development and comparison of air pollution exposure surfaces

Subscriber access provided by - Access paid by the | UCSB Libraries

Article

DEVELOPMENT AND COMPARISON OF AIR POLLUTION EXPOSURE SURFACES DERIVED FROM ON-ROAD MOBILE MONITORING AND SHORT-TERM STATIONARY SIDEWALK MEASUREMENTS Laura Minet, Rick Liu, Marie-France Valois, Junshi Xu, Scott Weichenthal, and Marianne Hatzopoulou Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.7b05059 • Publication Date (Web): 23 Feb 2018 Downloaded from http://pubs.acs.org on February 26, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Environmental Science & Technology is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 30

Environmental Science & Technology

1 2 3 4 5 6

DEVELOPMENT AND COMPARISON OF AIR POLLUTION EXPOSURE SURFACES DERIVED FROM ON-ROAD MOBILE MONITORING AND SHORT-TERM STATIONARY SIDEWALK MEASUREMENTS

7

Marianne Hatzopoulou1*

Laura Minet1, Rick Liu1, Marie-France Valois2, Junshi Xu1, Scott Weichenthal3,

8 1

9 10

2

Department of Civil Engineering, University of Toronto, Toronto, Ontario, Canada

Division of Clinical Epidemiology, Faculty of Medicine, McGill University, Montreal,

11 12 13

Quebec, Canada 3

Epidemiology, Biostatistics & Occupational Health, Faculty of Medicine, McGill University, Montreal, Quebec, Canada

14

*Corresponding author, Department of Civil Engineering, University of Toronto

15

35 St George Street, Toronto, ON M5S 1A4

16

Telephone: 416-978-0864. Email: [email protected]

1 ACS Paragon Plus Environment


Page 2 of 30

17

ABSTRACT

18

Land-use regression (LUR) models of air pollutants are frequently developed based on short-

19

term stationary or mobile monitoring approaches, which raises the question of whether these two

20

data collection protocols lead to similar exposure surfaces. In this study, we measured Ultrafine

21

Particles (UFP) and Black Carbon (BC) concentrations in Toronto during summer 2016, using

22

two short-term data collection approaches: mobile, involving 3,023 road segments sampled on

23

bicycles, and stationary, involving 92 sidewalk locations. We developed four LUR models and

24

exposure surfaces, for the two pollutants and measurement protocols. Coefficients of

25

determination (R2) varied from 0.434 to 0.525. Various small-scale traffic variables were

26

included in the mobile LUR. Pearson correlation coefficients between the mobile and stationary

27

surfaces were 0.23 for UFP and 0.49 for BC. We also compared the two surfaces using personal

28

exposures from a panel study in Toronto conducted during the same period. The personal

29

exposures differed from the outdoor exposures derived from the combination of GPS information

30

and exposure surfaces. For UFP, the median for personal outdoor exposure was 26,344 part/cm3,

31

while the cycling and stationary surfaces predicted medians of 31,201 and 19,057 part/cm3.

32

Similar trends were observed for BC, with median exposures of 1,764 (personal), 1,799 (cycling)

33

and 1,469 ng/m3 (stationary).


Page 3 of 30

34


TOC ART

35 36 37

1

38

Understanding the small-scale variations of Ultrafine Particles (UFP) and Black Carbon (BC) in

39

an urban area is crucial for accurate exposure assessment. However, the cost of sampling

40

equipment often puts a curb to large monitoring campaigns with numerous simultaneous

41

measurements. Land-use regression models (LUR) have been widely used to develop air

42

pollution exposure surfaces for various traffic-related air pollutants, which were then used in the

43

context of epidemiologic studies1–4. Early LUR models were developed for nitrogen dioxide

44

(NO2), primarily because of the ease of measuring NO2 using passive samplers and its common

45

use as a proxy for traffic-related air pollution. The recent availability of portable monitors for

46

UFP and BC has motivated the development of LUR models. Yet, the nature of the instruments

47

available to measure these two pollutants (e.g. high cost, power requirements, need for daily

48

maintenance), have made it difficult to conduct long-term measurements at fixed locations thus

49

giving rise to short-term monitoring campaigns. These campaigns have either taken place via

INTRODUCTION



Page 4 of 30

50

monitoring at fixed sites5–11, or via mobile monitoring, involving measurements conducted in

51

motion9,12–18. In addition to allowing the investigation of a spatially larger area, the costs of

52

sampling via mobile monitoring can be reduced5,19,20. However, the duration of sampling at each

53

location (usually represented by a road segment) is shorter than the time typically spent at fixed

54

locations, and it is difficult to quantify the number of repeated measures necessary in order to

55

obtain a representative average for each road segment.

