A comprehensive study of the COVID-19 impact on PM2.5 levels over the contiguous United States: A deep learning approach.

We investigate the impact of the COVID-19 outbreak on PM2.5 levels in eleven urban environments across the United States: Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, Phoenix, and Seattle. We estimate daily PM2.5 levels over the contiguous U.S. in March-May 2019 and 2020, and leveraging a deep convolutional neural network, we find a correlation coefficient, an index of agreement, a mean absolute bias, and a root mean square error of 0.90 (0.90), 0.95 (0.95), 1.34 (1.24) μg/m3, and 2.04 (1.87) μg/m3, respectively. Results from Google Community Mobility Reports and estimated PM2.5 concentrations show a greater reduction of PM2.5 in regions with larger decreases in human mobility and those in which individuals remain in their residential areas longer. The relationship between vehicular PM2.5 (i.e., the ratio of vehicular PM2.5 to other sources of PM2.5) emissions and PM2.5 reductions (R = 0.77) in various regions indicates that regions with higher emissions of vehicular PM2.5 generally experience greater decreases in PM2.5. While most of the urban environments ⸺ Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, and Seattle ⸺ show a decrease in PM2.5 levels by 21.1%, 20.7%, 18.5%, 8.05%, 3.29%, 3.63%, 6.71%, 4.82%, 13.5%, and 7.73%, respectively, between March-May of 2020 and 2019, Phoenix shows a 5.5% increase during the same period. Similar to their PM2.5 reductions, Washington DC, New York, and Boston, compared to other cities, exhibit the highest reductions in human mobility and the highest vehicular PM2.5 emissions, highlighting the great impact of human activity on PM2.5 changes in eleven regions. Moreover, compared to changes in meteorological factors, changes in pollutant concentrations, including those of black carbon, organic carbon, SO2, SO4, and especially NO2, appear to have had a significantly greater impact on PM2.5 changes during the study period.

Introduction

On the last day of 2019, China reported the first cases of an abnormal lung infection in Wuhan to the World Health Organization (WHO) (Ghahremanloo et al., 2021a). On January 7, 2020, WHO announced the identification of a new virus, severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), and on January 30, 2020, a worldwide public health emergency was declared by WHO (WHO, 2020). In the United States, the first case was reported to the Centers for Disease Control and Prevention on January 20, 2020, and the U.S. president declared a public health emergency on January 31, 2020 (Aubrey, 2020; Robertson, 2020). The coronavirus disease 2019 (COVID-19) is an extremely infectious respiratory illness, especially in those who have been in close contact with already infected individuals (Bherwani et al., 2020; Cascella et al., 2021). Therefore, its rapid spread can be controlled by implementing suitable social distancing measures and avoiding crowded places such as public transport, stadiums, and even family gathering spots (Bherwani et al., 2020; Gautam, 2020). Since the outbreak of the COVID-19 pandemic, many countries around the world have implemented lockdown or stay-at-home advisory strategies to combat the spread of the virus (Ghahremanloo et al., 2021a), and half of the world population has experienced a degree of lockdown by the end of March 2020 (Tosepu et al., 2020). Despite these lockdowns, or stay-at-home advisory strategies, as of December 2021, more than 5.3 million deaths and 276 million confirmed cases of the COVID-19 globally have been reported (https://covid19.who.int/).

Lockdown situations, and the following decrease in human activity, have led to major reductions in pollution levels, providing researchers with a unique opportunity to investigate the impact of reduced human activity on air pollution levels. Human activity, directly or indirectly, is one of the main factors responsible for various types of pollutant emissions into the atmosphere. One of the major pollutants in the atmosphere is particulate matter with an aerodynamic diameter of less than 2.5 μm (PM2.5) (Ghahremanloo et al., 2021b). Although mucociliary clearance eliminates the majority of the particulate matter we breathe, the PM2.5 fraction can reside inside the lungs and penetrate the circulatory system (Ghahremanloo et al., 2021b). Because of the serious health effects of PM2.5, such as respiratory disease (Polezer et al., 2018) and cardiovascular illness (Hayes et al., 2020), WHO has listed PM2.5 as a major public health threat (Lancet, 2006). A number of studies have reported significant reductions in concentrations of PM2.5 and other pollutants in various regions of the world, such as the United States (Chen et al., 2021; Pan et al., 2020; Zangari et al., 2020), China (Yin et al., 2021), East Asia (Ghahremanloo et al., 2021a), Europe (Baldasano, 2020; Tobías et al., 2020), and South America (Krecl et al., 2020) resulting from the COVID-19 lockdowns, or stay-at-home advisory strategies. For instance, Ghahremanloo et al. (2021a) used satellite images to examine the impact of COVID-19 on air pollution levels in the Beijing-Tianjin-Hebei (BTH) and Wuhan regions of China, and in Seoul, South Korea, and Tokyo, Japan. According to their results, column densities of NO2, HCHO, SO2, and CO in Wuhan, China decreased by 83%, 11%, 71%, and 4%, respectively, in February 2020 compared to those in February 2019. They also found a 62% decrease in aerosol optical depth (AOD) levels during the same period in Wuhan. Lal et al. (2020) also reported a substantial global decrease in the AOD (0.1–0.2) and the column densities of CO (<1.81 × 1018 molecules cm−2) and NO2 (1.2 × 1015 molecules cm−2) during February–March 2020 compared to those on similar days in 2019. By comparing surface concentrations of pollutants before and during the lockdown in Salé, Morocco, Otmani et al. (2020) found a major decrease in concentrations of PM10 (75%), SO2 (49%), and NO2 (96%). Moreover, COVID-19 lockdowns, or stay-at-home advisory policies resulted in PM2.5 reductions of almost 30% in central China (Xu et al., 2020), 5.9% in northern China (Bao and Zhang, 2020), and 35–39% in several cities in India (Chauhan and Singh, 2020; Mahato et al., 2020; Parida et al., 2021). Rodríguez-Urrego and Rodríguez-Urrego (2020) investigated 50 capital cities across the globe and reported both increased and decreased PM2.5 levels during the COVID-19 outbreak. According to their results, several capital cities such as Tokyo (Japan), Paris (France), London (Britain), Vienna (Austria), and Jakarta (Indonesia) experienced an increase in PM2.5 concentrations, while other cities such as Delhi (India), Dhaka (Bangladesh), Tehran (Iran), Bogota (Colombia), and Mexico City (Mexico) reported reductions. Since most of the PM2.5 ingredients are short-lived pollutants, it is also worth conducting similar studies to investigate the impact of the COVID-19 lockdowns on air pollution levels in receipt regions like the Arctic. Previous studies have shown a positive impact of the COVID-19 lockdown on other pollutants such as ozone (Bouarar et al., 2021; Khan et al., 2021; Tahir and Batool, 2020) in Arctic regions.

