Quantifying anomalies of air pollutants in 9 U.S. cities during 2020 due to COVID-19 lockdowns and wildfires based on decadal trends

On March 11, 2020, the World Health Organization declared the coronavirus disease 2019 (COVID-19) outbreak a pandemic. In response, many U.S. cities and states enacted stay-at-home orders, or “lockdowns,” to combat the spread of the disease. This led to a reduction in traffic, and thus on-road emissions, beginning in mid-March 2020 when a significant portion of the workforce began working from home (e.g., U.S. Census Bureau, 2022).

Many scientific papers have been published assessing the impacts these lockdowns had on air quality around the world. Several studies showed that ozone (O₃) in the free troposphere and at remote measurement sites in spring and summer 2020 decreased as a result of COVID-19 lockdowns worldwide (e.g., Cristofanelli et al., 2021; Miyazaki et al., 2021; Steinbrecht et al., 2021; Chang et al., 2022). Additionally, Gkatzelis et al. (2021) reviewed 219 scientific papers accepted for publication through September 30, 2020, that quantified the effects of COVID-19 lockdowns on air quality for mostly urban regions across the globe. They found concentrations of nitrogen dioxide (NO₂), particulate matter (PM_2.5) with a diameter less than 2.5 µm, and carbon monoxide (CO) decreased for every continental-scale region examined, but that O₃ generally increased due to the lockdowns. However, they found median O₃ decreased slightly in North America. Gkatzelis et al. (2021) recommended that future analyses account for the effects of meteorology on pollutant concentrations, since meteorology can alter the magnitudes of pollutant changes on a similar scale as emissions change due to lockdowns. For studies that accounted for meteorology, Gkatzelis et al. (2021) found that the decrease in pollutant concentrations correlated with the severity of the lockdown, whereas studies that did not account for meteorology did not show a correlation between lockdown severity and pollutant concentration decreases.

Numerous studies have specifically analyzed the effects of lockdowns in U.S. metropolitan areas using both satellite and in situ measurements (e.g., Bauwens et al., 2020; Goldberg et al., 2020; Parker et al., 2020; Bray et al., 2021; Keller et al., 2021; Kondragunta et al., 2021; Laughner et al., 2021; Poetzscher and Isaifan, 2021; Shi et al., 2021; Sokhi et al., 2021; Yang et al., 2021; Cooper et al., 2022; Jing and Goldberg, 2022). Further study has shown how the lockdowns affected regional to state scale concentrations (e.g., Archer et al., 2020; Bekbulat et al., 2021; Jaffe et al., 2022). Here, we summarize published findings for each of the species (NO₂, O₃, PM_2.5, and CO) and cities (New York, NY; Los Angeles, CA; Riverside, CA; Chicago, IL; Washington, DC; Dallas, TX; Atlanta, GA; Denver, CO; and Seattle, WA) we examine in this study. Studies that used satellite measurements typically used either the TROPOspheric Monitoring Instrument (TROPOMI) aboard the Sentinal-5 Precursor satellite for measurements of NO₂ or the Ozone Monitoring Instrument aboard the Aura satellite for measurements of NO₂ and O₃. Studies that used in situ measurements typically used products from the U.S. Environmental Protection Agency (EPA) AirData network or the Air Quality System (AQS). These data are typically from roadside monitors. Several studies used preliminary data before quality control and quality assurance had been performed; however, in this study, we use final data that have passed these steps. More details are given in the Supplementary Material (§1).

Previous studies using satellite- and ground-based monitoring data found NO_x emissions generally decreased in the United States during the COVID lockdowns. Reported changes during the most stringent lockdown period, typically March and April, ranged from −10% to −51% in New York (Bauwens et al., 2020; Goldberg et al., 2020; Bray et al., 2021; Keller et al., 2021; Kodgragunta et al., 2021; Poetzcher and Isaifan, 2021; Sokhi et al., 2021; Cooper et al., 2022), −10% to −33% in Los Angeles (Goldberg et al., 2020; Parker et al., 2020; Keller et al., 2021; Kondragunta et al., 2021; Poetzcher and Isaifan, 2021; Shi et al., 2021; Sokhi et al., 2021; Yang et al., 2021), +3% to −19% in Chicago (Bauwens et al., 2020; Goldberg et al., 2020; Poetzcher and Isaifan, 2021; Cooper et al., 2022; Jing and Goldberg, 2022), −12% to −29% in Washington, DC (Bauwens et al., 2020; Goldberg et al., 2020; Keller et al., 2021; Poetzcher and Isaifan, 2021; Cooper et al., 2022), +29% to −12% in Dallas (Goldberg et al., 2020; Cooper et al., 2022), −21% to −29% in Denver (Goldberg et al., 2020; Keller et al., 2021; Poetzcher and Isaifan, 2021), −12% to −34% in Atlanta (Goldberg et al., 2020; Keller et al., 2021; Kondragunta et al., 2021; Poetzcher and Isaifan, 2021; Cooper et al., 2022), and −34% in Seattle (Keller et al., 2021).

Previous studies found that O₃ both increased in some parts of the United States and decreased in others, but the median change was negative (Gkatzelis et al., 2021). The effects on O₃ were less straightforward due to the complex nonlinear chemistry of NO_x (= NO + NO₂) and volatile organic carbon (VOC) species that form O₃, such that a decrease in NO_x emissions could either increase or decrease O₃ production, especially as the lockdowns continued into the summer O₃ production season. Reported changes during the most stringent lockdown period ranged from +2% to +15% in New York (Keller et al., 2021; Shi et al., 2021; Sokhi et al., 2021), +5.7% to −13% in Los Angeles (Campbell et al., 2021; Keller et al., 2021; Shi et al., 2021; Sokhi et al., 2021; Yang et al., 2021), −8% to −18% in Chicago (Campbell et al., 2021), +0.2% in Washington, DC (Keller et al., 2021), −3% to −35% in Dallas (Campbell et al., 2021), +7.6% to −1% in Denver (Campbell et al., 2021; Keller et al., 2021), −0.9% in Atlanta (Keller et al., 2021), and +6.7% in Seattle (Keller et al., 2021). We note that the U.S. EPA National Ambient Air Quality Standard for O₃ is 70 ppbv, which is typically not exceeded during the springtime. However, O₃ adversely affects human and crop health below this standard as well, so even small increases in springtime O₃ concentrations could affect the most sensitive populations and crop yields (e.g., Jerrett et al., 2009; Van Dingenen et al., 2009; Malley et al., 2017; Wang et al., 2019).