56

LUR models and associated exposure surfaces are intrinsically dependent on the data

57

collected, e.g. the sampling locations, repeated observations, and the pollutant concentrations

58

associated with them. The influence of the data collection protocol on LUR models is therefore a

59

topical question. With the emergence of mobile data collection approaches, the question of

60

whether to promote the spatial extent at the expense of the duration of sampling becomes

61

pertinent since the two approaches may lead to different models. Besides spatial extent and

62

sampling duration, the two approaches differ in the location of the measurements: mobile

63

sampling inherently occurs on the road while stationary sampling often occurs on the sidewalk.

64

In the case of dissimilar results, knowing which approach leads to a more representative surface

65

is crucial to guide future studies.

66

In the context of UFP and BC, we explored the influence of the sampling approach on

67

LUR models and associated exposure surfaces. We designed short-term stationary and mobile

68

monitoring campaigns, occurring simultaneously during summer 2016 in Toronto, Canada. LUR

69

models and exposure surfaces were derived for both pollutants and based on the two data

70

collection approaches separately. The results of the exposure surfaces were compared with short-

71

term outdoor personal exposures collected in a panel study that occurred during the same time

72

period.


Page 5 of 30


73

2

74

2.1

75

Two simultaneous campaigns based on mobile and stationary protocols were conducted between

76

May and August 2016 in Toronto with the aim of optimising spatial coverage and temporal

77

coverage of a 7am to 7pm day.

METHODOLOGY Data collection

78

The short-term stationary monitoring campaign was based on 92 near-road sampling

79

locations (on the sidewalk, at the edge of the road), equally distributed between intersections and

80

mid-blocks. Each visit to a fixed point consisted of a 20-minute sampling duration, whereby

81

research assistants counted vehicles (i.e. passenger car, passenger truck, light commercial truck,

82

short haul truck, long haul truck, coach bus, school bus, transit bus, and street cars) going

83

through the intersection, or driving along the road. In contrast, the mobile data collection was

84

performed using bikes. Ten routes of 24 to 31 km each were designed, passing through most of

85

the fixed points, and representing a total of 270 km and 3,895 unique road segments (FIGURE

86

S1).

87

Each day was divided into three time blocks of 4 hours each (7am-11am, 11am-3pm, and

88

3pm-7pm). Measurements along biking routes and fixed points were conducted at least once per

89

time block for a minimum of five repetitions. We randomized the direction of the routes and the

90

order of fixed points sampling, as well as the starting time. More detail on the campaign design

91

is available in the supplemental information (SI).

92

2.2

93

Miniature Diffusion Size Classifiers (DiSCMini) and Microaethalometers (Microaeth model

94

AE51) were used to collect UFP and BC concentrations with time resolutions of 1 and 30

Instrumentation



Page 6 of 30

95

seconds respectively. Both instruments were carried in backpacks with a tube connected to them

96

and pointing outside of the bag.

97

Seven DiSCMinis and six Microaeths were used during the campaign. As such, all

98

instruments were collocated with a reference monitoring station (Particle Counter TSI 651 for

99

UFP and Aethalometer AE 633 for BC) by concomitantly sampling air from the same outdoor

100

source. We repeated this process before and after the campaign, over more than 48h each time.

101

The purpose of this collocation was to select a reference DiSCMini and a reference Microaeth

102

based on the Pearson correlation coefficients between the simultaneous measurements of all the

103

instruments. The DiSCMini and the Microaeth with the highest average correlation were chosen

104

as the reference, and the other instrument measurements were corrected based on simple linear

105

regressions between the reference measurements and the measurements to correct. Records from

106

the entire summer were then corrected based on the linear regressions determined.