Several studies, most of which have used either surface measurements or satellite images for their analysis, have examined the impact of the COVID-19 pandemic on air pollution levels. Although their pollutant surface observations at ground stations have been precise, they took place at a limited number of ground stations over the United States and globally. Thus, because of the limited spatial coverage of the ground stations, studies have not been able to determine precise surface pollution levels, especially those in urban and suburban regions, with significant changes in the spatial distribution of pollutants (Ghahremanloo et al., 2021b). In addition, satellite instruments that measure the column densities of pollutants are seriously limited at capturing surface concentrations of pollutants, required in public health studies. The high frequency of missing values in satellite images results in data loss, restricting high-quality analyses of pollution reductions resulting from reduced human activity (Lops et al., 2021). Such limitations further highlight an essential need to develop advanced models that capture the spatiotemporal distribution of pollutants at high resolutions, particularly in regions with limited or no surface measurements, to improve the accuracy of analyses regarding the impact of COVID-19 on air pollution levels. Numerous studies have leveraged machine learning (ML) and deep learning (DL) to capture the spatiotemporal distribution of pollution levels (de Hoogh et al., 2019; Eslami et al., 2020; Ghahremanloo et al., 2021b, 2021c; Hu et al., 2017; Lops et al., 2020; Park et al., 2020; Sayeed et al., 2020; Yeo et al., 2021). Ghahremanloo et al. (2021b) used random forest (RF) to achieve a correlation coefficient (R) and a mean absolute bias (MAB) of 0.83–0.90 and 1.47–1.77 μg/m3, respectively, in high-resolution PM2.5 estimation. Hu et al. (2017) also used RF to incorporate various predictor variables such as satellite AOD and meteorological factors to estimate PM2.5 levels over the contiguous United States (CONUS) in 2011. Their model also showed high accuracy (R = 0.89) of surface PM2.5 estimation over the CONUS. Moreover, Ghahremanloo et al. (2021c) leveraged a deep convolutional neural network (Deep-CNN) to estimate the spatiotemporal distribution of surface NO2 levels at high resolution. Their results revealed the promising performance of Deep-CNN at capturing surface NO2 levels, with an R and an MAB of 0.91 and 1.75 ppb, respectively.

This study presents a DL approach to investigating the impact of COVID-19 lockdowns, or stay-at-home advisory strategies on changes in PM2.5 concentrations in eleven metropolitan areas over the CONUS. As mentioned before, previous studies mainly used either ground stations or satellite images to examine changes in air pollution during the COVID-19 pandemic. However, when monitoring PM2.5 changes over large urban environments, both approaches have serious limitations, discussed in the previous paragraph. Therefore, this study leverages Deep-CNN to produce accurate daily PM2.5 maps at a 5-km spatial resolution over the CONUS to study changes in PM2.5 levels from March to May 2020, compared to similar days in 2019. The output of daily PM2.5 grids over the CONUS do not contain missing values to improve the quality of analyses. Moreover, we use Google Community Mobility Reports to study people's mobility changes during the study period. We believe that this study is the first to comprehensively investigate PM2.5 changes resulting from shutdowns and stay-at-home strategies during the COVID-19 outbreak in several metropolitan areas over the CONUS.

Study area and data

The contiguous United States (CONUS)

We analyze daily changes in PM2.5 concentrations in eleven urban environments over the CONUS (Fig. 1 ) from March 1 to May 31, 2020, compared to similar days in 2019. We have selected eleven urban environments ⸺ Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, Phoenix, and Seattle ⸺ based on their economic importance, pollution levels, populations, and areas. For the analyses, we have defined eleven boundaries around each urban environment to include all metropolitan areas, including downtown and suburban regions. Table 1 shows the latitude and longitude coordinates of the boundaries around the urban environments analyzed in this study, along with the exact dates of the lockdowns, or stay-at-home orders issued by each city (Wu et al., 2020). This table also displays the counties inside each region. Of note, the boundaries around three urban environments (Washington DC, New York, and Philadelphia) include more than one state. Therefore, the information of all included states appears in Table 1.

Fig. 1

Map of the CONUS (study areas). Pink dots represent the location of the eleven urban environments analyzed in this study. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Table 1

List of urban environments analyzed in this study. The Lat. and Long. columns refer to the latitude and longitude coordinates of the edges of the boundaries around each region. The fourth column lists the multiple counties inside each region, and the last column lists the effective dates of the stay-at-home/lockdown orders issued in each state.

Region	Lat. (°N)	Long. (°W)	Counties Included	Effective Dates
Washington DC	38.60 to 39.60	77.30 to 76.10	DC, Baltimore, Arlington, Alexandria, Prince George's, Howard, Anne Arundel	Apr. 1 - May 15 - DCMar. 30 - May 7 - MD
New York	40.30 to 41.10	74.60 to 73.30	New York, Kings, Queens, Bronx, Hudson, Richmond, Essex, Nassau, Union, Bergen	Mar. 22 - May 15 - NYMar. 24 - May 2 - NJ
Philadelphia	39.80 to 40.20	75.40 to 74.75	Philadelphia, Delaware, Camden, Montgomery, Bucks, Burlington	Apr. 1 - May 8 - PAMar. 24 - May 2 - NJ
Dallas	32.40 to 33.25	97.60 to 96.30	Dallas, Tarrant, Denton, Collin, Rockwall, Kaufman	Apr. 2 - Apr. 30 - TX
Los Angeles	33.40 to 34.60	119.0 to 117.0	Los Angeles, Orange, Riverside, Ventura	Mar. 19 - May 8 - CA
Chicago	41.45 to 42.50	88.20 to 87.25	Cook, DuPage, Lake, Will	Mar. 21 - Apr. 30 - IL
Boston	42.20 to 42.54	71.30 to 70.80	Suffolk, Norfolk, Middlesex	Mar. 24 - May 4 - MA
Houston	29.35 to 30.20	96.0 to 94.75	Harris, Fort Bend, Galveston	Apr. 2 - Apr. 30 - TX
Detroit	41.90 to 42.75	83.60 to 82.60	Wayne, Oakland, Macomb	Mar. 24 - May 15 - MI
Phoenix	33.17 to 33.98	112.5 to 111.5	Maricopa, Pinal	Mar. 31 - Apr. 30 - AZ
Seattle	47.12 to 48.06	122.6 to 122.0	King, Snohomish	Mar. 23 - May. 4 - WA

Datasets

In order to use Deep-CNN to estimate daily PM2.5 concentrations over the CONUS from March to May 2019 and 2020, we have incorporated several predictor variables, including the weighted average PM2.5 layer (WAPM), the weighted average NO2 layer (WANO2), model data from the Community Multiscale Air Quality (CMAQ) modeling system, meteorological factors, reanalysis data, land-use parameters, and the vegetation index. All the predictor variables are prepared over the CONUS during March–May 2019 and 2020. To train the Deep-CNN model, we use surface PM2.5 observations as the target variable, including 63,303 samples in 2019 and 62,181 in 2020. More information about the datasets used in this study is available in the supplementary document.

Methodology

Deep convolutional neural network (Deep-CNN)

To estimate surface PM2.5 concentrations over the CONUS from March to May 2019 and 2020, we trained one Deep-CNN model for each year. Both models consisted of seven layers, including one input layer, two convolutional layers, three fully-connected layers, and an output layer. We also added a dropout layer between the two convolutional layers to help reduce the overfitting issue (Ghahremanloo et al., 2021c). After feature engineering and finding the best combination of parameters, we retained only eleven predictor variables — WAPM, WANO2, CMAQ PM2.5, CMAQ HCHO, MERRA SO2, MERRA BC, RD, PUS, PD, surface elevation, and SLH ⸺ for the PM2.5 estimation.