Previous studies have found PM_2.5 decreased due to COVID-19 lockdowns. PM_2.5 is both directly emitted and formed from secondary chemistry, so the effects of decreasing on road emissions are not straightforward. Reported changes during the most stringent lockdown period ranged from −14% to −90% in New York (Bray et al., 2021; Shi et al., 2021; Sokhi et al., 2021; Lam et al., 2022) and −10% to −40% in Los Angeles (Parker et al., 2020; Shi et al., 2021; Sokhi et al., 2021; Yang et al., 2021; Lam et al., 2022).

Finally, previous studies found CO decreased during the COVID-19 lockdowns. Reported changes during the most stringent lockdown period ranged from −10% to −25% in New York (Bray et al., 2021; Sokhi et al., 2021) and −21% in Los Angeles (Sokhi et al., 2021).

Not all of the published analyses corrected for meteorological factors when determining the changes due to COVID lockdowns. Many compared measurements in 2020 to similar time periods averaged in previous years. This method averages the effects of meteorology during the comparison time period, but such averaging is not possible for the single year 2020, which may be subject to anomalous weather. For example, during the beginning of the COVID-19 lockdowns in Los Angeles, a Los Angeles Times article (Barboza, 2020) quoting a spokesperson for the local air quality management district noted that the city had particularly stormy weather at the onset of the lockdowns, with unusually deep boundary layers, which would have diluted pollutants in the Los Angeles Basin regardless of lockdown stringency. Further, since the lockdowns extended well into the summer of 2020, it is important to note unusual meteorology that took place in different parts of the United States. For example, in spring of 2020, there was above-average precipitation in portions of the West Coast, the Great Lakes region, and the Southeastern United States, whereas the Pacific Northwest and central Rockies had lower than average precipitation (Gleason et al., 2021). Additionally, in summer of 2020, there was above-average precipitation in the Great Lakes and the Mid-Atlantic regions of the United States, but below average precipitation in the West and Northeast. Finally, in fall, the largest area of wildfires since at least 2,000 burned in the Western United States (Gleason et al., 2021). The unusual weather and events impacted each of the cities studied here and thus the air quality in each of the cities.

In this work, we separate the effects of lockdowns from the effects of meteorology. We use 10 years of historical ambient air quality monitoring measurements of NO₂, O₃, PM_2.5, and CO from urban sites and meteorological data to determine the effects of the lockdown on urban air quality in New York City, NY; Los Angeles, CA; Riverside, CA; Chicago, IL; Denver, CO; Seattle, WA; Washington, DC; Atlanta, GA; and Dallas, TX. We chose the 9 cities to represent various regions of the United States, with the exception of Riverside, which is a known receptor site for Los Angeles emissions. We analyze these data with a multiple linear regression (MLR) model that accounts for meteorology in a 10-year climatology for each city. We further analyze the coefficients of the regression model to determine the long-term trends and weekend effect in each city. From the MLR, we compare changes in pollutant concentrations accounting for meteorology, which we then compare with the severity of the lockdowns and an emissions inventory that accounts for the COVID lockdowns. In the second part of our analysis, we examine hourly data from individual monitoring sites averaged for 2-week periods to see how the diel cycle of concentrations may have changed during the lockdowns.

Air quality monitoring data for NO₂, O₃, PM_2.5, and CO were downloaded from the U.S. EPA AQS website (see Supplemental Material §2.a.). For this study, we use 24-h mean NO₂, PM_2.5, and CO and maximum daily 8-h average (MDA8) O₃. Where practical, we use data from the core-based statistical area (CBSA), which averages data from numerous sites throughout an urban region (e.g., see Ghosal and Saha, 2021 and Figure S1). The CBSA is defined by the U.S. Census Bureau to represent a metropolitan region, and data from these regions are averaged, prepackaged, and available for download from the AQS website. When the CBSA data set contained too few measurement sites, such that step changes occurred in the time series as sites were added or removed from the average CBSA observations, we use data from the primary measurement site in that city. These sites typically had high-resolution CO data (i.e., precision <100 ppbv) and one or more other pollutant measurements. Meteorological data measured at large airports were downloaded from the National Oceanic and Atmospheric Administration’s (NOAA’s) National Centers for Environmental Information for each urban area studied (see Supplemental Material §2.b.). For this analysis, we use 24-h mean temperature, relative humidity, and windspeed and total daily precipitation. All pollutant and meteorological data are publicly available. We use the Oxford COVID-19 Government Response Tracker (OxCGRT) devised by the University of Oxford as a metric to determine the severity of lockdowns (Hale et al. 2021; see Supplemental Material §2.c.). We also compare our analysis with an inventory compiled from the Fuel-Based Inventory for Vehicle Emissions (FIVE; Harkins et al., 2021), the Fuel-based Oil and Gas inventory (Francoeur et al., 2021), and an adjusted near-real-time National Emissions Inventory for 2017 (NEI17; U.S. EPA, 2020), herein referred to as FIVE+NEI17 (He et al., n.d.). More details about inventory data sources are discussed in the Supplemental Material (§2.d.).

The goal of the MLR analysis is to determine the effects of COVID-19-related lockdowns on pollutant concentrations of NO₂, O₃, PM_2.5, or CO separate from the effects of meteorology for 9 U.S. cities. We account for the effects of meteorology on pollution concentration with 12 monthly MLR models fit to meteorological parameters from 2010 to 2019 and then predict what pollutant concentrations would have been in the absence of COVID-related lockdowns in 2020 based on meteorological data and the model fit coefficients. The estimated pollutant concentrations are then compared to the measured concentrations during the COVID-19 pandemic. We interpret the difference as a change in pollution concentrations due to the lockdowns.

