key: cord-1006187-1ykmej3b
authors: Ghahremanloo, Masoud; Lops, Yannic; Choi, Yunsoo; Jung, Jia; Mousavinezhad, Seyedali; Hammond, Davyda
title: A comprehensive study of the COVID-19 impact on PM(2.5) levels over the contiguous United States: A deep learning approach
date: 2022-01-14
journal: Atmos Environ (1994)
DOI: 10.1016/j.atmosenv.2022.118944
sha: b0fe36ceb454fa45614c219981ca7aaf7decf1bf
doc_id: 1006187
cord_uid: 1ykmej3b

We investigate the impact of the COVID-19 outbreak on PM(2.5) 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 PM(2.5) 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/m(3), and 2.04 (1.87) μg/m(3), respectively. Results from Google Community Mobility Reports and estimated PM(2.5) concentrations show a greater reduction of PM(2.5) in regions with larger decreases in human mobility and those in which individuals remain in their residential areas longer. The relationship between vehicular PM(2.5) (i.e., the ratio of vehicular PM(2.5) to other sources of PM(2.5)) emissions and PM(2.5) reductions (R = 0.77) in various regions indicates that regions with higher emissions of vehicular PM(2.5) generally experience greater decreases in PM(2.5). While most of the urban environments ⸺ Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, and Seattle ⸺ show a decrease in PM(2.5) 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 PM(2.5) reductions, Washington DC, New York, and Boston, compared to other cities, exhibit the highest reductions in human mobility and the highest vehicular PM(2.5) emissions, highlighting the great impact of human activity on PM(2.5) changes in eleven regions. Moreover, compared to changes in meteorological factors, changes in pollutant concentrations, including those of black carbon, organic carbon, SO(2), SO(4), and especially NO(2), appear to have had a significantly greater impact on PM(2.5) changes during the study period.

We analyze daily changes in PM2.5 concentrations in eleven urban environments over the CONUS 129 ( Figure 1 ) from March 1 to May 31, 2020, compared to similar days in 2019. We have selected eleven 130 urban environments -Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, 131

Philadelphia, Detroit, Phoenix, and Seattlebased on their economic importance, pollution levels, 132 populations, and areas. For the analyses, we have defined eleven boundaries around each urban environment 133 to include all metropolitan areas, including downtown and suburban regions. Table 1 shows the latitude and 134 longitude coordinates of the boundaries around the urban environments analyzed in this study, along with 135 the exact dates of the lockdowns, or stay-at-home orders issued by each city (Wu et al., 2020) . This table 136 also displays the counties inside each region. Of note, the boundaries around three urban environments 137 (Washington DC, New York, and Philadelphia) include more than one state. Therefore, the information of 138

all included states appears in Table 1 . To estimate surface PM2.5 concentrations over the CONUS from March to May 2019 and 2020, we trained 164

one Deep-CNN model for each year. Both models consisted of seven layers, including one input layer, two 165 convolutional layers, three fully-connected layers, and an output layer. We also added a dropout layer 166 between the two convolutional layers to help reduce the overfitting issue (Ghahremanloo et al., 2021c) . variables. Following Kline (2015) , to apply the VIF test on predictor variables, we set a VIF threshold equal 179

to 5. Table S1 shows the results of the VIF test, which excluded two predictor variables (air temperature 180 and surface pressure) from the data. After excluding parameters with large multicollinearity, we leveraged 

To validate the accuracy of the Deep-CNN at surface PM2.5 estimation, we use the ten-fold cross-validation 192 (10-CV) approach, which splits samples into ten non-overlapping groups, trains the model with nine groups, 193 and tests the accuracy of PM2.5 estimation with the remaining group (Ghahremanloo et al., 2021b; 194 Ghahremanloo et al., 2021c) . This process is repeated until all ten groups are selected as test data. To 195 evaluate the accuracy of the Deep-CNN, we use the correlation coefficient (R), the index of agreement 196 (IOA), the mean absolute bias (MAB), and the root mean square error (RMSE) at every cycle, and the 10-197

cycle mean of each metric shows the final performance of the model (Ghahremanloo et al., 2021c) . We also 198 used spatial cross-validation (spatial-CV) to evaluate the spatial accuracy of the Deep-CNN, especially in 199 areas with a small number of monitoring stations. The only difference between 10-CV and spatial-CV is 200 that spatial-CV splits samples according to the ground stations while 10-CV splits samples randomly over 201 all stations. 202 203

