key: cord-0743518-09kvh4bv
authors: Xiang, Jianbang; Austin, Elena; Gould, Timothy; Larson, Timothy; Shirai, Jeffry; Liu, Yisi; Marshall, Julian; Seto, Edmund
title: Impacts of the COVID-19 responses on traffic-related air pollution in a Northwestern US city
date: 2020-07-28
journal: Sci Total Environ
DOI: 10.1016/j.scitotenv.2020.141325
sha: f278be7e7ca8e912c5bb35dd4de662bb82408220
doc_id: 743518
cord_uid: 09kvh4bv

Abstract This study evaluates the COVID-19 impacts on traffic-related air pollution, including ultrafine particles (UFPs), PM2.5, black carbon (BC), NO, NO2, NOx, and CO in a Northwestern US city. Hourly traffic, air pollutants, and meteorological data on/near a major freeway in the downtown of Seattle, Washington, were collected for five weeks before and ten weeks after the Washington Stay Home Order (SHO) was enacted, respectively (February 17–May 31, 2020). The pollutants between pre- and post-SHO periods were compared, and their differences were statistically tested. Besides, first-order multivariate autoregressive (MAR(1)) models were developed to reveal the impacts specific to the change of traffic due to the COVID-19 responses while controlling for meteorological conditions. Results indicate that compared with those in the post-SHO period, the median traffic volume and road occupancy decreased by 37% and 52%, respectively. As for pollutants, the median BC and PM2.5 levels significantly decreased by 25% and 33%, relatively, while NO, NO2, NOx, and CO decreased by 33%, 29%, 30%, and 17%, respectively. In contrast, neither size-resolved UFPs nor total UFPs showed significant changes between the two periods, although larger particles (≥115.5 nm) decreased by 4–29%. Additionally, significant differences were found in meteorological conditions between the two periods. Based on the MAR(1) models, controlling for meteorological conditions, the COVID-19 responses were associated with significant decreases in median levels of traffic-related pollutants including 11.5–154.0 nm particles (ranging from −3% [95% confidence interval (CI): −1%, −4%] to −12% [95% CI: −10%, −14%]), total UFPs (−7% [95% CI: −5%, −8%]), BC (−6% [95% CI: −5%, −7%]), PM2.5 (−2% [95% CI: −1%, −3%]), NO, NO2, NOx (ranging from −3% [95% CI: −2%, −4%] to −10% [95% CI: −18%, −12%]), and CO (−4% [95% CI, −3%, −5%]). These findings illustrate that the conclusion of the COVID-19 impacts on urban traffic-related air pollutant levels could be completely different in scenarios whether meteorology was adjusted for or not. Fully adjusting for meteorology, this study shows that the COVID-19 responses were associated with much more reductions in traffic-related UFPs than PM2.5 in the Seattle region, in contrast to the reverse trend from the direct empirical data comparison.

The ongoing global pandemic of coronavirus disease caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has led to more than 355,000 premature deaths worldwide, including 100,000 in the United States (US) alone as of May 27, 2020 (Johns Hopkins University, 2020 . Washington State, located in the northwest of the US, first reported a COVID-19 confirmed case on January 20, 2020, and declared a state of emergency on March 3 and issued a Stay Home Order (SHO) on March 23 (see Appendix Fig. A1 for a timeline of events). As a result, the road traffic volume and patterns in Washington were altered. By April 3, traffic on Interstate 5 (I-5) fell by over 50% in both downtown Seattle and Everett (40 km to the north) as compared to typical traffic volumes in February (Washington State Department of Transportation, 2020b). The Stay Home Order, which lasted ten weeks, ended on May 31. Road traffic is a major urban pollutant source in the Seattle area based on multiple apportionment studies (Friedman, 2020; Larson et al., 2004; Maykut et al., 2003; Wu et al., 2007) . Considering the unprecedented decrease of urban traffic due to the COVID-19 responses, a relatively large decrease in urban air pollutant levels was anticipated in Washington and other J o u r n a l P r e -p r o o f 6 layer (Banks et al., 2015; Hu et al., 2010) , which plays an important role in air pollution formation and dispersion. Therefore, the impacts of meteorological factors may not be completely and accurately captured by the WRF model. Hence, it is of great value to develop an approach to account for these confounding factors fully. Road traffic emits a wide variety of pollutants and is one of the main contributors of urban ultrafine particles (UFPs; diameter ≤ 100 nm) (Harrison et al., 2011; Pant and Harrison, 2013; Vu et al., 2015) and black carbon (BC) (Miguel et al., 1998; Zheng et al., 2015) , no studies have examined the COVID-19 impacts on ambient UFPs and only one study in Milan, Italy involved BC (Collivignarelli et al., 2020) , to the authors' knowledge.

