Changes in air pollutants during the COVID-19 lockdown in Beijing:Insights from a machine-learning technique and implications for future control policy

Jiaao Hu , Yupng Pan , Yuxin H , Xiyuan Chi , Qianqian Zhang , Tao Song ,Wisho u Shn

a Collaborative Innovation Centre of Atmospheric Environment and Equipment Technology, Jiangsu Key Laboratory of Atmospheric Environment Monitoring and Pollution Control, School of Environmental Science and Engineering, Nanjing University of Information Science & Technology, Nanjing, China

b State Key Laboratory of Atmospheric Boundary Layer Physics and Atmospheric Chemistry (LAPC), Institute of Atmospheric Physics, Chinese Academy of Sciences,Beijing, China

c University of Chinese Academy of Sciences, Beijing, China

d National Meteorological Center, China Meteorological Administration, Beijing, China

e National Satellite Meteorological Center, China Meteorological Administration, Beijing, China

Keywords:Random forest model Air pollutants Meteorological normalization COVID-19 Emission control strategy

A B S T R A C T The COVID-19 lockdowns led to abrupt reductions in human-related emissions worldwide and had an unintended impact on air quality improvement. However, quantifying this impact is difficult as meteorological conditions may mask the real effect of changes in emissions on the observed concentrations of pollutants. Based on the air quality and meteorological data at 35 sites in Beijing from 2015 to 2020, a machine learning technique was applied to decouple the impacts of meteorology and emissions on the concentrations of air pollutants. The results showed that the real ( “deweathered ”) concentrations of air pollutants (expect for O 3 ) dropped significantly due to lockdown measures. Compared with the scenario without lockdowns (predicted concentrations), the observed values of PM 2.5 , PM 10 , SO 2 , NO 2 , and CO during lockdowns decreased by 39.4%, 50.1%, 51.8%, 43.1%, and 35.1%, respectively. In addition, a significant decline for NO 2 and CO was found at the background sites (51%and 37.8%) rather than the traffic sites (37.1% and 35.5%), which is different from the common belief. While the primary emissions reduced during the lockdown period, episodic haze events still occurred due to unfavorable meteorological conditions. Thus, developing an optimized strategy to tackle air pollution in Beijing is essential in the future.

1. Introduction

The coronavirus disease (COVID-19) broke out abruptly in December 2019 and spread rapidly across the world. In response to the COVID-19 crisis, most governments around the world introduced restrictions on behavior or lockdown measures. As a consequence, there has been a sharp slowdown in global economic growth and human-activity-related pollutant emissions. During the lockdown period, for instance, emissions of COdeclined by 6.9%, 12.1%, and 9.5% in China, Europe, and the U.S., respectively ( Liu et al., 2020 ). In addition, the concentrations of air pollutants also decreased greatly in most cities in the world, especially NO, with a reduction by 20%, 43.5%, 50%, 51%, 52.7%, and 54.3% in Baghdad ( Hashim et al., 2021 ), Rome ( Kumari and Toshniwal, 2020 ),Barcelona ( Baldasano, 2020 ), New York ( Zangari et al., 2020 ), Delhi( Mahato et al., 2020 ), and S?o Paulo ( Nakada and Urban, 2020 ), respectively. However, these ratios were mostly obtained via simple statistical analysis that compared concentrations of air pollutants before and after the lockdowns, or during the lockdowns, with the same periods in previous years ( Shi et al., 2021 ). This common statistical approach comes with a major caveat that meteorological conditions may have masked the real effect of changes in emissions on the concentrations of target air pollutants ( Zhang et al., 2020 ; Shi et al., 2021 ). Thus, there are difficulties in using such a method to explain the observed haze events during the lockdowns in some cities ( Huang et al., 2020 ).