First, we establish a 10-year climatology by fitting daily data reported to the EPA AQS for each month from 2010 to 2019 using an MLR model similar to those described by Venter et al. (2020) and de Foy and Schauer (2019):

where x is MDA8 O₃, or 24-h mean NO₂, PM_2.5, or CO; i represents each month; Y is the year minus 2010; T is the 24-h mean temperature; RH is the 24-h mean relative humidity; P is the total daily precipitation; WS is the 24-h mean wind speed; WE is a dummy variable that is 0 for a weekday (Monday–Friday) and 1 for a weekend (Saturday and Sunday); and ε is the residual. Due to the log-normal distributions of the PM_2.5 observations, we fit the natural logarithm of those data, similar to that done by de Foy and Schauer (2019). For cases with negative PM_2.5 data, we add an offset of the smallest integer greater than zero necessary to make all data positive before calculating the natural logarithm. The model is run for each month of the year, which results in 12 monthly sets of fit coefficients and eliminates the need for more complex trigonometric functions to account for annual cycles (e.g., de Foy, 2018). The R² values for monthly fits are provided in Figure S2.

Next, we use the coefficients from the MLR model run together with meteorological data for 2020 to predict what the pollutant concentrations would have been in 2020 without COVID-19 travel restrictions (herein referred to as “predicted”). We then compare these predicted concentrations to the measured concentrations to assess the impact of the lockdowns on pollutant levels in the 9 study cities. Uncertainties in the model are calculated by adding in quadrature the precision of the model determined by the average of the model’s residuals for a given time period and the standard error of the means of the observed and predicted data sets.

For much of this analysis, we bin the predicted and observed data into biweekly periods in order to (i) improve P values when comparing the distributions of daily data and (ii) ensure an equal number of weekend days, which is especially important when examining NO₂ changes due to the weekend effect. Distributions of predicted concentrations and the observed measurements are compared using a two-sided t test and examined at a level of P < 0.10 (|α| < 0.05). Measurements meeting this criterion are either higher or lower than predicted and fall within the 5% wings of the t-test distribution.

In the second part of our analysis, we compare diel, biweekly distributions of hourly NO₂, O₃, odd oxygen (O_x = NO₂ + O₃), PM_2.5, and CO from late February, late March/early April, late May/early June, and late August/early September in 2020 to the same biweekly distributions from 2016 to 2019. Data are analyzed from one measurement site in each urban area (see Supplemental Material §1 and Figure S1 for more details). O_x is examined to determine the effects that presumed NO_x emission reductions may have on O₃ titration, especially at night, and O₃ formation during the daytime. Although the O_x budget is more complicated at night with the formation of NO₃ and N₂O₅, we are limited by available monitoring data to the simplest definition of O_x (e.g., Brown et al., 2006; Womack et al., 2019).

For the diel analysis, we do not explicitly account for the effects of meteorology as we did with the MLR. However, calculating a 3-year biweekly mean from 2017 to 2019 averages the effects of meteorology and creates a short-term climatology that may then be compared with the COVID time period of 2020. Yet the 2020 data may still be affected by anomalous meteorology. We note that this is a drawback of several published analyses that assert a meteorological correction, even though this correction only applies to the pre-COVID comparison period but not to any anomalous meteorology that may have occurred during 2020.

The comparisons of the distributions of biweekly diel hourly data are made between 2016 and 2019 and 2020, similar to the MLR analysis. Two-sided t tests are run on the data and examined at P < 0.10. For PM_2.5, the analysis is done on the natural logarithm of the data, but the results are presented in linear space.

An example MLR for the Denver CBSA is shown in Figure 1. Meteorological data were not available for 2013, leading to a gap in the modeled output that year. The coefficients from Equation 1 are given in Table S1. Two results of note from this MLR are that NO₂ has decreased since 2010 in every month but January, and the weekend effect for NO₂ is negative in every month (Table S1). This is to be expected because NO_x emissions have been decreasing from on-road sources in the United States over the past decade (e.g., Yu et al., 2021), and diesel trucks, a significant source of on-road NO_x emissions, are less active on the weekends (e.g., see Pollack et al., 2012). These findings are frequently confirmed in the scientific literature and provide a check that our MLR is working as expected.

In the following sections, we plot for each species the mean biweekly concentration from 2010 to 2019, the daily observations and their biweekly means, and the daily predictions and their biweekly means from January 1 to October 31, 2020, along with the ratio of the observed to predicted biweekly mean, the state-level OxCGRT (Hale et al., 2021), and weekly total FIVE+NEI17 emissions (if applicable). FIVE+NEI17 emissions were calculated in a 3 × 3 grid cell (12 × 12 km) box surrounding the primary measurement site for each city (Figure S1). The OxCGRT stringency index and FIVE+NEI17 indicate when emissions are expected to be lower due to the lockdowns. Ratios of observed to predicted concentrations that are different at a P < 0.10 level using a 2-tailed t test are highlighted with red or light blue open circles, indicating the observations are higher or lower, respectively, than the predictions at this threshold.

Figure 2 shows the MLR results for 24-h mean NO₂. CBSA data sets were used for all cities except Seattle and Atlanta. Between March 1 and April 26, the observed biweekly mean NO₂ was lower than predicted for 16 of 36 biweekly periods and higher than predicted for zero biweekly periods, for all cities combined. Decreases in weekly mean NO₂ at the time of the initial lockdowns are apparent for Denver, Los Angeles, New York City, Seattle, and Washington, DC. The prediction for March 15–28 in New York was skewed by a large precipitation event followed by 4 days without reported meteorological data.

