
Figure 7. In the process of model selection and calculation, we processed all data as time series data. First, we selected climate data (temperature, precipitation, wind, and surface pressure) as the independent variables, as has been performed for regression modelling for air pollution estimates in many studies. We simulated daily and monthly data, respectively, and found that the R
2 of monthly data (R
2 = 0.51) performed better than the daily data fitting results (R
2 = 0.34); however, the model fitting results were not good enough for our prediction goals. Combined with the X11 time series decomposition model, it was further found that the climate data exhibits certain seasonal components, which are primarily of a cyclical nature, as can be seen in
Figure 3; this was especially evident with the monthly data. We, therefore, defined a trend term and seasonal dummy term for the multiple linear regression model in order to conduct the time-series prediction. The results showed that the monthly data fitting result has a better performance than previous experiments (R
2 = 0.80).
Since we used historical time-series data, the best method of validation is to use historical data from 2015–2018 as training data for prediction and to use 2019 observation data as verification (testing) data to compare that with the expected value from the model prediction results for 2019. The comparison results show that the overall accuracy reaches 97.71%.
Figure 7 shows the validation results, and it fits well.
This study confirmed the consistency between satellite and ground data by using a linear regression model. Although different instruments and sensors were used to measure NO
2 levels, the changes observed were relatively consistent and significantly correlated (R
2 = 0.65), as shown in
Figure 8. However, it is important to note that ground station data provide measurements at discrete points, resulting in non-continuous monitoring of the surface, and thus cannot reflect the spatial variations of NO
2 across the entire city. In contrast, satellite monitoring provides continuous observations of NO
2 concentrations in a “column” over urban space. This can lead to deviations in results when considering average, aggregated monthly correlation. The different spatial conditions, according to the mechanisms of air mass factor transmission, can also contribute to these deviations. Specifically, some models developed for converting satellite column data to ground-level concentrations, such as the GEOS-Chem chemical transport model (
https://geos-chem.seas.harvard.edu/, latest version accessed on 1 February 2023), have taken into account complex atmospheric transfer properties and a range of atmospheric conditions. It has been found that the shape factors of NO
2 have significant variability [
34], with peaks near the surface in urban regions due to local pollution sources and in the upper troposphere in remote regions due to lightning.
6. Conclusions
In this study, we employed time-series decomposition and regression-prediction models to evaluate the effects of COVID-19’s different containment policies on NO2 anomalies in Beijing, China. Unlike most previous studies, we detected the effects of seasonality and a long-term downward trend on NO2 changes and excluded them for analysis. This model design aimed to identify the specific effects due to lockdown measures. The results showed that observed changes exceeded expectations, with NO2 decreasing by −6.08% on average every year, and as much as −17.04% (95% CI, −7.71%~−24.67%, p < 0.001) in March 2020 when the strictest lockdown measures were in place. When compared with previous studies, our results are lower than most existing studies; It suggested that air quality policies and the local context should not be ignored when assessing the impacts of COVID-19 policies on changes in pollutant levels, and policymakers should beware of exaggerating the effects of the COVID-19 restrictions on the NO2 anomaly. In addition, the main novelty and contributions also lie in that this study succeeds in elucidating the differences in human activity containment and quantifying their contributions to NO2 anomalies.
Information about factors that reduce NO2 emissions during COVID-19 can be helpful in designing policy responses. As seen in this study, the restrictions on human activities, especially those related to transport, played a dominant role in the observed reduction of NO2. By removing seasonal and trend impacts, the study revealed that NO2 levels rebounded in the second half of 2020, especially in October and November, with peaks appearing in November earlier than in previous years (higher than expected according to the natural cycle). This suggests that the government’s recovery policies, for example, to stimulate production activities, should also take into consideration the “retaliatory” emissions. This study is expected to help decision-makers to identify the effects and sensitivity of NO2 to different policy responses, thereby allowing them to enact more precise policies that balance air pollution control and measures to support post-COVID-19 recovery. We have identified the dominant policies and assessed their corresponding magnitudes of influence at the urban-scale level. This evidence may help other regions to examine the impact of their own policies.
The complementary of historical ground-based data with satellite datasets enables a more comprehensive evaluation of NO
2 anomalies than monitoring with a single source, by leveraging the strengths of both monitoring datasets. More specifically, in this study, the available long-term ground-station observation data are used to construct a time-series model to eliminate compound and mixed effects, such as seasonal and structural emission control effects, to extract partial anomalies attributed to strict policy measures. The current Sentinel 5P has no such long historical records prior pandemic applicable for temporal analysis, and here it was mainly used to map spatial variations of NO
2 in 2019–2020 and helped with better interpretations of the time-series results. In addition, the ground station data represent measurements at discrete points, providing non-continuous surface monitoring, and cannot reflect the spatial distribution or variations of NO
2. Instead, satellite monitoring enables continuous observation of NO
2 concentrations in “column” over space, though it is not analogous to ground-station monitoring. In particular, it can detect the spatial heterogeneity of the NO
2 level and give spatially resolved information for the urban structure that single ground monitors or aggregated average data cannot reflect, as we discussed in
Section 4.3. With advancements in satellite technologies, the TROPOMI satellite instrument can infer ground-level NO
2 concentrations at a finer spatial resolution using specific models and obtain continuous observations over a longer duration.
This study focused on evaluating the effects of pandemic restrictions on NO2 anomalies, but it also has some limitations. One is that the forecast of atmospheric trace gases is a complex task, which requires expert knowledge in atmospheric sciences. This calls for greater cooperation between researchers in different fields in the future. Furthermore, the study used monthly average data to detect NO2 anomalies during COVID-19 and the model results were well-fitted, while the analysis and simulations for daily changes in NO2 require more complex models, which can handle a 5-year time scale or even longer, as well as the effects of emerging events, climate change, and other uncertainties. Some less obvious potential sources of pollution should also be given more attention. For example, telecommuting decentralized office activities and reduced travel activities, but stay-at-home orders are likely to have increased households’ consumption of energy. Another consideration is whether less usage of public transport will increase the demand for private cars and generate more pressure on air quality. Next, comparison studies between countries will be conducted to explore different policy-driven forces, and the applications of this methodology may further examine the variability over regions and time. It would also be useful to fine-tune these methods to assess the policy sensitivity of other types of air pollutants.