While atmospheric chemistry and transport models can decouple the effect of emission changes from meteorology ( Vu et al., 2019 ), this method still suffers from the lack of a timely emissions inventory to reflect the real-world changes ( Wang et al., 2020 ). As an alternative approach, machine learning offers a reliable way to quantify changes in air quality due to emissions and meteorological factors ( Shi et al., 2021 ).Indeed, a random forest (RF) method was successfully used to assess the short-term changes in selected cities in China due to the COVID-19 lockdowns ( Wang et al., 2020 ). However, this study did not decouple the impacts of meteorology on the observed concentrations. Since the measured concentrations are determined by both emissions and meteorology, it is essential to decouple the effects of meteorology to understand the link between emissions and interventions ( Zhang et al., 2020 ).

Here, we employed a novel machine learning technique based on an RF algorithm to evaluate the impacts of the COVID-19 lockdowns in early 2020 on air quality in the megacity of Beijing. The selected city covers a range of air pollution levels, from highly to less polluted regimes. The 35 monitoring sites were divided into urban, suburban,background, and transport (road traffic) sites to better understand the impacts of various emissions on air quality changes. This study aims to(1) document the concentrations of air pollutants in early 2020 and the same periods in the past five years, (2) decouple the effects of meteorology from short-term emission changes on air quality, and (3) predict the concentrations of air pollutants without lockdown measures. Our findings can be used to elucidate the changes in air pollutants at different site types during the pandemic and provide guidance for policymakers in designing mitigation strategies towards improving air quality in future.

2. Data and methods

2.1. Data sources

The major air pollutants of particulate matter (PMand PM),sulfur dioxide (SO), nitrogen dioxide (NO), ozone (O), and carbon monoxide (CO) at 35 stations in Beijing were used in this study ( Fig. 1 , Table S1). The data were downloaded from the Beijing Municipal Ecological and Environmental Monitoring Center(h ttp://www.bjmemc.com.cn/). In this study, hourly mean concentrations were used for the air pollutants, except O 3 , for which the maximum daily 8-h average concentrations were calculated for each period.

Fig. 1. Monitoring sites for air quality and meteorology in Beijing.

The hourly meteorological parameters including temperature, relative humidity, wind speed and wind direction, and atmospheric pressure, were obtained from the National Meteorological Information Center (h ttp://data.cma.cn). In total, there are 20 meteorological stations in Beijing ( Fig. 1 , Table S2). Note that the sites for monitoring the meteorology were not always collocated with those for air quality observations. To address this mismatch, the nearest meteorological site for each air quality site was assigned based on the principle of proximity.The distance between the air quality and meteorological sites was calculated based on the Geosphere R package (Table S3), the corresponding matched results of which are shown in Table S1.

Fig. 2. Interannual variations of air pollutants from 1 January to 30 April 2015—20.

2.2. Definition of the study period

In this study, we divided the whole period into four stages as follows: P1 (1—19 January 2020) was the early stage before the outbreak of COVID-19; P2 (24—30 January 2020) was the Spring Festival when the First-Level Public Health Emergency Response was started; P3 (31 January to 9 February 2020) was the prolonged holiday, which coincided with strictest lockdown measures; and P4 (10 February to 30 April 2020) saw closed-offmanagement in the communities implemented and industries gradually restored.

2.3. RF models

RF is a machine learning algorithm that uses multiple trees to train and predict samples ( Breiman, 2001 ). Two basic scenarios were performed following previous study ( He et al., 2021 ). First, meteorological normalization was used to exclude the impact of meteorological conditions on the concentrations of air pollutants ( Grange et al., 2018 ). And second, a prediction experiment was used to predict concentrations of air pollutants under the scenario without lockdown ( Wang et al., 2020 ).The configurations of these scenarios were as follows:

2.3.1.

Meteorological

normalization

experiment

We established 35 models (RF1) based on each air quality site to achieve meteorological normalization. In RF1, hourly concentrations of air pollutants were the dependent variables; and the meteorological parameters, time predictors, and regional transport parameters served as the independent predictors (Table S4). RF1 was trained on datasets from 1 January to 30 April 2015—20. The training set accounted for 70% of the data, with the remaining 30% used as the testing set.