Feature selection

Several studies have addressed the negative impact of multicollinearity on model performance (Kroll and Song, 2013; Wei et al., 2019). Therefore, it is crucial to properly remove predictor variables that are highly correlated with others. To find and exclude predictor variables with a high degree of multicollinearity, this study applied the variance inflation factor (VIF) (Ghahremanloo et al., 2021c; Kline, 2015; Wei et al., 2019). Previous studies (Cenfetelli and Bassellier, 2009; Ghahremanloo et al., 2021c; Kline, 2015; Kock and Lynn, 2012) have suggested various values as a VIF threshold, indicating multicollinearity among predictor variables. Following Kline (2015), to apply the VIF test on predictor variables, we set a VIF threshold equal to 5. Table S1 shows the results of the VIF test, which excluded two predictor variables (air temperature and surface pressure) from the data. After excluding parameters with large multicollinearity, we leveraged SHapley Additive exPlanations (SHAP) (Lundberg and Lee, 2017; SHAP, 2019) to find the best combination of predictor variables for surface PM2.5 estimation. To this end, we removed less important parameters based on the SHAP feature importance scores of all the chosen parameters in the pre-trained models (Ghahremanloo et al., 2021c; Liu et al., 2020). SHAP calculates the importance of a feature by estimating PM2.5 levels with and without the feature (Lundberg and Lee, 2017) and also leads to the interpretation of the output of black-box models such as the Deep-CNN (Ghahremanloo et al., 2021c; García and Aznarte, 2020). Among the available SHAP-based explanation methods, the Deep SHAP is used in this study because it performs better with DL models (Lundberg and Lee, 2017; SHAP, 2019).

Model evaluation

To validate the accuracy of the Deep-CNN at surface PM2.5 estimation, we use the ten-fold cross-validation (10-CV) approach, which splits samples into ten non-overlapping groups, trains the model with nine groups, and tests the accuracy of PM2.5 estimation with the remaining group (Ghahremanloo et al., 2021b, 2021c). This process is repeated until all ten groups are selected as test data. To evaluate the accuracy of the Deep-CNN, we use the correlation coefficient (R), the index of agreement (IOA), the mean absolute bias (MAB), and the root mean square error (RMSE) at every cycle, and the 10-cycle mean of each metric shows the final performance of the model (Ghahremanloo et al., 2021c). We also used spatial cross-validation (spatial-CV) to evaluate the spatial accuracy of the Deep-CNN, especially in areas with a small number of monitoring stations. The only difference between 10-CV and spatial-CV is that spatial-CV splits samples according to the ground stations while 10-CV splits samples randomly over all stations.

Google Community Mobility Reports

Google Community Mobility Reports (https://www.google.com/covid19/mobility/) release information about temporal mobility trends across regions for various types of places such as grocery stores and pharmacies, parks, transit stations, workplaces, and residential areas. The baseline day for each region is the median value from the five-week period from January 3 to February 6, 2020, in that region. It should be noted that the baseline is not a single value but consist of seven individual baseline values for each day in each category. Therefore, the same number of visitors on two different days could result in various percentage changes. The category “Parks” refers to places such as public gardens, castles, national forests, campgrounds, and observations decks, and the category “Transit Stations” represents places such as subway stations, seaports, taxi stands, highway rest stops, and car rental agencies. Although all categories measure changes in the total number of visitors, the “Residential” category shows only changes in duration. According to Google, “These reports are created with aggregated, anonymized sets of data from users who have turned on the location history setting, which is off by default.” It should be noted that the Google Community Mobility Reports are not based on quality-assured data sources. Therefore, these data should be used with caution.

Results and discussion

Validation of the resampled CMAQ and NLDAS datasets

Figs. S1 and S2 represent the validation results of the original and resampled CMAQ datasets in 2019 and 2020, respectively, showing moderate accuracy of CMAQ at estimating daily surface concentrations of PM2.5 and HCHO over the CONUS. Figs. S3–S6 also display validation results of NLDAS meteorological factors against surface observations in 2019 and 2020. Results show the very high capability of NLDAS at modeling various meteorological factors, including surface pressure, air temperature, specific humidity, and LRad over the study area. Previous studies (Ghahremanloo et al., 2021c; Luo et al., 2003) have also confirmed the high accuracy of NLDAS meteorological factors over the CONUS. The accuracy of NLDAS wind speed (R-2019 = 0.58 and R-2020 = 0.54) and SRad (R-2019 = 0.69 and R-2020 = 0.81), however, was lower than that of the other variables. The figures also reveal that cubic convolution interpolation, used to resample CMAQ and NLDAS datasets, slightly improved the accuracy of several parameters, especially with regard to the MAB. For instance, this approach decreased the MAB of surface pressure in 2019 from 3.69 mbars (mb) to 3.46 mb. The resampling approach also improved the accuracy of the other variables ⸺ LRad, SRad, air temperature, and CMAQ PM2.5 ⸺ in 2019 and 2020.

Table 2 shows the validation results of the Deep-CNN at estimating daily concentrations of surface PM2.5 from March to May 2019 and 2020 over the CONUS. With an R, IOA, MAB, and RMSE of 0.90 (0.90), 0.95 (0.95), 1.34 (1.24) μg/m3, and 2.04 (1.87) μg/m3, respectively, the Deep-CNN shows promising performance at surface PM2.5 estimation in 2019 (2020). A number of studies have estimated surface PM2.5 levels in the CONUS. Using a two-stage spatial statistical model, Hu et al. (2014) obtained an R of 0.82 in an estimation of surface PM2.5 in the southeastern United States. Lee et al. (2016) also leveraged a mixed-effects model to achieve a relatively high accuracy (R = 0.74–0.85) at estimating PM2.5 levels in California from 2006 to 2012. Ghahremanloo et al. (2021b) used RF to estimate surface PM2.5 levels over Texas from 2014 to 2018 and found an R ranging from 0.83 to 0.90 and an MAB ranging from 1.47 to 1.77 μg/m3. The results of spatial-CV, listed in Table 2, also show the spatial performance of the Deep-CNN at estimating surface PM2.5 levels in 2019 (2020) with an R, IOA, MAB, and RMSE of 0.83 (0.84), 0.91 (0.91), 1.73 (1.66) μg/m3, and 2.63 (2.47) μg/m3, respectively. Since the model is validated against stations it has not been trained with, however, the spatial-CV accuracy of Deep-CNN is lower than that determined by 10-CV. This finding is in agreement with the findings of previous studies (Ghahremanloo et al., 2021c; Li et al., 2017; Park et al., 2020). In Fig. 2 , estimated PM2.5 is compared to corresponding PM2.5 observations, further highlighting the strong capacity of Deep-CNN to estimate surface PM2.5 concentrations.

Table 2

Results of the ten-fold cross-validation (10-CV) and spatial cross-validation (spatial-CV), which show the performance of two Deep-CNN models at estimating surface PM2.5 levels in 2019 and 2020. The evaluation metrics include the correlation coefficient (R), the index of agreement (IOA), the mean absolute bias (MAB), and the root mean square error (RMSE). The MAB and RMSE are in μg/m3, and % refers to the error (i.e., MAB or RMSE) value divided by mean observed PM2.5.