Near-roadside sampling was conducted in Seattle, Washington, preceding and following the Washington State SHO. Pollutant data collected at a roadside monitoring station in Seattle was combined with roadway traffic and meteorology data to better understand any changes in air quality related to decreased traffic. This paper aims to evaluate the impacts of the COVID-19 responses on traffic and corresponding traffic-related air pollution (TRAP), including UFPs, J o u r n a l P r e -p r o o f 7 2. Methods

The Washington State Department of Transportation (WSDOT) operates multiple loop detector stations along main routes in Washington, including I-5, which connects Canada, Washington, Oregon, California, and Mexico from north to south. The hourly traffic data from and road occupancy (%) defined as a percent of the time a short space on the road is occupied by vehicles (Hall, 1996) . Traffic data were excluded as invalid or unusable under one or more of the following conditions: (1) flagged as errors by the TRACFLOW system; (2) volume or occupancy was less than 0; (3) time when the loop detectors were working in only one direction (either northbound or southbound).

J o u r n a l P r e -p r o o f 8 Size-resolved nanoparticles were measured at 1-minute intervals from March 7 to May 31, 2020, using a NanoScan SMPS Nanoparticle Sizer (Model 3910, TSI Inc., MN) deployed at the same site by our team from University of Washington. The NanoScan measures particle number concentrations (PNCs) over a size range of 10 to 420 nm particles with a resolution of 13 size bins (11.5 nm, 15.4 nm, 20.5 nm, 27.4 nm, 36.5 nm, 48.7 nm, 64.9 nm, 86.6 nm, 115.5 nm, 154.0 nm, 205.4 nm, 273.8 nm, 365.2 nm) . Pollutant data were excluded under one or more of the following conditions: (1) flagged as errors by the WAQA system or the NanoScan instrument; (2) pollutant level was less than 0; (3) incomplete hourly measures by the NanoScan (< 50% minute data). We additionally limited the size-resolved data from the NanoScan to the size bins < 205.4 nm as the larger bins reported more than 50% of zeros.

Hourly meteorological data, including ambient wind direction and speed, as well as the temperature at the Seattle-10 th &Weller monitoring station, were obtained from the WAQA system. Hourly ambient relative humidity (RH) and precipitations were obtained from the BFI station (about 7.5 km south of the Seattle-10 th & Weller site) in Washington State Automated Surface Observing System (ASOS) Network (Iowa Environmental Mesonet, 2020) as they were not routinely monitored at the Seattle-10 th & Weller site. J o u r n a l P r e -p r o o f 9 2020) and the post-SHO period (ten weeks, March 23 -May 31, 2020). The week right before the Washington SHO enacted was defined as Week 0 (March 16 -22, 2020) . Note that size-resolved PNC data measured from NanoScan were not available for Weeks (-4) to (-3), while all other traffic, pollutant, and meteorological data were available for the entire fifteen-week period.

Total UFP number concentrations (NCs) were calculated by summing the PNCs of eight size bins (median sizes ranging from 11.5 -86.6 nm). The PNC data were aggregated into hourly means after applying the aforementioned exclusion criteria. Then, the fifteen-week traffic, pollutant, and meteorological data were merged based on date and hourly time.

None of the traffic, pollutant, and meteorological variables were normally distributed based on Shapiro-Wilk tests. Therefore, Wilcoxon two-sample rank-sum tests were conducted for each traffic, pollutant, and meteorological variable between the pre-and post-SHO periods to compare the differences between these two periods. Wilcoxon effect sizes, which indicate the strength of the differences were calculated using the "coin" and "rstatix" packages in R (R Core Team, 2013).

Road occupancy is an equivalent variable to traffic density, which can be expressed as traffic volume divided by speed (Hall, 1996) . Theoretically, TRAP levels at a near-road monitoring site should be directly related to the traffic density rather than volume, although Journal Pre-proof J o u r n a l P r e -p r o o f 10 traffic density and volume are highly correlated. Additionally, road occupancy in the adjacent areas between the west and east of the air monitoring station should be much more similar than traffic volume. Therefore, the change of road occupancy was used as the primary indicator for TRAP impacts due to the COVID-19 responses in the present study.