To remove the short-term meteorological impacts, meteorological datasets were randomly selected and replaced 1000 times from 2015 to 2020, rather than early 2020, to generate a new meteorological dataset.Specifically, meteorological parameters at a selected hour of a particular day in the 2020 meteorological data were replaced randomly with meteorological data from 2015 to 2020 within 2 weeks before and after the selected data point ( Vu et al., 2019 ; Zhang et al., 2020 ). For example,the new input meteorological data at 0600 LST 15 January 2020 were randomly selected from the observed data at 0600 LST on any date from 1 to 29 January of any year in 2015—20.

The above 1000 pieces of data were fed into the RF1 model to predict the concentrations of air pollutants. The 1000 predicted concentrations were then aggregated using the arithmetic mean to calculate the final meteorological normalized concentration (referred to as the “deweathered ”concentration). Relative changes (

R

) between the observed concentrations (

C

) and deweathered concentrations (

C

)for each pollutant were calculated using the following equation:

2.3.2.

Prediction

experiment

We further established 24 (4 site types ×6 pollutants) models (RF2)to predict hourly concentrations of air pollutants without lockdown. In RF2, we used the same dependent variables and independent predictors as in RF1. The training set accounted for 70% of the data, and the remaining 30% was the testing set. The RF2 model was trained on datasets from 1 January to 30 April 2015—19 (note that the dataset in 2020 was not used in this stage). Then, the RF2 model was used to predict the concentrations of air pollutants during the COVID-19 outbreak from 1 January to 30 April 2020. Relative changes (

R

) between the observed concentrations (

C

) and predicted concentrations (

C

) for each pollutant were calculated using the following equation:

In general, the RF models performed well for most sites. A detailed model evaluation, including the coefficients of determination (

R

), the fraction of predictions within a factor of 2 (FAC2), mean bias, normalized mean bias, root-mean-square error (RMSE), and Pearson correlation coefficient (PCC), is displayed in Tables S5 and S6. In brief, the model performance indicated a low bias, low RMSE, and high PCC. For RF1,

R

values ranged from 0.77 to 0.86 and FAC2 was greater than 0.76.For RF2,

R

values ranged from 0.57 to 0.87 and FAC2 was greater than 0.71.

3. Results

3.1. Changes in observed concentrations of air pollutants

We first analyzed the annual variation of air pollutants from 2015 to 2020, selecting the period from 1 January to 30 April as a whole ( Fig. 2 ).With the exception of O, the mean concentrations of PM, PM, SO,NO, and CO during January—April gradually decreased from 2015 to 2020, with a reduction of 9.2%, 10.4%, 16.4%, 8.8%, and 9.8% per year, respectively. This declining trend is indicative of the remarkable achievements with respect to the implementation of clean air actions in recent years ( Wang et al., 2019 ; Huang et al., 2020 ).

In addition, we calculated the overall mean concentrations of each pollutant for the period 2015—19 and compared them with those in 2020( Fig. 3 ). As shown, the concentrations of most air pollutants (except O)were relatively low in 2020, particularly those of SOand NO. In summary, the mean concentrations of PM, PM, SO, NO, and CO in 2020 were 31.5%, 41.0%, 69.2%, 36.8%, and 34.6% lower than the overall mean concentrations of 2015—19, respectively. The substantial decrease in air pollutants in 2020 was likely due to the reduction in anthropogenic emissions after the outbreak of COVID-19 ( Pei et al., 2020 ),although some pollution episodes still occurred unexpectedly (discussed below).

Fig. 3. Temporal variations of air pollutants in Beijing. The light gray area indicates 2015—19 variations in concentrations, with their average given by the thick gray line. The color line in each subfigure shows the hourly concentration of each pollutant in 2020. The color shadowed areas represent the different periods of P1(white, before lockdown), P2 (pink, Spring Festival), P3 (blue, during lockdown), and P4 (yellow, after lockdown).