Year	10-CV				Spatial-CV
	R	IOA	MAB (%)	RMSE (%)	R	IOA	MAB (%)	RMSE (%)
2019	0.90	0.95	1.34 (19)	2.04 (29)	0.83	0.91	1.73 (25)	2.63 (37)
2020	0.90	0.95	1.24 (19)	1.87 (28)	0.84	0.91	1.66 (25)	2.47 (37)

Fig. 2

Scatterplots of the ten-fold cross-validation (10-CV) results showing the performance of the deep convolutional neural network (Deep-CNN) at estimating surface concentrations of PM2.5 over the CONUS from March to May 2019 (a) and 2020 (b). The R, IOA, and the MAB refer to the correlation coefficient, index of agreement, and the mean absolute bias, respectively.

Fig. 3 displays the spatial distribution of the estimated PM2.5 concentrations at a 5-km spatial resolution over the CONUS from March to May 2019 and 2020. The estimated PM2.5 levels are higher in the eastern CONUS than in the western regions, particularly over urban environments such as Phoenix, Houston, and Los Angeles. This finding is consistent with the findings of previous studies (Di et al., 2016; Hu et al., 2017; Park et al., 2020) that estimated PM2.5 over the CONUS. Moreover, Fig. 3 shows that the Deep-CNN successfully captures PM2.5 concentrations on most highways over the CONUS. Fig. 3 also plots the mean PM2.5 concentrations at ground stations over the estimated PM2.5 maps from March to May 2019 and 2020 to show the performance of the Deep-CNN in PM2.5 estimation. Results reveal that estimated and observed PM2.5 concentrations exhibit a similar pattern over the study area, indicating the promising capability and accuracy of the Deep-CNN at estimating the spatiotemporal distribution of PM2.5 levels over the CONUS.

Fig. 3

The spatial distribution of mean estimated PM2.5 levels over the CONUS from March to May 2019 (top) and 2020 (bottom). The circles show the mean observed PM2.5 concentrations from the EPA stations during the study period, and the background maps represent the estimated PM2.5 levels using deep learning (DL).

Fig. 3 reveals relatively high PM2.5 concentrations in southern Texas in 2019 and 2020. In 1994, the governments of the United States, Mexico, and Canada signed the North American Free Trade Agreement (NAFTA) (Text - H.R.3450 - 103rd Congress) to create a free trade zone. Since that time, NAFTA has fostered the growth and accumulation of industrial regions and rapid population growth in regions close to the U.S.-Mexico border (Karnae and John, 2019). In light of these changes, the U.S. EPA predicted that air pollution would be a significant problem in regions close to the border (Karnae and John, 2019). Chow et al. (2000) also investigated the cross-boundary transport of air pollution across the U.S.-Mexico border, concluding that the pollution levels transported from Mexico to the U.S. were three times as high as those flowing in the opposite direction. Several studies have investigated the impact of long- and short-range transport on air pollution levels in various regions of the world (Jaffe et al., 2004; Liang et al., 2004; Pouyaei et al., 2020). There are various sources impacting PM2.5 levels in southern Texas. Karnae and John (2019) used positive matrix factorization (PMF) to investigate the percentage contributions of various aerosols influencing PM2.5 levels in southern Texas. According to their study, secondary sulfate, mainly from southern, southwestern, and southeastern regions, is the dominant aerosol contributing to PM2.5 concentrations in this region. Another major contributor is sea salt sulfates from the Gulf of Mexico (southeast) and the hypersaline Laguna Madre (south-southwest) in Mexico (Karnae and John, 2019). The transport of pollutants to coastal regions increases acid formation (Karnae and John, 2019). Although sea salt is mainly composed of sodium chloride (NaCl), elevated acidity results in dechlorination or dehalogenation, in which the Cl− in sea salt is replaced by sulfate and nitrate, producing sea salt sulfate (Hasheminassab et al., 2014). The third major contributor is fresh marine aerosols from the Gulf of Mexico, elevating PM2.5 levels in southern Texas (Karnae and John, 2019).

Fig. S7 presents the importance of the features used in PM2.5 estimation and describes the influence of each predictor variable on the model output (i.e., estimated surface PM2.5 concentrations). The most important variable in the model is the WAPM. This finding is in agreement with the findings of previous studies (Hu et al., 2017; Park et al., 2020). The remaining important features are the RD, PUS, PD, surface elevation, SLH, CMAQ PM2.5, CMAQ HCHO, WANO2, MERRA SO2, and MERRA BC. In Fig. S7a in the supplement, red designates the highest value of the feature while blue represents the lowest. The high (low) SHAP value of a feature means that the feature increases (decreases) the output (i.e., estimated PM2.5 levels) value (Ghahremanloo et al., 2021c). According to Fig. S7a, the WAPM, RD, PUS, PD, CMAQ PM2.5, CMAQ HCHO, WANO2, MERRA SO2, and MERRA BC positively impact the output (i.e., estimated PM2.5 levels) of the Deep-CNN model, although a degree of mixed impact (i.e., positive and negative) occurs in some features such as the PUS. The RD, PUS, and PD refer to vehicles and human activity, major sources of pollutants such as NO2 and SO2 (Ghahremanloo et al., 2021a), both of which are the precursors of nitrate and sulfate, respectively, two of the main ingredients of PM2.5 (Ghahremanloo et al., 2021b). The only variable negatively impacting the model output is surface elevation. Generally, less human activity takes place at higher altitudes with stronger winds, providing conditions favorable to pollution dispersion (Ghahremanloo et al., 2021b; Hu et al., 2014). As all three gases — HCHO, NO2, and SO2 — are precursors of PM2.5 (Huang et al., 2022), they positively impact estimates of PM2.5 concentrations. As BC is another main component of PM2.5 (Targino et al., 2016; Viidanoja et al., 2002), higher BC usually results in elevated PM2.5 levels.

Reduction in mobility and its impact on PM2.5 levels

Table 3 represents the percentage changes in human mobility to the various categories of places, including grocery stores and pharmacies, workplaces, residential areas, transit stations, and parks in the eleven regions from March to May 2020, compared to the baseline (median value from January 3 to February 6, 2020). This table also lists the percentage changes in the mean estimated PM2.5 concentrations over the eleven urban environments in March–May 2020, compared to similar days in 2019. Although the baseline for the mobility changes is not March to May 2019, the mobility change data can still represent the amount of changes in human activity and contribute to the interpretation of PM2.5 changes. According to Table 3, three urban environments (Washington DC, New York, and Boston) experienced the highest reductions in mean PM2.5 levels (i.e., −21.1%, −20.7%, and −18.5%, respectively) in 2020, compared to those in 2019. Accordingly, these regions had the highest percentage reductions in the number of visitors to grocery stores-pharmacies, workplaces, and transit stations. The highest percentage increases in the amount of time people stayed in their residential areas also occurred in Washington DC (18.44%), New York (20.34%), and Boston (20.43%), further highlighting the impact of mobility changes and stay-at-home policies on PM2.5 changes.