Some additional anomalous events are apparent in the data as well. Wildfires affected air quality in Seattle and Los Angeles in September, when the ratio of observed to predicted NO₂ was often greater than one. This ratio was also higher in Atlanta from September 13–26; however, we attribute this to the MLR underpredicting NO₂ due to the anomalous effects of Hurricane Sally rather than unusually high observations (NOAA National Weather Service, 2020).

Figure 3 shows the MLR results for MDA8 O₃. CBSA data sets were used for all cities except Seattle. Between March 1 and April 26, the observed biweekly mean MDA8 O₃ was lower than predicted for 10 of 36 biweekly periods, and higher than predicted for zero biweekly periods, for all cities combined. Decreases in weekly mean MDA8 O₃ at the time of the initial lockdowns are apparent for Chicago, Los Angeles, Riverside, and Washington, DC, and perhaps Atlanta, Dallas, Denver, and Seattle. The prediction for March 15–28 in New York was again skewed by a large precipitation event followed by 4 days without reported meteorological data, as it was for NO₂.

The generally lower-than-predicted O₃ in the 9 cities is contrary to what might be expected from locally produced photochemical O₃. In the winter months, O₃ production is expected to be NO_x-saturated, such that a decrease in NO_x emissions would lead to an increase in O₃ (e.g., Jin et al., 2017; Womack et al., 2019; Coggon et al., 2021). Additionally, a decrease in NO_x emissions would lessen the local effect of O₃ titration, which would also increase O₃. However, the data clearly indicate a general decrease in O₃ in the 9 cities, which is consistent with the results compiled by Gkatzelis et al. (2021) for North America. The results are also consistent with those from Chang et al. (2022), who found a mean anomaly of −2.8 ppbv in the free troposphere over North America in 2020.

Some further anomalous events are also apparent in the data. Wildfires adversely affected air quality in Denver in August and in Los Angeles and Riverside in September, when the ratio of observed to predicted O₃ was often greater than 1. This is consistent with a recent box-modeling study of local and distant wildfire emissions producing O₃ in urban environments (Rickly et al., 2023).

Figure 4 shows the MLR results for 24-h mean PM_2.5. CBSA data sets were used for all cities. Between March 1 and April 26, the observed weekly mean PM_2.5 was lower than predicted for 10 of the 36 biweekly periods and higher than predicted 1 biweekly period, for all cities combined. Decreases in biweekly mean PM_2.5 at the time of the initial lockdowns are apparent for Los Angeles, New York City, Riverside, and Washington, DC, and perhaps Denver. The prediction for March 15–28 in New York was skewed as before.

Several anomalous events are readily apparent in the PM_2.5 data. First, a Saharan dust storm affected the southern cities (Atlanta and Dallas) in late June by transporting PM_2.5 to those regions (Francis et al., 2020). Atlanta had 4 days of 24-h mean PM_2.5 greater than 23 μg/m³ and well above predicted; however, the other 10 days between June 21 and July 4 were similar to the predictions, and therefore, the P value of the ratio of observed: predicted greater than one is not less than 0.10 during this period. Second, Independence Day celebrations on July 4, 2020, were apparent in observations from Chicago, Los Angeles, Riverside, Seattle, and Washington, DC, where 24-h mean PM_2.5 was 20, 44, 17, 17, and 34 μg/m³, respectively (Seidel and Birnbaum, 2015). However, none of those events were strong enough to make the biweekly ratio P < 0.10. Finally, wildfires noticeably affected air quality in Denver in August and September (Rickly et al., 2023), and Los Angeles, Riverside, and Seattle in September, when the ratio of observed to predicted PM_2.5 was often greater than one at P < 0.10.

Figure 5 shows the MLR results for 24-h mean CO. CBSA data sets were used for all cities except Atlanta, Chicago, Dallas, Seattle, and Washington, DC. Between March 1 and April 26, the observed biweekly mean CO was lower than predicted for 15 of the 36 biweekly periods and higher than predicted for zero biweekly periods, for all cities combined. Decreases in biweekly mean CO at the time of the initial lockdowns are apparent for Riverside and perhaps Chicago, Denver, Los Angeles, New York, and Washington, DC. The prediction for March 15–28 in New York was skewed by the large precipitation event and meteorological data gap, as it was for NO₂, O₃, and PM_2.5.

Anomalous events are also apparent in the CO data. Wildfires affected air quality in Denver in August and September and Los Angeles, Riverside, and Seattle in September, when the ratio of observed to predicted CO was often greater than 1 at P < 0.10.

The MLR analysis indicates that the COVID-19 lockdowns decreased concentrations of NO₂, O₃, PM_2.5, and CO in most of the cities studied, both collectively and individually. Next, we compare the MLR with the FIVE+NEI17 (He et al., n.d.) inventory for the 2-week period with the lowest inventory emissions during the lockdowns: March 29–April 11.

Figure 6 shows the ratios of the MLR observed:predicted and FIVE+NEI17 2020:2019 for March 29–April 11. O₃ is also included for completeness. The ratios are converted to a percentage change ([ratio − 1]×100). A background has been subtracted from the MLR observed and predicted CO before calculating the ratio in order to best compare to a change in emissions. We define the CO background as the intercept of a linear least squares fit to the CO and NO₂ observations between March 15 and April 15, 2019 (Table S2). This method (i) does not require the subtraction of an arbitrary background and (ii) has the added benefit of accounting for any offsets or biases the observations may have compared to their true value. The backgrounds for NO₂ and PM_2.5 are assumed to be zero for the comparison with FIVE+NEI17. We note that the attempts to define the backgrounds add uncertainty to the comparison, especially if the background levels have changed during the COVID time period. Uncertainties plotted in Figure 6 are determined by quadrature addition of the standard error of the mean observation, the standard error of the mean prediction, and the difference in the standard errors of the mean for the observation and prediction for the same time period in 2019. The MLR ratios are outlined in black when CBSA data are used and gray when data from only one site are used.