Due to the potential autocorrelation among observations from the time-series air pollution measurement, multivariate autoregressive (MAR) models were used to analyze traffic-pollutant associations between road occupancy and each pollutant level (Holmes et al., 2020; Neumaier and Schneider, 2001) . Based on the partial autocorrelation function (PACF) computing for each pollutant, the p orders for the MAR(p) models were set to 1 for all pollutants (Fig. A3) , which means the pollutant at time t was based on the immediately preceding value at time t-1.

Depending on data distribution and statistical tests, pollutant concentrations were natural log-transformed in the MAR(1) models. With road occupancy as the main traffic-related indicator, the models (Equation (1)) were fitted for each pollutant in which the dependent variable was the natural log-transformed pollutant level at time t (log(y t )), and the independent variables included the natural log-transformed pollutant level at lag hour 1 (log(y t-1 )), traffic-related indicator (traffic t ), and meteorological variables such as temperature (T t ), relative humidity (RH t ), precipitation (P t ), wind speed (WS t ), and a category variablewind direction (WD t ). According to the layout of I-5 near the air quality monitoring station (Fig. A2) , wind direction (degrees from the north) was categorized based on a northwest-southeast line (i.e., 135 -315 degrees as the reference category (level = 0, mostly west wind), and the rest as the other Journal Pre-proof J o u r n a l P r e -p r o o f 11 category (level = 1, mostly east wind)). Outliers were determined by using Cook's distance value.

The observations with Cook's distance value greater than 0.5 were excluded from the regression analysis (Cook, 1977; Yerramshetty and Akkus, 2008) .

where β 0 -β 7 are the coefficients from the MAR(1) model; ε is the residual.

The estimates of β 2 determined from the MAR(1) models were used to predict the median percent changes of pollutant levels due to the COVID-19 responses, as shown in Equation (2):

where Δ traffic is the median absolute change of road occupancy due to the COVID-19 responses.

Based on pairwise comparisons using Wilcoxon rank-sum tests, the median road occupancy was not statistically different (p > 0.2) among Weeks (-4), (-3), and (-2), but statistically different (p < 0.05) between Week (-2) and Week (-1) (see more in the Results section). As there were no significant differences for the median hourly traffic between Week (-2) and previous weeks, and all data were available in Week (-2), Week (-2) was taken as the reference week when there were no significant impacts on the median traffic from the COVID-19 responses. Therefore, Δ traffic was calculated as the difference of median hourly traffic between the post-SHO weeks and Week (-2), J o u r n a l P r e -p r o o f 12 as shown in Equation (3):

( 2) post SHO Week traffic traffic traffic

where traffic post-SHO and traffic

Week (-2) are the median hourly road occupancy during Weeks 1 to 10 and Week (-2), respectively.

The median absolute changes of pollutant levels due to the COVID-19 responses were calculated by multiplying the percent changes with the median hourly concentration in Week (-2) for each pollutant, respectively.

In the main model, data in Week 0 were excluded as there probably existed significant air pollution emission sources from regional events in that week (see more in the Results section).

The sensitivity of the MAR(1) results to the data in Week 0 was examined by including the data in Week 0. Additionally, based on the wind rose plot (Appendix Fig. A4 ), the sensitivity to the wind direction categorizing was examined by categorizing the wind direction based on a north-south line into the west (180 -360 degrees, reference category) and the east (0 -180 degrees). Also, five more models were assessed: (1) without considering the autocorrelation effects in the time-series observations (excluding the log(y t-1 ) variable in the MAR(1) model), (2) with a second-order multivariate autoregressive (MAR(2)) model, (3) For all statistical tests, p = 0.05 indicated statistical significance in this study. All calculations and figures were made using "data.table", "stats", "mgcv", "coin" and "rstatix", "tidyverse", "openair","ggplot2", "ggpubr" and "leaflet" packages in R, Version 3.3.0 embedded in RStudio Version 1.1.456. significantly lower than that in the pre-SHO period, the traffic from Weeks 1 to 10 in the post-SHO period gradually increased. By the end of the post-SHO period (Week 10), the median traffic TOV volume and road occupancy were 68% and 54% of those in Week (-2), respectively.