Table 3

Percentage changes in human mobility to various types of places, including grocery-pharmacy, workplaces, transit stations, and parks in the eleven study regions between March–May of 2020 and the baseline (median value from the five-week period ranging from January 3 to February 6, 2020). The column “Residential” refers to percentage changes in the amount of time people stayed in their residential areas compared to the baseline. The column “PM2.5 Change” also shows the percentage changes in the mean PM2.5 levels between March–May of 2020 and 2019. All numbers represent percentages.

Region	PM2.5 Change	Grocery-Pharmacy	Workplace	Residential	Transit Stations	Parks
Washington DC	−21.1	−14.40	−43.69	18.44	−47.28	16.53
New York	−20.7	−14.57	−47.19	20.34	−53.35	−3.64
Boston	−18.5	−16.25	−47.63	20.43	−57.38	13.14
Detroit	−13.53	−11.35	−44.72	17.26	−32.22	63.68
Chicago	−8.05	−3.51	−37.21	15.09	−31.79	19.32
Seattle	−7.73	−11.33	−45.46	17.86	−44.98	26.78
Dallas	−6.71	−4.36	−36.48	14.64	−24.67	14
Philadelphia	−4.82	−11.09	−39.75	15.76	−44.45	11.94
Houston	−3.63	−1.49	−34.37	14.26	−13.09	9.17
Los Angeles	−3.29	−10.39	−37.33	17.14	−41.66	−23.56
Phoenix	5.5	−7.00	−34.13	12.32	−31.16	−11.15

In addition, Table 3 displays a 5.5% increase in the mean PM2.5 concentrations in Phoenix between March–May of 2020 and 2019. In fact, Phoenix, compared to other regions, also experienced the lowest percentage reduction in the number of individuals who went to their workplaces (−34.13%) and the lowest percentage increase (12.32%) in the amount of time people stayed in residential areas. Phoenix also had one of the lowest percentage reductions in the number of visitors to other places, including grocery stores and pharmacies (−7%) and transit stations (−31.16%). These findings indicate that human activity in Phoenix did not decline as much as it did in other urban environments. The correlation coefficient between PM2.5 percentage changes in all regions and corresponding percentage changes in the “Grocery-Pharmacy”, “Workplace”, “Residential”, “Transit Stations”, and “Parks” categories were 0.67, 0.83, −0.87, 0.61, and −0.34, respectively. These findings indicate that the longer people stayed in their residential areas and the more they avoided visiting grocery stores, pharmacies, workplaces, and transit stations, the more significant the PM2.5 reduction was in the area; however, the relatively small negative correlation between the percentage changes in PM2.5 and the percentage changes in the “Parks” category appears to be inconsistent, calling for further research in this regard. Fig. S8 shows a decrease followed by an increase in the percentage change of human mobility to parks in almost all regions. The increase in park visitation can be attributed to the people's desire to visit parks and partake in other recreational activities when lockdown mandates were gradually being eased. Studies have found evidence of both the negative impact of lockdowns on mental health across several social groups and globally (Adams-Prassl et al., 2020; Banks and Xu, 2020; Elmer et al., 2020; Rossi et al., 2020) and the positive impact of access to public green spaces and parks on mental wellbeing and physical activity (Bedimo-Rung et al., 2005; Wood et al., 2017).

Another possible factor for the increase in the number of people visiting parks is the seasonal transition from winter to spring. The negative correlation between percentage changes in PM2.5 levels in all regions and the corresponding percentage changes in the “Parks” category indicates that as more people avoided visiting parks, PM2.5 levels tended to increase. As the Google mobility report does not specify the mode of transport (e.g., vehicular or walking) people used to visit parks, studies could analyze the mode of transportation that people tended to use to visit green and recreational spaces as the lockdowns were gradually eased.

Fig. S8 also displays the daily percentage changes in human mobility to the various types of places in the eleven regions between March–June of 2020 and the baseline days. According to the figure, human mobility in all regions began to decrease during the first days of March 2020, with the highest percentage reductions occurring in the number of visitors to workplaces (a 40.7% mean reduction) and transit stations (a 38.4% mean reduction). Visits to grocery stores/pharmacies peaked before the sudden decrease during the first half of March in all regions, which can be attributed to people flocking to grocery stores and pharmacies in need of items.

The relationship between PM2.5 percentage changes in March–May 2020 compared to 2019 and vehicular PM2.5 (i.e., the ratio of vehicular PM2.5 to other sources of PM2.5) in the eleven urban environments over the CONUS is shown in Fig. 4 . As mentioned in Section 2.2.8, we were using the 2017 NEI to calculate vehicular PM2.5. The figure shows a significant correlation (R = −0.77) between PM2.5 percentage changes and vehicular PM2.5, indicating that regions with more vehicular PM2.5 experienced greater PM2.5 reductions in 2020 compared to 2019. For instance, PM2.5 levels in Washington DC decreased 21.1% in March–May 2020 compared to 2019, and 8.82% of the total PM2.5 emissions in this region were attributed to vehicle use; however, in Phoenix, with a 5.5% increase in PM2.5 levels, only 4.15% of PM2.5 emissions were from vehicles. Fig. 4 also shows the upper and lower limits of the 95% confidence interval (CI) for the linear regression line of the population. Since we use a small sample size (i.e., eleven urban environments) from the population, the 95% CI shows the area in which, with 95% probability, the true linear regression line of the total population lies. The CI shows the uncertainty of the relationship between PM2.5 percentage changes in March–May 2020 compared to 2019 and vehicular PM2.5. However, a small significance F (almost 0.0058) of the linear regression in Fig. 4 indicates that the linear relationship between PM2.5 percentage changes and vehicular PM2.5 is significant.

Fig. 4

Scatterplot showing the relationship between percentage changes of PM2.5 between March–May of 2020 and 2019 and vehicular PM2.5 (i.e., the ratio of vehicular PM2.5 to other sources of PM2.5) in the eleven urban environments over the CONUS, including Washington DC (DC), New York (NY), Boston (BO), Chicago (CH), Los Angeles (LA), Houston (HO), Philadelphia (PHI), Detroit (DE), Phoenix (PH), Dallas (DA), and Seattle (SE). The Upper CI and Lower CI refer to the upper and lower limits of the 95% confidence interval, respectively.

Impact of meteorology and precursors on changes in PM2.5

Table 4 shows the relationship between daily PM2.5 changes and daily variations in several parameters, including air temperature, specific humidity, surface pressure, PBLH, and surface concentrations of dust, sea salt, BC, OC, SO2, SO4, and NO2 between March–May of 2020 and 2019, in the eleven urban environments over the CONUS. The table lists the correlation coefficients of all parameters with PM2.5 in 2019 and 2020 (henceforth referred to as R2019-PM2.5 and R2020-PM2.5, respectively) along with their absolute and percentage changes occurring between March–May of 2020 and 2019.

Table 4

Percentage (%) and absolute (Diff.) changes in parameters, including air temperature (T), specific humidity (Hum), surface pressure (Pres), the planetary boundary layer height (PBLH), and surface concentrations of dust, sea salt (Salt), black carbon (BC), organic carbon (OC), SO2, SO4, and NO2 in the eleven urban environments over the CONUS between March–May of 2020 and 2019. The table also shows the correlation coefficient of all parameters with PM2.5 in 2019 (R2019) and 2020 (R2020). The units for the parameters are as follows: T (k), Hum (10−3 kg/kg), Pres (millibars), PBL (m), Dust (10−11 kg/m3), Salt (10−11 kg/m3), BC (10−11 kg/m3), OC (10−11 kg/m3), SO2 (10−11 kg/m3), SO4 (10−11 kg/m3), NO2 (ppb), and PM2.5 (μg/m3).