The MLR results generally match the changes in emissions from FIVE+NEI17 and those presented by Gkatzelis et al. (2021). During the 2-week period when the lockdowns were most severe, NO₂, O₃, PM_2.5, and CO decreased in 8, 9, 6, and 8 of the 9 study cities, respectively. The geometric mean decreases for the 9 cities are presented in Table 1 and compared with the geometric mean decreases in FIVE+NEI17 for those cities, and the median decreases for North America reported by Gkatzelis et al. (2021) for comparison. The range from the MLR is a 1σ geometric mean multiplied/divided by the geometric standard deviation. The MLR results agree to within the 1σ range with all the changes in FIVE+NEI17 and with the median values reported by Gkatzelis et al. (2021) with the exception of O₃, which was slightly lower. The MLR is not sensitive enough to determine if the decrease in PM_2.5 was due to a decrease in primary emissions, or a change in secondary production.

The results of the MLR in this study can be compared with those from previous studies using any date range specified by the previous study. Figure 7 plots the changes in NO₂ in Los Angeles and New York as determined by this and previous studies. Results for the other cities with multiple NO₂ measurement sites in their CBSA are shown in Figure S3 (Bauwens et al., 2020; Chen et al., 2020; Goldberg et al., 2020; Habibi et al., 2020; Keller et al., 2021; Poetzscher and Isaifan, 2021; Cooper et al., 2022; Jing and Goldberg, 2022). Error bars are the standard errors of the mean added in quadrature for this study and the reported uncertainties for previous studies. The changes are generally clustered about the 1:1 line with a few exceptions. For example, the Shi et al. (2021) and Cooper et al. (2022) studies reported less of a decrease in NO₂ than most other studies for Los Angeles and New York, including this one (Figure 7). The Shi et al. (2021) study used “deweathered” data similar to the method provided by Grange and Carslaw (2019) and Vu et al. (2019), which is difficult to compare with our method’s results, since machine learning methods do not produce fit coefficients for input variables that can be compared directly to those from MLR. Cooper et al. (2022) used TROPOMI satellite data and compared April 2020 measurements to April 2019 without accounting for the effects of meteorology. However, some studies compared well to the MLR. The NO₂ results from Goldberg et al. (2020), who used TROPOMI satellite data and corrected for meteorology, agreed well for Los Angeles and New York. Further, the R² for all 6 cities that overlap between the Goldberg et al. (2020) study and this one is 0.85.

Figure S4 shows the results of the change in NO₂ from previous studies plotted against the results of this work’s MLR for all cities, categorized by both meteorological correction (left panel) and whether the study used in situ versus remote sensing data (right panel). The MLR correlated best with studies that used both in situ data and corrected for meteorology (R² = 0.42), studies that used in situ data regardless of meteorological correction (R² = 0.41), and studies that corrected for meteorology, not including the outlying Cooper et al. (2022) study (R² = 0.38). This is not surprising, because our MLR used in situ data and corrected for meteorology, often used the same data sets as other studies. However, larger discrepancies remain when comparing in situ data with satellite data, even with the good agreement with the Goldberg et al. (2020) study mentioned above.

Figures S5–S7 show the results of the changes in O₃, PM_2.5, and CO, respectively, from previous studies plotted against the results of this work’s MLR for all cities with multiple measurement sites in their respective CBSAs (Chen et al., 2020; Habibi et al., 2020; Parker et al., 2020; Xiang et al., 2020; Bray et al., 2021; Huang et al., 2021; Keller et al., 2021; Shi et al., 2021; Sokhi et al., 2021; Wei, 2021; Yang et al., 2021; Ghahremanloo et al., 2022).

Our MLR results can also be compared with the OxCGRT stringency index. Figure 8 shows the ratio of biweekly means of observed to predicted concentrations, beginning with February 15–28, when most state stringency indices were zero, and ending April 26–May 9, once most states had settled into a consistently high lockdown. As with Figures 2–5, the ratios are plotted in gray when the observations are statistically similar to the predictions, red when the observations are greater than the predictions at P < 0.10, and light blue when the observations are less than the predictions at P < 0.10. The decrease in NO₂ correlates with the stringency of the lockdowns, similar to that reported by Gkatzelis et al. (2021), with a decrease in the ratio of −0.0022 ± 0.0006 per unit increase in the stringency index and an r value of −0.48. However, PM_2.5, O₃, and CO are not well-correlated with the stringency index, although there are far more lower-than-predicted observations at P < 0.10 than higher for PM_2.5 and CO.

The MLR results are consistent with our understanding of a reduction in emissions due to the COVID-19 lockdowns, but they also explain other features of urban emissions and chemistry between 2010 and 2019. For example, the MLR fits linear trends of each pollutant from 2010 to 2019. Annual trend coefficients for the summer months for locations with CBSA data are plotted in Figure 9. The trend coefficient for CO was negative for each summer month in the 4 cities, where CBSA data are available: Denver, Los Angeles, Riverside, and New York City. The NO₂ trend coefficient was either slightly negative or near zero for all cities. Similarly, the trend coefficient for PM_2.5 was negative or near zero, with the exception of August in Denver, where it increased at 3% ± 3% per year. Finally, the MDA8 O₃ trend coefficient is negative in the summer for Atlanta, New York City, and Washington, DC; however, it is positive in the early summer months in Chicago, Dallas, and Riverside, and the Los Angeles MDA8 O₃ trend coefficient was positive for every summer month.