107

Garmin 800 GPS and the GPS mobile application Strava were used for the mobile

108

monitoring campaign and they were set up to record GPS coordinates every second.

109

2.3

110

The BC measurements were corrected with the Optimized Noise-Reduction Algorithm (ONA)

111

developed by the U.S. EPA to improve BC measurements using Aethalometers21.

Data processing

112

We also attempted to adjust UFP and BC measurements for daily variations in air quality.

113

A fixed station located at street level next to a downtown university building recorded hourly

114

UFP and BC concentrations. We used these measurements to develop a correction factor similar

115

to other studies5,18. This correction factor was specific to each day, as shown in equation (1). All

116

measurements were multiplied by the correction factor (after correction by collocation, and ONA

117

for BC). 6 ACS Paragon Plus Environment

Page 7 of 30


118 Median UFP ሺor BCሻ concentration over the campaign

119

Correction Factor = Median UFP ሺor BCሻ concentration over the day of measurement

(1)

120

The fixed points and cycling databases were then developed separately. For each visit to

121

a fixed point, the 20 minutes of sampling were averaged. Meteorological variables (temperature,

122

relative humidity, wind speed) corresponding to the measurement time and day were retrieved

123

from the Toronto International Airport station. Traffic data were also stored. Subsequently, all

124

visits to a same fixed point were averaged, providing average UFP and BC concentrations, and

125

average meteorological conditions and traffic counts encountered at that point.

126

For the cycling measurements, a few GPS records were sometimes missing due to signal

127

losses. We therefore linearly interpolated the missing coordinates. The second step was to

128

associate each GPS point with a road segment of the bikeways network provided by the City of

129

Toronto on its Open Data portal (June 2016), which includes the road network of the city to

130

which cycle tracks and trails not accessible by vehicles were added. For this step, we first drew

131

the prescribed biking routes on ArcMap 10.4.1 before snapping the GPS points to the closest

132

road segment of the route drawn to limit the potential mismatch of points with road segments.

133

Then, UFP and BC values were associated with each GPS point based on the time stamp. BC

134

concentrations were recorded every 30 seconds as an average of the 30 previous seconds. We

135

therefore associated a same BC concentration with the 30 corresponding seconds of sampling.

136

Hourly meteorological variables were also recorded. We first averaged UFP and BC

137

concentrations, and meteorological variables by visit to a segment. Subsequently, we averaged

138

all visits to a road segment in order to obtain unique UFP and BC values, and meteorological

139

variables. Our final database consisted of 3,023 rows, equivalent to the number of road segments

140

covered at least once during the campaign. 7 ACS Paragon Plus Environment


Page 8 of 30

141

2.4

142

Built environment and land use variables were computed using ArcMap 10.4.1 for all the fixed

143

points and road segments (TABLE S5).

Land use and built environment characteristics

144

First, we calculated distance variables as the distance between the point or the segment

145

and a specific feature: distance to the shore, to the Central Business District (CBD) and to the

146

closest airport, rail line, highway, major road, as well as the nearest Nitrogen Oxides (NOx) and

147

Particulate Matter (PM) emitting chimney. Other variables computed were within buffers of

148

sizes 50m, 100m, 200m, 300m, 500m, 750m and 1000m created around each fixed point or road

149

segment, and normalized by the area of the buffer. Areas of building footprint and of

150

commercial, governmental, industrial, parks, residential, open area, and waterbody land uses

151

were computed, as well as lengths of highway, major road, road and bus line. The number of

152

intersections, bus stops, trees, NOx and PM emitting chimneys within the buffers were also

153

calculated. The number of inhabitants within the 500m, 750m and 1000m buffers was

154

determined. The average and the maximum building height were determined within buffers of

155

25m, 50m and 100m. Finally, we computed the average hourly traffic volume within the various

156

buffers.