		T	Hum	Pres	PBLH	Dust	Salt	BC	OC	SO2	SO4	NO2	PM2.5
Washington DC	R2019	0.07	0.16	0.20	−0.36	0.10	−0.16	0.45	0.37	0.38	0.31	0.42	–
	R2020	0.13	0.12	0.21	−0.16	0.25	−0.14	0.50	0.32	0.49	0.00	0.63	–
	Diff.	−1.56	−1	−9.31	64.00	20.9	36	−8.53	−64.5	−4.51	−0.57	2.42	−1.37
	%	−0.55	−14.7	−0.01	8.05	14.62	22.22	−8.81	−17.4	−7.70	−0.17	−23.6	−21.1
New York	R2019	−0.08	0.06	0.18	−0.37	0.02	−0.01	0.57	0.40	0.53	0.32	0.64	–
	R2020	0.13	0.17	0.12	−0.45	−0.04	0.00	0.65	0.68	0.51	0.11	0.70	–
	Diff.	−0.65	−0.7	−63.6	87.9	22.4	39.3	−8.3	−45.9	−52.9	−29.4	−3.62	−1.52
	%	−0.23	−10.2	−0.06	12.53	17.54	17.87	−7.69	−13.7	−5.36	−7.94	−26.6	−20.7
Boston	R2019	−0.35	−0.25	0.41	−0.19	−0.07	−0.14	0.64	0.36	0.71	0.09	0.70	–
	R2020	0.06	0.05	0.18	−0.37	−0.05	0.04	0.59	0.54	0.67	0.23	0.51	–
	Diff.	0.17	−0.4	−89	50.3	22.4	104	−4.50	−6.69	13.5	−23.2	−2.57	−1.18
	%	0.06	−7.33	−0.09	7.04	21.80	42.22	−6.24	−2.82	3.37	−8.34	−28.1	−18.5
Los Angeles	R2019	0.75	0.14	0.00	−0.26	0.27	0.01	0.71	0.75	0.59	0.29	0.44	–
	R2020	0.81	0.11	−0.13	−0.61	0.44	0.42	0.78	0.62	0.62	0.12	0.39	–
	Diff.	0.84	−0.1	42.2	−4.27	32.8	26.2	2.4	22.3	−13.7	−29.8	−1.54	−0.22
	%	0.29	−1.99	0.04	−0.61	19	16.95	3.76	9.70	−1.69	−11.4	−15.6	−3.29
Chicago	R2019	−0.14	−0.15	0.32	−0.22	0.17	−0.46	0.54	0.29	0.62	0.31	0.72	–
	R2020	0.02	0.03	−0.03	−0.38	−0.07	−0.35	0.63	0.45	0.56	0.48	0.60	–
	Diff.	0.24	0	20.70	44.60	6.22	17.3	−5.20	−66.9	−66.5	3.22	−2.37	−0.66
	%	0.09	−0.12	0.02	6.95	3.56	15.06	−5.94	−19.3	−8.42	0.87	−17.4	−8.05
Houston	R2019	0.27	0.30	−0.16	−0.02	0.11	0.47	0.20	0.24	0.18	0.09	0.00	–
	R2020	0.33	0.36	−0.32	−0.13	−0.28	0.30	0.30	0.31	−0.01	−0.10	0.10	–
	Diff.	1.73	0.2	−24.2	15.15	12.4	−1.51	2.30	23.6	−19.6	−13.5	−1.36	−0.34
	%	0.59	1.79	−0.02	2.12	8.31	−25.2	7.21	9.06	−2.95	−6.31	−18.7	−3.63
Philadelphia	R2019	−0.09	0.02	0.22	−0.23	0.02	−0.19	0.57	0.30	0.50	0.34	0.51	–
	R2020	0.13	0.13	0.10	−0.22	0.04	−0.05	0.47	0.57	0.44	0.07	0.66	–
	Diff.	−1.24	−0.9	−37.5	67.5	26.8	35	−14.8	−157	−70.8	−3.54	1.67	−0.33
	%	−0.44	−12.6	−0.04	8.48	20.1	21.23	−9.94	−31	−8.59	−1.04	−16	−4.82
Detroit	R2019	−0.13	−0.10	0.34	−0.21	0.09	−0.38	0.59	0.24	0.62	0.18	0.74	–
	R2020	0.02	0.05	0.06	−0.52	0.01	−0.29	0.75	0.59	0.65	0.47	0.69	–
	Diff.	−0.06	−0.1	12.7	60.2	−17.9	15	−4.30	−71.7	−19.4	−40.9	−3.1	−1.16
	%	−0.02	−2.4	0.01	8.87	−9.17	13.11	−6.48	−24	−2.39	−10.3	−24.1	−13.5
Phoenix	R2019	0.34	−0.04	0.23	−0.37	−0.11	−0.38	0.65	0.32	0.73	0.37	0.59	–
	R2020	0.81	−0.36	−0.25	0.15	0.28	0.07	0.38	0.70	0.00	0.09	0.39	–
	Diff.	1.06	0.1	103	−19.2	−29.7	24.8	5.88	22.6	7.57	14.2	1.73	0.28
	%	0.36	2.17	0.11	−1.50	−4.28	33.43	10.52	13.06	3.72	11.39	−13.2	5.50
Dallas	R2019	0.00	−0.02	0.22	−0.14	−0.11	0.18	0.20	0.17	0.48	0.36	0.38	–
	R2020	0.27	0.24	−0.21	−0.13	−0.26	0.22	0.25	0.27	0.37	0.10	0.19	–
	Diff.	1.42	0	4.35	−20.2	20.1	−37.2	−0.7	−40.8	−6.64	−6.47	−0.68	−0.54
	%	0.49	0.75	0.00	−2.76	11.03	−20	−1.48	−11.5	−1.82	−2.56	−11.6	−6.71
Seattle	R2019	0.17	0.13	0.17	−0.18	0.25	−0.29	0.35	0.47	0.56	−0.17	0.56	–
	R2020	−0.11	−0.24	0.16	−0.41	0.23	−0.22	0.52	0.52	0.81	−0.07	0.69	–
	Diff.	−1.05	−0.4	81.1	77.2	3.07	8.09	−3.44	−39.2	−9.68	−1.65	3.42	−0.42
	%	−0.37	−6.9	0.08	12.10	3.11	5.99	−5.15	−16	−3.97	−1.08	−24.3	−7.73