Google Community Mobility Reports (https://www.google.com/covid19/mobility/) release information 205 about temporal mobility trends across regions for various types of places such as grocery stores and 206 pharmacies, parks, transit stations, workplaces, and residential areas. The baseline day for each region is 207 the median value from the five-week period from January 3 to February 6, 2020, in that region. It should 208 be noted that the baseline is not a single value but consist of seven individual baseline values for each day 209 in each category. Therefore, the same number of visitors on two different days could result in various 210 percentage changes. The category "Parks" refers to places such as public gardens, castles, national forests, 211 campgrounds, and observations decks, and the category "Transit Stations" represents places such as subway 212 stations, seaports, taxi stands, highway rest stops, and car rental agencies. Although all categories measure 213 changes in the total number of visitors, the "Residential" category shows only changes in duration. 214

According to Google, "These reports are created with aggregated, anonymized sets of data from users who 215 have turned on the location history setting, which is off by default." It should be noted that the Google 216 Community Mobility Reports are not based on quality-assured data sources. Therefore, these data should 217 be used with caution. Table 3 represents the percentage changes in human mobility to the various categories of places, including 327 grocery stores and pharmacies, workplaces, residential areas, transit stations, and parks in the eleven regions 328 from March to May 2020, compared to the baseline (median value from January 3 to February 6, 2020). percentage changes in all regions and corresponding percentage changes in the "Grocery-Pharmacy", 347 "Workplace", "Residential", "Transit Stations", and "Parks" categories were 0.67, 0.83, -0.87, 0.61, and 348 -0.34, respectively. These findings indicate that the longer people stayed in their residential areas and the 349 more they avoided visiting grocery stores, pharmacies, workplaces, and transit stations, the more significant 350 the PM2.5 reduction was in the area; however, the relatively small negative correlation between the 351 percentage changes in PM2.5 and the percentage changes in the "Parks" category appears to be inconsistent, 352

calling for further research in this regard. Figure S8 The results in Table 4 show that Washington DC experienced the highest reduction in PM2. The relationship between PM2.5 and temperature (R2020-PM2.5 = 0.13) in Washington DC, however, was 431 weak, indicating that the temperature was unlikely to have impacted PM2.5 levels in this region in 2020. shows that one of the greatest reductions in human mobility to various categories of places occurred in New 445

York. Table 4 shows that PM2.5 levels in New York are strongly associated with NO2 concentrations 446 (R2020-PM2.5 = 0.70), which decreased by 26.6% in 2020 compared to 2019, indicating that the decrease 447 in NO2 was main reason for the decrease in PM2.5 levels in this region. In addition, PM2.5 reduction in New 448

York could have partly been due to small decreases in concentrations of BC (R2020-PM2.5 = 0.65), OC 449 (R2020-PM2.5 = 0.68), and SO2 (R2020-PM2.5 = 0.51) resulting from the relatively strong correlations 450 between these pollutants and PM2.5. In addition, New York saw an increase of 87.9 m (12.53%) in the PBLH 451 and OC (R2020-PM2.5 = 0.62) also increased by 19%, 16.95%, 3.76%, and 9.70%, respectively, while SO4 469 (R2020-PM2.5 = 0.12) and NO2 (R2020-PM2.5 = 0.39) decreased by 11.4% and 15.6%, respectively, in Los Table 4 shows only 3.63% decrease in PM2.5 levels in Houston during the study period in 2020 compared 477 to 2019. A 1.73 °K increase in temperature (R2020-PM2.5 = 0.33) between 2020 and 2019 could have 478 contributed to the slight increase in PM2.5 levels in Houston. Despite the increase in the levels of BC (7.21%; 479 R2020-PM2.5 = 0.30) and OC (9.6%; R2020-PM2.5 = 0.31) in Houston, SO2, SO4, and NO2 decreased by 480 2.95%, 6.31%, and 18.7%, respectively, during the study period. The results, however, show no relationship 481 between PM2.5 levels and concentrations of SO2 (R2020-PM2.5 = -0.01), SO4 (R2020-PM2.5 = -0.1), and NO2 482 (R2020-PM2.5 = -0.1) in Houston in 2020. Accordingly, Figure 4 and Table 3 Phoenix is the only region in this study that experienced an increase (5.5%) in PM2.5 levels during March-497