While a decline in the annual trend was found, two severe haze episodes were also seen during 24—28 January and 6—13 February in 2020, with higher PM, PM, and CO concentrations than those of 2015—19. These two episodes coincided with unfavorable meteorological conditions, e.g., high relative humidity, low wind speed, and lowered mixed-layer height ( Le et al., 2020 ). One implication here is that the meteorological conditions may have masked the real changes in air pollutants, leading to the unexpected increase in observed concentrations. In the next section, to better understand the real changes in concentrations for each pollutant, we try to decouple the meteorological impacts.

3.2. Changes in deweathered concentrations of air pollutants

Meteorological normalization based on RF1 provides us with an opportunity to quantify the concentrations of air pollutants without the impact of meteorological conditions. Fig. 4 shows the deweathered concentrations against observations of each air pollutant at the four site types in Beijing, with detailed statistics listed in Table S7.

As shown in Fig. 4 and Fig. S1, the deweathered concentrations of air pollutants show a similar temporal pattern with the observations at different site types. Except for the P2 period, there is a minor difference between the deweathered and observed concentrations, with an

R

of only a few percent or higher. This result indicates that the meteorological conditions in 2020 did not differ significantly to those during 2015—19.During the period of P2, however, the deweathered concentrations of air pollutants were significantly lower than their observed values, indicating unfavorable meteorological conditions for pollutant dispersion( Zhang et al., 2020 ). In other words, these unfavorable meteorological conditions in P2 caused an increase by 92.0%—126.4%, 59.6%—87.0%,10.7%—25.3%, 15.2%—32.4%, and 42.2%—65.7% for the concentrations of PM, PM, SO, NO, and CO, respectively ( Fig. 4 ). Clearly, particulate concentrations had the most significant increase, with the

R

changing from 92% at the background sites to 126.4% at the urban sites.

After meteorological normalization, the deweathered concentrations of air pollutants, to some extent, reflected the real changes in emissions of the target air pollutants. As shown in Fig. 4 , the deweathered concentrations of air pollutants (except O) decreased significantly from P2 to P3 at all site types. Compared with the P2 period, the deweathered concentrations of PM, PM, SO, NO, and CO in P3 decreased by 15.8%—22.7%, 25.1%—53.6%, 27.6%—37.7%, 9.5%—15.6%, and 11.1%—12.5%, suggesting the large reduction was due to lockdown measures rather than meteorology.

3.3. Changes in predicted concentrations of air pollutants

After decoupling the effects of meteorology, we confirmed that the concentrations of most air pollutants decreased because of the lockdown measures during P3, while the impact level of lockdown on air quality remains unclear. Thus, here we applied RF2 to predict the concentrations of air pollutants assuming no lockdown measures. The predicted results are shown in Fig. 5 and Table S8.

With the exception of PM(in P2) and O 3 , the predicted concentrations of other pollutants were higher than the observations at most sites in P2—P4 (Fig. S2 and Fig. S3). The findings indicate that the decrease in primary pollutants can be attributed to the lockdown measures in Beijing ( Li et al., 2020 ; Wang et al., 2020 ). Similar changes in air pollutants during the COVID-19 pandemic have been found at other cities worldwide (Table S9).

Fig. 4. Observed and “deweathered ”concentrations of each air pollutant at four site types in different periods in Beijing. Numbers on bars indicate relative changes between the observed and deweathered concentrations.

Fig. 5. Observed and predicted concentrations of air pollutants at four site types in different periods in Beijing. Numbers on bars indicate relative changes between the observed and predicted concentrations.