The MLR calculates monthly the mean effect that weekends have on the concentration of the 4 studied species. The weekend coefficient shows that NO₂ decreased on the weekend in every city, as expected due to less commercial diesel traffic (e.g., Cleveland et al., 1974) in what is known as the “weekend effect.” However, the effect this has on O₃ is less certain, depending if a region is in a NO_x-limited or NO_x-saturated (also known as VOC-limited) regime. To study the weekend effect, we run the MLR on data from the primary measurement sites (i.e., not CBSA data) for each city that had both 24-h mean NO₂ and hourly O₃ measurements. We do so by calculating a 24-h mean O₃ value from hourly O₃ data reported to the AQS, since the prepackaged data sets from the AQS site do not properly calculate a 24-h mean from the hourly data. We perform the MLR analysis on hourly average O_x (= NO₂ + O₃), and then as a check run separate MLR analyses on NO₂ and 24-h mean O₃ to test whether the weekend coefficients sum to the same coefficient as for O_x. Figure 10 shows the monthly weekend coefficients for 24-h mean O_x, NO₂, and O₃ and for the 9 study cities, as well as the sum of the NO₂ and O₃ weekend coefficients. As expected, the sum of the NO₂ and O₃ coefficients is within the error bars of the O_x coefficient. We interpret titration to be the cause when the weekend O_x effect is zero, as it typically is in the winter months when photochemical production of O₃ is at a minimum. In the summer, the effects of weekends on O₃ are positive, but typically not more than the decreases in NO₂. As a result, the weekend effect on summertime O_x is generally negative. We imply from this that O₃ production near these measurement sites is NO_x-limited. For example, Atlanta, Dallas, Denver, and Washington, DC, have less O_x in the summer on the weekends, as do Chicago and New York City in August. Interestingly, the North Main St. site in Los Angeles measures more O_x during September weekends than expected from titration alone, although the other summer months are mainly driven by titration effects. This type of analysis could be done in future work with downtown sites, which capture emissions, and downwind receptor sites, which capture the results of O₃ chemistry. We note that if some locations use a molybdenum converter for NO₂ measurements (Dunlea et al., 2007), this would bias O₃ + NO₂ high relative to the true O_x.

In this section, we display hourly diel plots from single measurement sites in the 9 study cities for NO₂, O₃, O_x, PM_2.5, and CO. Data are binned biweekly for the following time periods: February 15–28, before the most stringent lockdowns took effect; March 29–April 11, when most cities were at maximum lockdown stringency; May 24–June 6, when most cities were at a well-established lockdown stringency; and August 30–September 12, a late summer period. Data from 2020 are compared with the 4-year mean and variability from 2016 to 2019. Note that this analysis differs from the MLR analysis because meteorological effects and annual trends are not accounted for. Distributions between 2020 and 2016–2019 are compared with a 2-sided t test at a P < 0.10 level, as done for the MLR analysis above, and displayed in Figures 11–13 and S8–S10. The distributions in 2020 that were higher than 2016–2019 at P < 0.10 are displayed with an upward-pointing triangle, and those that were lower at P < 0.10 are displayed with a downward-pointing triangle. Similarly, at the side of each plot, an upward-pointing black triangle indicates that the distribution of the 24-h mean observations was higher than the MLR-predicted values at P < 0.10, and a downward-pointing triangle indicates the observations were lower than the MLR-predicted values at P < 0.10 during the same biweekly period.

NO_x emissions were expected to decrease due to the COVID-19 lockdowns (He et al., n.d.).

During the prelockdown phase, none of the cities experienced different NO₂ concentrations at P < 0.10 than predicted by the MLR (Figure 2). The Chicago measurement site was the only location that experienced multiple hours of lower NO₂ at P < 0.10 in the diel average over this time (Figure 11), but this effect was not seen for the CBSA data set and MLR analysis. Data were not available in Dallas for this time period.

At the time of the most stringent lockdowns when the MLR analysis shows Denver, Los Angeles, New York, and Washington, DC, had lower NO₂ in the CBSA at P < 0.10 (Figure 2), the diel profiles for these cities also had lower hourly means at P < 0.10 throughout the day (Figure 11). Riverside provides a good example of how the 2016–2019 average differs from the MLR analysis, which accounts for meteorology. Figure 2 shows that for March 29–April 11, NO₂ in Riverside was well below the 2010–2019 mean, and similarly, Figure 11 shows it was also well below the 2016–2019 mean. However, the MLR predicted lower NO₂ due to the cooler weather Riverside experienced during this time (see also Figure S7), so the lower values in the diel plot are likely due to meteorology, as predicted by the MLR.

For the late spring time period, Chicago, Denver, New York, Seattle, and Washington, DC, had both lower NO₂ than predicted by the MLR at the P < 0.10 threshold (Figure 2), and large portions of the day with lower hourly NO₂ than the previous 4-year averages (Figure 11). Atlanta, Dallas, and Los Angeles also had lower diel NO₂ for large portions of the day (Figure 11), but the MLR NO₂ observed: predicted ratios <1 did not meet the P < 0.10 threshold (Figure 2).

By late summer, only Washington, DC, had lower NO₂ at P < 0.10 in the MLR analysis (Figure 2), which is reflected by the lower NO₂ at P < 0.10 during the early morning hours (Figure 11). The diel plots show lower NO₂ at P < 0.10 for significant portions of the day in Atlanta, Chicago, Dallas, Denver, New York, and Seattle, but not in the MLR analysis. The wildfires that impacted Los Angeles, Riverside, Denver, and Seattle at this time did not impact the diel NO₂ profiles at the P < 0.10 level; however, they did for the following week in Los Angeles and Seattle in the MLR analysis.

Here, we examine hourly O₃ data rather than the MDA8 O₃ because the MDA8 calculation is not equally weighted throughout the day.

From the MLR analysis, Dallas, New York, and Washington, DC, had higher MDA8 O₃ than predicted by the MLR for the late February time period (Figure 3). However, the diel profiles for these cities had only 1 or 2 h that were above the 2016–2019 means at the P < 0.10 threshold. The diel profiles for the other cities mostly agreed with the 2016–2019 mean, with the exception of lower O₃ in the morning hours in Los Angeles (Figure 11), which may have been due to warmer than average temperatures (Figure S7).

Diel differences during the lockdown period were variable. From March 29 to April 11, the observations of MDA8 O₃ were lower than predicted in Chicago, Los Angeles, and Seattle. The diel profiles for Chicago and Los Angeles show several hours during the day had lower O₃ than the previous 4-year means (Figure 11). However, for Seattle, the O₃ was generally higher than for 2016–2019 for most of the day, even though the MLR analysis resulted in lower O₃ than predicted. This was due to meteorological effects, with the MLR predicting higher than average O₃ for that time period.