The concentrations of BC, PM 2.5 , and gaseous pollutants seem to be generally higher in the pre-SHO period than those in the post-SHO period, while the trend is unclear for UFPs. It can be clearly seen that the trend of weekly variation of all pollutants does not mirror that of traffic, especially for Week 0, where an unexpected spike in most pollutants is observed. Relatively large drops in concentrations were observed in most pollutants, except in 11.5 -20.5 nm particles, from Weeks 0 to 1. For instance, median UFPs, PM 2.5 , and NO x concentrations decreased by 43%, 75%, and 52%, respectively. The lower wind speed in Week 0 relative to that in Weeks (-1) and 1 possibly contributed to the spike in that week. However, this cannot explain why BC, PM 2.5 , NO 2 , and NO x levels in Week 0 were significantly higher than those in Week (-3) even though wind speed and other meteorological variables were comparable in the two weeks, while traffic volume and road occupancy in Week 0 decreased by 25% and 53% relative to those in

Week (-3), respectively. It can be inferred that there probably existed significant air pollution emission sources from regional events in Week 0.

As regional events showed a large impact on pollutants at Seattle-10 th &Weller monitoring and from 23% to 11% (a drop of ~52%), respectively. As for pollutants, the median BC and PM 2.5 levels decreased by 25% and 33%, relatively, while NO, NO 2 , NO x , and CO decreased by 33%, 29%, 30%, and 17%, respectively. In contrast, although larger particles (≥ 115.5 nm) decreased by 4 -29%, neither size-resolved UFPs nor total UFPs show significant changes between the two periods. As for meteorological conditions, there was more wind from the west (degree > 180) in the post-SHO period. Compared with those in the pre-SHO period, median wind speed and temperature increased by 0.3 m/s (19%) and 6 °C, respectively, while relative humidity decreased by 7%. Median precipitations were both 0 μm in the two periods, while mean precipitation increased by 21 μm (122%) in the post-SHO period. Wilcoxon tests indicate that the differences between pre-and post-SHO periods are significant for all traffic, pollutants except J o u r n a l P r e -p r o o f 18 (changes ranging from 0 to 1%), further confirming that the MAR(1) models are good enough to capture the autocorrelation effects (Fig. A10) . Including an interaction between the amount of road occupancy and the two WD categories did not improve the model, and the coefficients for (Table A4) .

Despite the variations in pollutant percent change predictions among all the models, the relative magnitudes among different pollutants are generally unchanged. It further indicates traffic contributed more to UFPs, BC, and NO, than PM 2.5 and CO. The main models generally have the larger R 2 and smaller Akaike and Bayes information criteria compared with the sensitivity analysis models, and the results from the main models are statistically significant.

Therefore, the results from the main model analyses are primarily reported.

J o u r n a l P r e -p r o o f

Previous studies have examined the impacts of the COVID-19 responses on ambient PM 2.5 and/or some other gaseous pollutants in China (Bauwens et al., 2020; Chen et al., 2020; Li et al., 2020; Shi and Brasseur, 2020; Xu et al., 2020) , India (Sharma et al., 2020) , Italy (Collivignarelli et al., 2020) , Brazil (Nakada and Urban, 2020) , and NO 2 in Western Europe and major

Northeastern US cities (Bauwens et al., 2020) . Additionally, an ongoing study has estimated the impacts on PM 2.5 and O 3 across the US, as well as NO 2 in three US cities, including Seattle (Bekbulat et al., 2020) . Another recent study tried to exclude meteorological impacts on the change of air pollution during the COVID outbreak by using the WRF model (Zhao et al., 2020) .

However, the impacts of meteorological factors may not be completely and accurately excluded due to the inherent limitations of the WRF model (Zhao et al., 2020) . To the authors' knowledge, the present study is the first to examine the impacts of the COVID-19 responses on size-resolved and total UFPs, and the first to reveal the impacts on TRAP specific to the change of traffic after fully adjusting for meteorological variations.