The results in Table 4 show that Washington DC experienced the highest reduction in PM2.5 (−21.1%) in March–May of 2020 compared to 2019. The table shows no major changes in meteorological factors in Washington DC between 2020 and 2019, except for air temperature (R2020-PM2.5 = 0.13), which decreased by 1.56 K in 2020. There is usually a positive correlation between PM2.5 and air temperature since an increase in temperature can promote the conversion of NO2 and SO2 to nitrate and sulfate (Lin et al., 2019; Ghahremanloo et al., 2021b), respectively, increasing PM2.5 levels. Increased temperature can also result in more biogenic emissions, which, in turn, leads to more PM2.5 concentrations (Di et al., 2016; Ghahremanloo et al., 2021b). It should be noted that elevated temperature can also promote air convection resulting in dispersion situations that reduce pollution levels (Luo et al., 2017; Pouyaei et al., 2021; Yang et al., 2017). The relationship between PM2.5 and temperature (R2020-PM2.5 = 0.13) in Washington DC, however, was weak, indicating that the temperature was unlikely to have impacted PM2.5 levels in this region in 2020. PM2.5 is more strongly associated with other parameters, particularly surface NO2 concentrations (R2020-PM2.5 = 0.63), with a 2.42 ppb (−23.64%) decrease in 2020 compared to 2019 in Washington DC. PM2.5 concentrations also show a moderate correlation with BC (R2020-PM2.5 = 0.50), OC (R2020-PM2.5 = 0.32), and SO2 (R2020-PM2.5 = 0.49), with a decrease of 8.81%, 17.40%, and 7.7%, respectively, in 2020 compared to 2019. In fact, as BC, OC, and SO4 make up a portion of PM2.5, any changes they undergo directly influence PM2.5 concentrations (Bell et al., 2007; Ghahremanloo et al., 2021b). SO2 and NO2 are also precursors of sulfate and nitrate, which comprise a large portion of PM2.5. Therefore, as vehicular emissions are main source of NO2, SO2, SO4, OC, and BC concentrations in the atmosphere (Cao et al., 2006; Ghahremanloo et al., 2021b; Ni et al., 2014; Yang et al., 2018), a decrease in vehicular use and emissions, especially in Washington DC, with relatively high levels of vehicular PM2.5, can lead to reduced PM2.5 levels, as shown in Section 4.3.

In New York, PM2.5 concentrations decreased by 20.7% between March–May of 2020 and 2019. Table 3 shows that one of the greatest reductions in human mobility to various categories of places occurred in New York. Table 4 shows that PM2.5 levels in New York are strongly associated with NO2 concentrations (R2020-PM2.5 = 0.70), which decreased by 26.6% in 2020 compared to 2019, indicating that the decrease in NO2 was main reason for the decrease in PM2.5 levels in this region. In addition, PM2.5 reduction in New York could have partly been due to small decreases in concentrations of BC (R2020-PM2.5 = 0.65), OC (R2020-PM2.5 = 0.68), and SO2 (R2020-PM2.5 = 0.51) resulting from the relatively strong correlations between these pollutants and PM2.5. In addition, New York saw an increase of 87.9 m (12.53%) in the PBLH (R2020-PM2.5 = −0.45) during March–May 2020 compared to 2019. Since pollutants are distributed throughout a higher volume of air, an elevated PBLH can reduce air pollution levels (Su et al., 2018). Thus, the PBLH could have been partially responsible for the PM2.5 reduction in New York. A 7.04% increase in the PBLH (R2020-PM2.5 = −0.37) may also have been partially responsible for the PM2.5 reduction (−18.5%) in Boston in 2020. Most of the reduction in PM2.5, however, can be attributed to decreased concentrations of BC (−6.24%; R2020-PM2.5 = 0.59), OC (−2.82%; R2020-PM2.5 = 0.54), sulfate (−8.34%; R2020-PM2.5 = 0.23), and particularly NO2 (−28.1%; R2020-PM2.5 = 0.51). It should be noted that Washington DC, New York, and Boston had the highest percentage decreases in the number of visitors to various categories of places and the highest percentage increases in the amount of time people stayed in their residential areas during March–May 2020 compared to baseline, resulting in large reductions in PM2.5 levels. Compared to other regions in this study, these urban environments also had relatively high vehicular PM2.5.

In Los Angeles, unlike in Washington DC, New York, and Boston, PM2.5 levels only slightly declined (−3.29%) during March–May 2020 compared to 2019. Table 4 shows a 0.84 K increase in the temperature in Los Angeles between March–May of 2020 and 2019. The strong relationship between PM2.5 and temperature in 2019 and 2020 (R2019-PM2.5 = 0.75 and R2020-PM2.5 = 0.81, respectively) indicates that temperature strongly impacted the PM2.5 levels in Los Angeles, although temperature increase is not significant in 2020. Dust (R2020-PM2.5 = 0.44), sea salt (R2020-PM2.5 = 0.42), BC (R2020-PM2.5 = 0.78), and OC (R2020-PM2.5 = 0.62) also increased by 19%, 16.95%, 3.76%, and 9.70%, respectively, while SO4 (R2020-PM2.5 = 0.12) and NO2 (R2020-PM2.5 = 0.39) decreased by 11.4% and 15.6%, respectively, in Los Angeles between 2020 and 2019. Because of their moderate to high correlation with PM2.5 changes, the increases in the levels of temperature, dust, sea salt, BC, and OC could, to some extent, have been responsible for the slight reduction of PM2.5 in Los Angeles. Results show that in Chicago, BC (R2020-PM2.5 = 0.63), OC (R2020-PM2.5 = 0.45), SO2 (R2020-PM2.5 = 0.56), and NO2 (R2020-PM2.5 = 0.60) decreased by 5.94%, 19.3%, 8.42%, and 17.4%, respectively, between 2020 and 2019, reducing PM2.5 levels by 8.05% between the two years.

Table 4 shows only 3.63% decrease in PM2.5 levels in Houston during the study period in 2020 compared to 2019. A 1.73 K increase in temperature (R2020-PM2.5 = 0.33) between 2020 and 2019 could have contributed to the slight increase in PM2.5 levels in Houston. Despite the increase in the levels of BC (7.21%; R2020-PM2.5 = 0.30) and OC (9.6%; R2020-PM2.5 = 0.31) in Houston, SO2, SO4, and NO2 decreased by 2.95%, 6.31%, and 18.7%, respectively, during the study period. The results, however, show no relationship between PM2.5 levels and concentrations of SO2 (R2020-PM2.5 = −0.01), SO4 (R2020-PM2.5 = −0.1), and NO2 (R2020-PM2.5 = −0.1) in Houston in 2020. Accordingly, Fig. 4 and Table 3 also show relatively low vehicular PM2.5 in Houston, compared to other urban environments, along with relatively small changes in the number of visitors to various categories of places. Houston experienced the lowest percentage reduction in the number of people visiting grocery stores and pharmacies (−1.49%) (Table 3) and one of the lowest percentage reductions in other categories, explaining the relatively small reduction in PM2.5 levels in March–May 2020 compared to 2019. Moreover, the increase in the amount of time people stayed at their residential areas in Houston was relatively small compared to other regions. Philadelphia also experienced a 4.82% decrease in PM2.5 concentrations between 2020 and 2019, possibly influenced by decreases in BC (−9.94%; R2020-PM2.5 = 0.47), SO2 (8.59%; R2020-PM2.5 = 0.44), and especially OC (−31%; R2020-PM2.5 = 0.57) and NO2 (−16%; R2020-PM2.5 = 0.66). In Detroit, a 13.5% decrease in PM2.5 concentrations between March–May of 2020 and 2019 can also be attributed to decreases in the concentrations of BC (−6.48%; R2020-PM2.5 = 0.75), OC (−24%; R2020-PM2.5 = 0.59), SO2 (−2.39%; R2020-PM2.5 = 0.65), SO4 (−10.3%; R2020-PM2.5 = 0.47), and NO2 (−24.1%; R2020-PM2.5 = 0.69). In addition, an 8.87% increase in the PBLH (R2020-PM2.5 = −0.52) in 2020 may have led to decreases in the PM2.5 levels in Detroit.