May 2020 compared to similar days in 2019. Table 4 shows that the concentrations of BC (R2020-PM2.5 = 498 0.38), OC (R2020-PM2.5 = 0.70), SO2 (R2020-PM2.5 = 0.0), and SO4 (R2020-PM2.5 = 0.09) increased by 499 10.52%, 13.06%, 3.72%, and 11.39%, respectively, between 2020 and 2019. Although the results show no 500 relationship between PM2.5 and concentrations of SO2 and SO4, the correlation of PM2.5 with BC and 501 especially OC in 2020 appears to be strong. Furthermore, the significant correlation between PM2.5 and 502 temperature in 2020, along with a 1.06 °K increase in temperature, could have increased PM2.5 503 concentrations in this region. It should be noted that NO2 levels decreased by 13.2% (R2020-PM2.5 = 0.39) 504

in Phoenix between 2020 and 2019 and that Phoenix had the lowest rate of vehicular PM2.5 emissions 505 ( Figure 4 ) than the other urban regions in this study. The city also had some of the lowest reductions in 506 human mobility in the "Grocery-Pharmacy" (-7%), "Workplace" (-34.13%, the lowest reduction), and 507 "Transit Stations" (-31.16%) categories. In addition, Phoenix showed the lowest percentage increase in the 508 amount of time people stayed in their residential areas, further highlighting the relatively poor application 509 and enforcement of stay-at-home strategies in this region, which potentially limited the reduction of PM2.5 510 in Phoenix. 511

The results showed a decrease of 6.71% in PM2.5 levels in Dallas in March-May 2020 compared to previous 512

year and a decrease of 20%, 1.48%, 11.5%, 1.82%, 2.56%, and 11.6% in the concentrations of sea salt 513 (R2020-PM2.5 = 0.22), BC (R2020-PM2.5 = 0.25), OC (R2020-PM2.5 = 0.27), SO2 (R2020-PM2.5 = 0.37), 514 SO4 (R2020-PM2.5 = 0.10), and NO2 (R2020-PM2.5 = 0.19), respectively, in Dallas between March-May of 515 2020 and 2019. However, the 1.42 °K increase in temperature, with a relatively low correlation with PM2.5 516 (R2020-PM2.5 = 0.27) in 2020, could have contributed to the slight increase in PM2.5 in this urban area. The 517 results listed in Table 3 also reveal that reductions in mobility in Dallas were not as significant as they were 518 in other areas, and people spent relatively less time in residential areas in Dallas than they did in other 519 regions. In Seattle, PM2.5 levels decreased by 7.73% during March-May 2020 compared to similar days in 520 2019. Table 4 

The main purpose of this study was to examine the impact of the COVID-19 outbreak on PM2.5 levels in corresponding percentage changes in the number of visitors to "grocery stores and pharmacies", 540 "workplaces", and "transit stations" were 0.67, 0.83, and 0.61, respectively. Moreover, percentage changes 541 in PM2.5 and the amount of time people stayed in their residential areas during the pandemic were also 542 strongly correlated (R = -0.87), indicating greater reductions in PM2.5 levels in urban environments in which 543 people stayed longer in their residential areas. The relatively strong correlation between vehicular PM2.5 544 and percentage changes in PM2.5 (R = -0.77) in study regions showed greater PM2.5 reductions in regions 545 with higher vehicular PM2.5 emissions. Among the eleven urban environments, Washington DC, New York, 546

and Boston experienced the greatest reductions in their mean PM2.5 levels (i.e., -21.1%, -20.7%, and -18.5%, 547 respectively) between March-May of 2020 and 2019. These same three urban environments also 548 experienced greater reductions in human mobility and one of the highest vehicular PM2.5 emissions 549 compared to other regions in this study. All of the other urban environments (Los Angeles, Chicago, 550

Houston, Dallas, Philadelphia, Detroit, and Seattle) experienced decreases in PM2.5 levels by 3.29%, 8.05%, 551

3.63%, 6.71%, 4.82%, 13.5%, and 7.73%, respectively, in 2020 compared to 2019. An exception, however, 552

is Phoenix, with a 5.5% increase in PM2.5 concentrations during the same period. According to the results, 553 changes in the concentrations of pollutants, BC, OC, SO2, SO4, and especially NO2, had a significantly 554 higher impact on changes in PM2.5 concentrations in 2020 than meteorological factors, although there was 555 a strong relationship between air temperature and PM2.5 levels in Los Angeles (R2020-PM2.5 = 0.81) and 556

Phoenix (R2020-PM2.5 = 0.81). The increase in PM2.5 levels (5.5%) in Phoenix could be attributed to 557 increased levels of BC, OC, and air temperature and the relatively low rate of vehicular PM2.5 emissions in 558 this region. In addition, one of the lowest reductions in human mobility to various categories of places 559 occurred in Phoenix, further explaining the increased PM2.5 concentrations in this urban environment. 