For the late spring period, the observations of MDA8 O₃ were lower in Chicago, Denver, and New York than the MLR predicted. The diel profile for Denver was also lower than the previous 4-year mean, but not so for Chicago and New York.

For the late summer period, only Washington, DC, had lower MDA8 O₃ than predicted by the MLR analysis at P < 0.10. The diel profile also had lower O₃ throughout the afternoon at a P < 0.10 threshold. The diel profile for Atlanta revealed most afternoon hours had lower O₃ than the 2016–2019 mean, while the lower MDA8 O₃ in the MLR had P = 0.113. Dallas also had many hours of hourly O₃ less than the 2016–2019 mean at P < 0.10. The MLR also predicted much lower MDA8 O₃ relative to the 2010–2019 mean during this time (Figure 3), and although the observations were also lower, they were higher than predicted. Thus, meteorology likely played a significant role in the low O₃ in Dallas at this time, which is also shown by the lower-than-average temperatures for this period (Figure S7). Finally, the higher O₃ at P < 0.10 in Riverside in the afternoon (Figure 11) may have been due to early wildfire impacts (e.g., Robinson et al., 2021), which affected MDA8 O₃ in both Riverside and Los Angeles the following 2 biweekly periods (Figure 3).

Changes in O₃ may be caused by several phenomena: (1) changes in NO, which titrates O₃ to form NO₂, (2) a different photochemical formation regime, or (3) changes in background O₃ entering the urban area. One way to examine changes due to titration with NO is to examine odd oxygen (O_x = O₃ + NO₂), especially at nighttime when there is little or no photochemical formation of O₃. Here, we add hourly NO₂ data with hourly O₃ data to calculate hourly O_x data. We expect O_x concentrations to have decreased in many U.S. cities during the lockdowns, since both NO₂ and median O₃ decreased in North America (Gkatzelis et al., 2021).

For the prelockdown time period, most cities had similar O_x concentrations as compared with 2016–2019. Calculated O_x data were not available in Dallas due to the missing NO₂ data during this time period.

At the height of the COVID-19 lockdowns, O_x was found only to be lower than the 2016–2019 mean at P < 0.10 and never higher. However, neither Atlanta nor Seattle had lower O_x at P < 0.10 during this time, despite being lower for most every hour of the day. This is to be expected from reduced NO_x emissions, since hourly NO₂ was also only lower, never higher, at the P < 0.10 level (Figure 11). O₃, however, was higher at night in New York, Riverside, and Seattle and lower during the daytime in Chicago, Dallas, Los Angeles, and Riverside (Figure 12). The nighttime values in Riverside indicate titration were the main reason for lower NO₂ but higher O₃ at night, with O_x similar to the 2016–2019 mean. Lower daytime O_x in Riverside can be attributed to lower O₃ and lower temperatures relative to 2016–2019 and possibly the lower NO₂ in Los Angeles. This is similar to NO₂ (Figure 11), which was only lower and never higher at the P < 0.10 level.

By late spring, O_x was lower in most places and lower in Atlanta, Denver, and Los Angeles at P < 0.10 for a large portion of the day.

In late summer, O_x was lower throughout the day in Atlanta, Dallas, Denver, and Washington, DC, at P < 0.10, but higher for much of the day in Riverside at P < 0.10. This may indicate that a changing chemical regime with fewer NO_x emissions affects O₃ formation in these cities. However, it is also important to note that Denver and Dallas experienced cooler temperatures through this 2-week period (Figure S7).

For the diel analysis, the PM_2.5 data distributions were offset by the nearest integer to make all data positive, log-transformed and statistically analyzed, then transformed back into linear space for the plots, as was done for the MLR (Figure S8). As with O₃, changes in PM_2.5 may be due to both primary and secondary effects. Although primary emissions are expected to decrease during the lockdowns, the effects of reduced NO_x and VOC emissions on the secondary formation of PM_2.5 are expected to be more variable.

For the prelockdown period, Atlanta, Dallas, and Denver had lower observed PM_2.5 in their CBSAs than predicted from the MLR at P < 0.10. However, only the Dallas measurement site had lower PM_2.5 at P < 0.10 in the diel profile from a single site. Los Angeles had higher CBSA PM_2.5 in the MLR analysis at P < 0.10, and had higher PM_2.5 at P < 0.10 at the North Main St. measurement site.

At the height of the lockdowns, only Washington, DC, had lower CBSA PM_2.5 in the MLR analysis at P < 0.10, and in the diel profile, most hours of the day had lower PM_2.5 than 2016–2019. However, Los Angeles and Seattle had lower hourly PM_2.5 at P < 0.10 for some or most of the day. This may correspond with lower temperatures limiting secondary production.

In the late spring period, only Denver had lower CBSA PM_2.5 in the MLR analysis at P < 0.10, and in the diel profile, the evening and nighttime hours had lower PM_2.5 at P < 0.10. Seattle also had lower PM_2.5 at P < 0.10, but may be due to measurement error, as some of the data reported were at or below zero.

Wildfires significantly impacted the MLR analysis for Los Angeles, Riverside, and Seattle from August 30 to September 12. This is also apparent in the diel profile for the Rubidoux site at Riverside and the North Main St. site in Los Angeles. Once again, the measurement site in Seattle reported many values below zero. Finally, the measurement site in Chicago did not report data for this time period.

CO emissions were expected to decrease with decreasing traffic emissions during the COVID lockdowns (He et al., n.d.).

CO in late February was lower in the MLR analysis at P < 0.10 for Chicago and Los Angeles. Several hours in the diel profile had lower CO for the measurement site in Chicago (Figure S9), which is expected since we did not use the CBSA data, but the North Main St. site in Los Angeles experienced higher CO for much of the day in contrast to the CBSA data used in the MLR analysis.

By March 29–April 11, CO was lower in the MLR analysis at P < 0.10 for Chicago, Denver, and Riverside. The diel profiles of CO for Denver and Riverside also have lower CO throughout the day. The diel profile was lower during much of the day at the measurement sites in New York, Los Angeles, and Washington, DC, although not in the MLR analysis.