Phoenix is the only region in this study that experienced an increase (5.5%) in PM2.5 levels during March–May 2020 compared to similar days in 2019. Table 4 shows that the concentrations of BC (R2020-PM2.5 = 0.38), OC (R2020-PM2.5 = 0.70), SO2 (R2020-PM2.5 = 0.0), and SO4 (R2020-PM2.5 = 0.09) increased by 10.52%, 13.06%, 3.72%, and 11.39%, respectively, between 2020 and 2019. Although the results show no relationship between PM2.5 and concentrations of SO2 and SO4, the correlation of PM2.5 with BC and especially OC in 2020 appears to be strong. Furthermore, the significant correlation between PM2.5 and temperature in 2020, along with a 1.06 K increase in temperature, could have increased PM2.5 concentrations in this region. It should be noted that NO2 levels decreased by 13.2% (R2020-PM2.5 = 0.39) in Phoenix between 2020 and 2019 and that Phoenix had the lowest rate of vehicular PM2.5 emissions (Fig. 4) than the other urban regions in this study. The city also had some of the lowest reductions in human mobility in the “Grocery-Pharmacy” (−7%), “Workplace” (−34.13%, the lowest reduction), and “Transit Stations” (−31.16%) categories. In addition, Phoenix showed the lowest percentage increase in the amount of time people stayed in their residential areas, further highlighting the relatively poor application and enforcement of stay-at-home strategies in this region, which potentially limited the reduction of PM2.5 in Phoenix.

The results showed a decrease of 6.71% in PM2.5 levels in Dallas in March–May 2020 compared to previous year and a decrease of 20%, 1.48%, 11.5%, 1.82%, 2.56%, and 11.6% in the concentrations of sea salt (R2020-PM2.5 = 0.22), BC (R2020-PM2.5 = 0.25), OC (R2020-PM2.5 = 0.27), SO2 (R2020-PM2.5 = 0.37), SO4 (R2020-PM2.5 = 0.10), and NO2 (R2020-PM2.5 = 0.19), respectively, in Dallas between March–May of 2020 and 2019. However, the 1.42 K increase in temperature, with a relatively low correlation with PM2.5 (R2020-PM2.5 = 0.27) in 2020, could have contributed to the slight increase in PM2.5 in this urban area. The results listed in Table 3 also reveal that reductions in mobility in Dallas were not as significant as they were in other areas, and people spent relatively less time in residential areas in Dallas than they did in other regions. In Seattle, PM2.5 levels decreased by 7.73% during March–May 2020 compared to similar days in 2019. Table 4 lists a 5.15%, 16%, 3.97%, 1.08%, and 24.3% reductions in the concentrations of BC (R2020-PM2.5 = 0.52), OC (R2020-PM2.5 = 0.52), SO2 (R2020-PM2.5 = 0.81), SO4 (R2020-PM2.5 = −0.07), and NO2 (R2020-PM2.5 = 0.69) in Seattle between March–May of 2020 and 2019. Decreases in the concentrations of BC, OC, SO2, and NO2, along with their strong relationship to PM2.5, could explain the 7.73% decrease in PM2.5 concentrations in Seattle between March–May 2020 and the previous year. In addition, an increase in the PBLH (R2020-PM2.5 = −0.41) of 12.1% between 2020 and 2019 could have been partially responsible for the reduction of PM2.5 in 2020.

Conclusion

The main purpose of this study was to examine the impact of the COVID-19 outbreak on PM2.5 levels in eleven urban environments (Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, Phoenix, and Seattle) across the United States. We used a Deep-CNN to accurately capture the spatiotemporal distribution of PM2.5 levels at a 5-km spatial resolution over the CONUS in March–May 2019 and 2020. The Deep-CNN model showed promising accuracy at PM2.5 estimation, with an R, IOA, MAE, and RMSE of 0.90 (0.90), 0.95 (0.95), 1.34 (1.24) μg/m3, and 2.04 (1.87) μg/m3, respectively, in March–May 2019 (2020). Results of the Google Community Mobility Reports and the estimated PM2.5 levels revealed a high correlation between changes in human mobility and PM2.5 concentrations, indicating that regions with more reductions in human mobility also experienced more reductions in PM2.5 levels. The correlation coefficient between PM2.5 percentage changes in all regions and corresponding percentage changes in the number of visitors to “grocery stores and pharmacies”, “workplaces”, and “transit stations” were 0.67, 0.83, and 0.61, respectively. Moreover, percentage changes in PM2.5 and the amount of time people stayed in their residential areas during the pandemic were also strongly correlated (R = −0.87), indicating greater reductions in PM2.5 levels in urban environments in which people stayed longer in their residential areas. The relatively strong correlation between vehicular PM2.5 and percentage changes in PM2.5 (R = −0.77) in study regions showed greater PM2.5 reductions in regions with higher vehicular PM2.5 emissions. Among the eleven urban environments, Washington DC, New York, and Boston experienced the greatest reductions in their mean PM2.5 levels (i.e., −21.1%, −20.7%, and −18.5%, respectively) between March–May of 2020 and 2019. These same three urban environments also experienced greater reductions in human mobility and one of the highest vehicular PM2.5 emissions compared to other regions in this study. All of the other urban environments (Los Angeles, Chicago, Houston, Dallas, Philadelphia, Detroit, and Seattle) experienced decreases in PM2.5 levels by 3.29%, 8.05%, 3.63%, 6.71%, 4.82%, 13.5%, and 7.73%, respectively, in 2020 compared to 2019. An exception, however, is Phoenix, with a 5.5% increase in PM2.5 concentrations during the same period. According to the results, changes in the concentrations of pollutants, BC, OC, SO2, SO4, and especially NO2, had a significantly higher impact on changes in PM2.5 concentrations in 2020 than meteorological factors, although there was a strong relationship between air temperature and PM2.5 levels in Los Angeles (R2020-PM2.5 = 0.81) and Phoenix (R2020-PM2.5 = 0.81). The increase in PM2.5 levels (5.5%) in Phoenix could be attributed to increased levels of BC, OC, and air temperature and the relatively low rate of vehicular PM2.5 emissions in this region. In addition, one of the lowest reductions in human mobility to various categories of places occurred in Phoenix, further explaining the increased PM2.5 concentrations in this urban environment.

CRediT authorship contribution statement

Masoud Ghahremanloo: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Yannic Lops: Data curation, Formal analysis, Software, Writing – review & editing. Yunsoo Choi: Conceptualization, Funding acquisition, Methodology, Supervision, Writing – review & editing. Jia Jung: Data curation, Writing – original draft, Writing – review & editing. Seyedali Mousavinezhad: Data curation, Writing – review & editing. Davyda Hammond: Funding acquisition, Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.