In the main analyses, the model predictions were based on the median reduction of road occupancy between the pre-and post-SHO periods. Although the mean hourly road occupancy in J o u r n a l P r e -p r o o f 22 (Bauwens et al., 2020; Li et al., 2020; Shi and Brasseur, 2020; Xu et al., 2020) and in other parts of the US (Bauwens et al., 2020; Bekbulat et al., 2020) . This can be explained by five reasons: (1) the results herein are specific to the change in traffic while the other studies included the changes in all anthropogenic sources; (2) the results herein were fully adjusted for meteorology while those in the other studies were not; (3) the contributions of traffic to different air pollutants tend to vary spatially and temporally; (4) the MAR(1) models used in the present study tend to underestimate the impacts of current-hour traffic on next-hour air pollution, as discussed above;

and (5) the results in the present study are based on medians while those in some previous studies were based on means (Bauwens et al., 2020; Li et al., 2020) . Due to the right-skewed distributions of air pollutant levels, the mean reductions were generally larger than median reductions. With the mean reduction of road occupancy as the prediction metric, the predicted percent reductions in pollutants would be 1 -7% larger compared with the main model ( Fig.  Figures   Fig. 1 . Weekly boxplots of traffic, air pollutants, and meteorological data from the Seattle-10 th & Weller monitoring site in the pre-and post-SHO periods. Week 1 was March 23-29, 2020. The size-resolved particles were not available in Weeks (-4) and (-3). All other pollutants and meteorology parameters were measured by government facilities continuously. Outliners (larger than Q3+1.5*IQR or smaller than Q1-1.5*IQR) were excluded from the figure.

Definition of abbreviations: TOV = total vehicle; 11.5nm-154.0nm = 11.5nm-154.0 nm particles; PNC = particle number concentration; BC = black carbon; WD = wind direction (degrees from the north); WS = wind speed; T = temperature; RH = relative humidity; SHO = Stay Home Order; Q1 = 25 th quantile; Q3 = 75 th quantile; IQR = interquartile range. 

Air pollution meteorology and dispersion

Impact of emissions from the Los Angeles port region on San Diego air quality during regional transport events

Performance evaluation of the boundary-layer height from lidar and the Weather Research and Forecasting model at an urban coastal site in the north-east Iberian Peninsula

Impact of coronavirus outbreak on NO2 pollution assessed using TROPOMI and OMI observations

and ozone air pollution levels have not dropped consistently across the US following societal COVID response

Estimating the contribution of regional transport to PM2. 5 air pollution in a rural area on the North China Plain

Air pollution reduction and mortality benefit during the COVID-19 outbreak in China. The Lancet Planetary Health

Lockdown for CoViD-2019 in Milan: What are the effects on air quality?

Detection of influential observation in linear regression

Source apportionment of PM2.5 at two Seattle chemical speciation sites

Traffic stream characteristics

PMF analysis of wide-range particle size spectra collected on a major highway

Applied time series analysis for fisheries and environmental data

Evaluation of Three Planetary Boundary Layer Schemes in the WRF Model

Biogenic versus anthropogenic sources of CO in the United States

Ozone as a function of local wind speed and direction: Evidence of local and regional transport

COVID-19 dashboard

Source apportionment of indoor, outdoor, and personal PM2.5 in Seattle, Washington, using positive matrix factorization

Air quality changes during the COVID-19 lockdown over the Yangtze River Delta Region: An insight into the impact of human activity pattern changes on air pollution variation

Source apportionment of PM2.5 at an urban IMPROVE site in Seattle

On-road emissions of particulate polycyclic aromatic hydrocarbons and black carbon from gasoline and diesel vehicles

COVID-19 pandemic: Impacts on the air quality during the partial lockdown in São Paulo state

Estimation of parameters and eigenmodes of multivariate autoregressive models

Estimation of the contribution of road traffic emissions to particulate matter concentrations from field measurements: A review

R: A language and environment for statistical computing

Effect of restricted emissions during COVID-19 on air quality in India

The response in air quality to the reduction of Chinese economic activities during the COVID-19 outbreak

Principal component and canonical correlation analysis for examining air pollution and meteorological data

Review: Particle number size distributions from seven major sources and implications for source apportionment studies

Washington State 2014 comprehensive emissions inventory

Washington Air Quality Advisory

Cabinet data sources

What's going on with Seattle area traffic with COVID-19

Source apportionment of PM2.5 and selected hazardous air pollutants in Seattle

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

The authors wish to express special thanks to Mr. Manouchehr Goudarzi and Dr. Hui Dong from Washington State Department of Transportation (WSDOT) for helping the authors

 (Table A3) , Model comparison (Table A4) .