The impact of the coronavirus lockdown on 568 mental health: evidence from the US

Temperature-vegetation-soil 570 moisture dryness index (TVMDI). Remote sensing of environment

Trump Declares Coronavirus A Public Health Emergency And Restricts 572

COVID-19 lockdown effects on air quality by NO2 in the cities of Barcelona and 576

The mental health effects of the first two months of lockdown and social 578 distancing during the Covid-19 pandemic in the UK (No. W20/16)

Does lockdown reduce air pollution? Evidence from 44 cities in northern 580 China

The significance of parks to physical activity 582 and public health: a conceptual model

Spatial and temporal variation in 584 PM2.5 chemical composition in the United States for health effects studies. Environmental health 585 perspectives

Exploring dependence of COVID-587 19 on environmental factors and spread prediction in India

The computation of equivalent potential temperature

Ozone Anomalies in the Free Troposphere During the COVID-19 Pandemic

Review of the Governing Equations, Computational Algorithms, and Other 595 Components of the Models-3 Community Multiscale Air Quality (CMAQ) Modeling System

Inventory of black carbon and organic carbon emissions from 598 China

Features, evaluation, and 600 treatment of coronavirus (COVID-19)

Interpretation of formative measurement in information systems 602 research

Population Density, Revision 11

Decline in PM2.5 concentrations over major cities around the world 607 associated with COVID-19

Differential Impacts of COVID-19 Lockdowns on 609 PM2. 5 across the United States

Cross-border transport and spatial variability of suspended particles in Mexicali and California's Imperial 612 Valley

Predicting fine-scale daily NO2 for 2005-2016 incorporating OMI satellite data across Switzerland. 615 Environmental science & technology

Assessing PM2.5 617 exposures with high spatiotemporal resolution across the continental United States. Environmental science 618 & technology

MOD13C1 MODIS/Terra Vegetation Indices 16-Day L3 Global 0.05Deg CMG V006 620

Processes DAAC

MODIS vegetation index user's guide (MOD13 623 series)

Students under lockdown: Comparisons of students' social 625 networks and mental health before and during the COVID-19 crisis in Switzerland

A real-time hourly ozone prediction system using 628 deep convolutional neural network

Technical Support Document (TSD) preparation of emissions inventories for 630 the version 6.3, 2011 emissions modeling platform

Preparation of 633 Emissions Inventories for the Version 7.1 -2016 North American Emissions Modeling Platform

Shapley additive explanations for NO2 forecasting

COVID-19: air pollution remains low as people stay at home. Air Quality, Atmosphere 637 & Health

Estimating daily 639 high-resolution PM2.5 concentrations over Texas: Machine Learning approach

Impact of the COVID-19 outbreak 642 on air pollution levels in East Asia

Deep learning estimation of daily ground 644 level NO2 concentrations from remote sensing data

Soil moisture estimation using land surface 647 temperature and soil temperature at 5 cm depth

Long-term source apportionment of 649 ambient fine particulate matter (PM2. 5) in the Los Angeles Basin: A focus on emissions reduction from 650 vehicular sources

air pollution and cause-specific cardiovascular disease mortality

Annual application 655 and evaluation of the online coupled WRF-CMAQ system over North America under AQMEII phase 2

Estimating ground-level PM2.5 concentrations in the Southeastern United States using MAIAC AOD 662 retrievals and a two-stage model

Residential 664 building materials: An important source of ambient formaldehyde in mainland China

Long-667 range transport of Siberian biomass burning emissions and impact on surface

Source apportionment of PM2. 5 measured in South Texas near USA-670 Mexico border

Cubic convolution interpolation for digital image processing

COVID-19 pandemic and its positive impacts on environment: an 674 updated review

Principles and practice of structural equation modeling

Lateral collinearity and misleading results in variance-based SEM: An 677 illustration and recommendations

Drop in urban air pollution from 679 COVID-19 pandemic: Policy implications for the megacity of São Paulo

Impact of multicollinearity on small sample hydrologic regression models

The dark 684 cloud with a silver lining: Assessing the impact of the SARS COVID-19 pandemic on the global 685 environment

WHO's global air-quality guidelines

Enhancing the applicability of satellite remote sensing 688 for PM2.5 estimation using MODIS deep blue AOD and land use regression in California, United States

Estimating ground-level PM2.5 by fusing 691 satellite and station observations: a geo-intelligent deep learning approach

Long-range 694 transport of Asian pollution to the northeast Pacific: Seasonal variations and transport pathways of carbon 695 monoxide