The MLR analysis revealed lower CO at P < 0.10 in Denver, Los Angeles, and Washington, DC, from May 24 to June 6, and the diel profile at the Denver Camp site was also lower at P < 0.10 throughout the day. CO was also lower at the Pfizer site in New York at P < 0.10 for much of the day. However, although the Los Angeles CBSA experienced lower CO during this time, CO at the North Main St. site was similar in 2020 as it was from 2016 to 2019.

In late August, CO was higher in the MLR analysis at P < 0.10 for Los Angeles, New York, Riverside, and Seattle. Los Angeles, Riverside, and Seattle were all affected by wildfire emissions at this time, which can be seen in the diel profile from the Rubidoux site in Riverside. The North Main St. site in Los Angeles did not report hourly data for this time period. Washington, DC, had lower CO than predicted by the MLR analysis, and the diel profile had lower CO around noon.

It is difficult to directly compare the diel profiles with the MLR analysis. First, the MLR analysis used CBSA data where available, whereas the diel profiles are from single measurement sites in each city. Second, even when the MLR analysis used data from a single measurement site, it determined a 10-year climatology and predicted 2020 values after accounting for meteorology. The diel profiles instead compared 2020 data, uncorrected for meteorology, with data from 2016 to 2019, which averages out the effects of meteorology over that 4-year period. Even so, the diel profiles are useful for determining the time of day where observations that differed from those predicted by the MLR analysis likely occurred. For example, in the late summer, Washington, DC, saw less NO₂ and less O₃ according to the MLR at the P < 0.10 threshold, and the diel plots show that the most significant reductions in O₃ were in the afternoon, which indicates that lower NO_x emissions led to less photochemical production of O₃. Similarly, Los Angeles saw less NO₂ according to the MLR at the P < 0.10 threshold in the early spring, but the diel profile shows that the reduction occurred throughout the day. O₃ was also lower in Los Angeles at the P < 0.10 threshold according to the MLR, and the diel profile showed O₃ was lower in midday, which may indicate some photochemical O₃ formation even in early spring. This analysis therefore indicates where a more in-depth chemical modeling study of these cities would be useful for air quality managers in these regions.

The COVID-19 pandemic led numerous state and local governments in the United States to enact lockdowns to control the spread of the virus. These lockdowns resulted in fewer on-road emissions when a large portion of the workforce began working from home. In this study, we examined pollutant concentrations of NO₂, O₃, PM_2.5, and CO from Atlanta, GA, Chicago, IL, Dallas, TX, Denver, CO, Los Angeles, CA, New York, NY, Riverside, CA, Seattle, CA, and Washington, DC, in 2020 using roadside monitoring data to see how they may have changed during the pandemic. We used 11 years of pollutant and meteorological data and an MLR model to account for meteorology and found geometric mean concentrations of NO₂, MDA8 O₃, PM_2.5, and CO decreased in the 9 cities between March 29 and April 11 by 19%, 8%, 8%, and 17%, respectively, consistent with previous literature and a fuel-based near-real-time emissions inventory. The decreases in concentrations were poorly correlated with lockdown stringency index severity with the exception of NO₂, but the number of lower concentrations far outweighed the number of higher concentrations than predicted in the absence of the lockdowns at P < 0.10, which we attribute to the COVID-19 lockdowns.

Results of changes in NO₂ due to COVID-19 lockdowns reported by previous studies for the 9 cities examined here correlated best with our MLR when they used in situ data and corrected for meteorology. It remains to be seen if comparisons between studies using satellite data and in situ data will improve as satellite instrumentation and retrieval algorithms improve in the future.

We further found the linear trends of NO₂, PM_2.5, and CO decreased in the summer months between 2010 and 2019 for all of the cities studied, which indicates on-road emissions controls have worked in reducing emissions in the past decade. However, MDA8 O₃ trends varied over this time. This indicates the nonlinear effects that changing emissions have on O₃ production and the importance of increasing background O₃. An analysis of weekend decreases in NO₂ was consistent with previous literature; however, the weekend increases in O₃ were typically dominated by changing titration effects. Finally, we found that western U.S. cities were impacted by wildfire emissions, often to a degree greater than the decreases due to the COVID-19 lockdowns.

All data are free and publicly available at https://www.epa.gov/aqs for the chemical data and https://www.ncei.noaa.gov/pub/data/noaa/isd-lite/ for the meteorological data.

The supplemental files for this article can be found as follows:

Supplemental Material.pdf

JP, KCA, ORC, CH, and K-LC were supported by the NOAA cooperative agreements with CIRES, NA17OAR4320101, and NA22OAR4320151.

The authors declare no competing interests.

Contributed to conception and design: JP, BCM, SSB.

Contributed to acquisition of data: KCA, JP.

Contributed to analysis of data: JP, KCA.

Contributed to interpretation of the data: JP, KCA, BCM, CH, AMM, AOL, ORC, K-LC, SSB.

Drafted and/or revised the article: JP, KCA, BCM, CH, AMM, AOL, ORC, K-LC, SSB.

Approved the submitted version for publication: JP, KCA, BCM, CH, AMM, AOL, ORC, K-LC, SSB.

How to cite this article: Peischl, J, Aikin, KC, McDonald, BC, Harkins, C, Middlebrook, AM, Langford, AO, Cooper, OR, Chang, K-L, Brown, SS. 2023. Quantifying anomalies of air pollutants in 9 U.S. cities during 2020 due to COVID-19 lockdowns and wildfires based on decadal trends. Elementa: Science of the Anthropocene 11(1). DOI: https://doi.org/10.1525/elementa.2023.00029

Domain Editor-in-Chief: Detlev Helmig, Boulder AIR LLC, Boulder, CO, USA.

Associate Editor: Allen Goldstein, University of California Berkeley, Berkeley, CA, USA

Knowledge Domain: Atmospheric Science