Note that during the P1 period, there is no remarkable difference between the observed and predicted concentrations of any pollutant. This result may indicate that the emissions in P1 were not markedly different between 2020 and 2015—19. After the series of lockdown measures adopted in P3, however, a significant

R

of 36.5%—42.3%, 48.4%—52.5%, 48.6%—54.4%, 37.1%—51.0%, and 33.5%—37.8% was found for PM, PM, SO, NO, and CO, respectively. Similar decreases in air pollutants in P3 have been reported in Hebei Province, China, and the whole of Korea ( Jiang et al., 2021 ; Ju et al., 2021 ). With the resumption of “normal life ”and the reopening of industries in P4, on the other hand, the

R

shows a decreasing tendency compared to P3, indicating the increased emissions with the resumption of economic activities.Since there are 35 air quality monitoring sites, we further compared the

R

of each pollutant at different site types. The

R

of PMand SO2 in P3 at the transport sites were 42.3% and 54.4%, respectively,which was higher than those at other site types. Notably, PMhad the most significant relative change at the urban sites, due to the limitations imposed on the transportation and construction industries ( Zhang et al.,2017 ; Cheriyan et al., 2020 ).

Note that previous studies highlighted the reduction of NO 2 concentrations at transport sites in response to lockdown measures such as in Shanghai (from 58.4 to 39.4 μg m, decrease of 32.5%) ( Wu et al.,2021 ), Rome (50.8—22.4 μg m, 55.9%), Nice (46.9—14.6 μg m,68.9%) ( Sicard et al., 2020 ), and Madrid (42.7—15.9 μg m, 62.7%)( Baldasano, 2020 ). However, there are few reports on background sites.In our study, the most significant change for NOwas found at the background sites (31.4—15.4 μg m, 51%) rather than the transport sites(55.3—34.8 μg m, 37.1%). In addition, a slightly larger decline in CO at the background sites (37.8%) than transport sites (35.5%) was also found in this study. These findings indicated that the reduced traffic emissions in urban areas may have a wide impact on regional air quality during the pandemic ( Sicard et al., 2020 ).

3.4. Implications for future control policy

The COVID-19 pandemic provided a natural experiment to evaluate the impacts of human activities on air pollutants. We found that most air pollutants decreased significantly —by 39.4%, 50.1%, 51.8%, 43.1%,and 35.1% for PM, PM, SO, NO, and CO, respectively —during the lockdown period, due to substantial reductions in anthropogenic emissions. The largest reduction was found at transportation sties for PMand SO, at background sites for CO and NO, and at urban sites for PM. These reduction ratios may represent the upper limit of emission control for Beijing at the present economic and technological level.

Although the air quality has improved gradually in recent years and pollution levels reached a minimum in 2020 for most air pollutants, several severe haze events still occurred around the Spring Festival, with higher PM, PMand CO concentrations than those of 2015—19. These unexpected pollution episodes coincided with unfavorable meteorological conditions rather than enhanced emissions. In other words, the unfavorable meteorological conditions may have offset the decreases in concentrations of air pollutants caused by the reduction in emissions.Thus, stricter mitigation strategies in reducing emissions under stagnant weathers are still needed in the future.

Different from the decline trends of the other air pollutants, the concentration of Owere more stable or increased in recent years. In addition, the observed concentrations of Oduring P2 and P3 were 41.9%—84.3% and 11.3%—38.6% higher than its predicted values. This pattern was also different from the other air pollutants that a large reduction was found during lockdown periods. Some studies argue that the imbalanced reductions of NOversus volatile organic compounds (VOCs) lead to higher VOCs-NOratio, and hence enhanced the unintended increase of O( Sicard et al., 2020 ; Zhang et al., 2021 ). Besides, the reduction in particulate matter was also suggested to increase Oconcentrations via the heterogeneous chemical processes ( Li et al., 2017 ). Thus, future control policies will require a coordinated and balanced approach toward Oand PM, taking into full account both primary emissions and secondary processes.

Disclosure Statement

No potential conflict of interest was reported by the authors.

Acknowledgment

This work was supported by the National Natural Science Foundation of China (Grant number 42077204 ) and the National Key Research and Development Project (Grant number 2017YFC0210103),with data support provided by the National Earth System Science Data Center, National Science & Technology Infrastructure of China( http://www.geodata.cn ).

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.aosl.2021.100060 .