Satellite-derived 697 correlation of SO2, NO2, and aerosol optical depth with meteorological conditions over East Asia from 698

Spatiotemporal distributions of surface ozone 700 levels in China from 2005 to 2017: A machine learning approach

Real-time 7-day forecast of pollen counts using a deep 702 convolutional neural network

Application of a Partial 704

Convolutional Neural Network for Estimating Geostationary Aerosol Optical Depth Data

A unified approach to interpreting model predictions

Spatiotemporal pattern of PM2.5 709 concentrations in mainland China and analysis of its influencing factors using geographically weighted 710 regression

Validation of the North American land data assimilation system (NLDAS) retrospective forcing over the 713 southern Great Plains

Effect of lockdown amid COVID-19 pandemic on air quality 715 of the megacity Delhi

Global patterns of current and 717 future road infrastructure

A review on black carbon emissions, worldwide 719 and in China

Covid-19 lockdown on PM10, SO2 and NO2 concentrations in Salé City (Morocco). Science of the total 722 environment

Air quality implications of 724 COVID-19 in California

Improvement in air quality and its impact on land surface temperature in major urban areas across India 727 during the first lockdown of the pandemic

Estimating PM2.5 concentration of the 729 conterminous United States via interpretable convolutional neural networks

Using the 732 satellite-derived NDVI to assess ecological responses to environmental change

An Indirect Data Assimilation Scheme for Deep Soil Temperature in the 735

Land Surface Model

Development of a Land Surface Model. Part II: Data Assimilation

Assessing the impact of PM2. 5 on respiratory disease using artificial neural networks

Concentration Trajectory Route of Air 744 pollution with an Integrated Lagrangian model (C-TRAIL Model v1. 0) derived from the Community 745

Multiscale Air Quality Model (CMAQ Model v5. 2)

Development 747 and Implementation of a Physics-Based Convective Mixing Scheme in the

Quality Modeling Framework

MERRA: NASA's modern-era retrospective analysis for research and applications

Trump's Snowballing China Travel Claim

Air quality during the COVID-19: PM2.5 analysis 755 in the 50 most polluted capital cities in the world

COVID-19 757 pandemic and lockdown measures impact on mental health among the general population in Italy. Frontiers 758 in psychiatry

Using a deep convolutional neural 760 network to predict 2017 ozone concentrations, 24 hours in advance

Relationships between the planetary boundary layer height and surface 763 pollutants derived from lidar observations over China: regional pattern and influencing factors. 764 Atmospheric Chemistry and Physics

COVID-19: Healthy environmental impact for public safety and menaces 766 oil market

Hotspots of black carbon and PM2. 5 in an urban area and relationships to traffic characteristics

North American Free Trade Agreement Implementation

Changes in air quality during the lockdown in Barcelona (Spain) one month into the SARS-CoV-2 774 epidemic

Correlation between 776 weather and Covid-19 pandemic in Jakarta

Organic and black carbon in PM2. 5 and PM10: 1 year of data from an urban site in Helsinki, Finland. 779 Atmospheric Environment

Estimating 1-km-resolution 781 PM2.5 concentrations across China using the space-time random forest approach

2020. Coronavirus Disease

World Health Organization Geneva

The Fire INventory from NCAR (FINN): a high resolution global model to estimate the emissions from 788 open burning

Estimating emissions from fires in North America for air quality modeling

Global Emissions of Trace Gases, Particulate Matter, 793 and Hazardous Air Pollutants from Open Burning of Domestic Waste

Public green spaces and positive mental health-796 investigating the relationship between access, quantity and types of parks and mental wellbeing

Coronavirus lockdowns and 799 stay-at-home orders across the U.S

NLDAS 802 primary forcing data L4 hourly 0.125× 0.125 degree V002. Goddard Earth Sciences Data and Information 803 Services Center (GES DISC)

Continental-805 scale water and energy flux analysis and validation for the North American Land Data Assimilation System 806 project phase 2 (NLDAS-2): 1. Intercomparison and application of model products

Impact of the 809 COVID-19 event on air quality in central China

Evaluating the impact of urban traffic investment on SO2 emissions 811 in China cities

The relationships between PM2.5 and 813 meteorological factors in China: seasonal and regional variations

Efficient PM2.5 forecasting using geographical 816 correlation based on integrated deep learning algorithms

Opposite impact of emission 818 reduction during the COVID-19 lockdown period on the surface concentrations of

Air quality changes in New York City 821 during the COVID-19 pandemic