157

Land use, building footprint, road network, rail lines, and aerodrome shapefiles were

158

extracted from the DMTI Spatial Inc. Database (2014). We retrieved the rail line, bus routes, bus

159

stops, trees and building height shapefiles from the City of Toronto Open Data portal (June

160

2016). We also downloaded on that portal the Neighbourhood shapefile and the Demographics

161

data of 2011 that we combined to determine the number of inhabitants within each

162

neighbourhood, a homogenous distribution of the inhabitants within each neighbourhood being

163

hypothesized. The shapefiles locating the NOx and PM emitting chimneys were downloaded


Page 9 of 30


164

from the National Pollutant Release Inventory (NPRI) website of the Government of Canada.

165

Finally, average traffic was generated between the hours of 6am and 7pm using EMME, a traffic

166

assignment model for the Greater Toronto and Hamilton Area (GTHA). The Origin-Destination

167

(OD) matrix of trips from the Transportation Tomorrow Survey (year 2011) was used and a

168

static traffic assignment method was employed to generate a traffic volume on the roads of the

169

GTHA.

170

2.5

171

We developed separate UFP and BC LUR models based on the fixed points and cycling data

172

using the programming language R and its statistical package. The normal logarithms of the UFP

173

and BC concentrations were computed.

Land use regression models

174

Two procedures based on the literature were tested and applied to both the temporally

175

and non-temporally adjusted UFP and BC concentrations. The first procedure, later called the

176

forward model, was based on the methodology presented by Eeftens et al.22, which was also

177

followed by Kerckhoffs et al.9 and Sabaliauskas et al.16. The steps of the algorithm were as

178

follows: 1) The potential predictors were sorted based on the adjusted R2 of the simple linear

179

regression between the predictor and ln(BC) or ln(UFP). 2) When considering non-temporally

180

adjusted pollutant concentrations, the initial model consisted in the best combination of

181

meteorological variables identified, i.e. the combination providing the highest adjusted R2. When

182

considering temporally adjusted pollutant concentrations, no predictor was added to the initial

183

model. 3) The variables were added one at a time to the initial model based on the order

184

determined in step 1. Variables were kept only if including them increased the adjusted R2 by

185

more than 1%, if the sign of the new variable in the model was meaningful, and if adding the

186

new variable did not alter the signs of the variables already in the model. At this step, we 9 ACS Paragon Plus Environment


Page 10 of 30

187

checked whether the new variable was from the same category as a predictor already included in

188

the model (i.e. same variable but different buffer). If it was the case, we tested which variable

189

was more appropriate by keeping one of the variables and checking the one that led to the

190

highest increase in the adjusted R2. Also, variables that were eliminated in a previous step were

191

always tested again. 4) Once an intermediate model was reached, i.e. when all variables had been

192

tested and adding any of them did not improve the adjusted R2, we removed one by one the

193

variables that had a p-value higher than 0.1 starting with the variable with the highest p-value. 5)

194

We checked the Variance Inflation Factor (VIF) of the variables and removed the variables with

195

a VIF higher than 3, starting with the variable with the highest VIF.

196

The second procedure, called the backward model, was based on the paper of

197

Weichenthal et al.16. The steps followed are as follows: 1) The potential predictors were sorted

198

based on the p-value of the simple linear regression between the predictor and ln(BC) or

199

ln(UFP). 2) All the variables that were significant (i.e. p-value ≤ 0.05) were added to the initial

200

model. If two variables were highly correlated (i.e. Spearman correlation coefficient ≥ 0.8), then

201

the variable presenting the highest R2 in the univariate regression with the pollutant was kept. 3)

202

Variables in the regression with a coefficient sign opposed to what was expected were removed.

203

4) Variables that were not statistically significant were removed one at a time and only if doing

204

so did not increase the Root Mean Square Error (RMSE) by more than 1%. 5) The non-

205

significant variables remaining (i.e. p-value higher than 0.1) were removed starting with the

206

variable with the highest p-value even if doing so increased the RMSE by more than 1%.

207

In addition, we performed a 100-fold cross-validation. We randomly selected 90% of the

208

points or road segments, and applied to them the forward and backward models developed

209

previously to determine new coefficients. We then applied the new model to the 10% hold-out


Page 11 of 30


210

sample, and calculated the RMSE and Pearson correlation coefficient between the predicted and

211

measured concentrations of this sample. We repeated this 100 times. We based our model choice

212

on the adjusted R2 of the model, as well as on the median adjusted R2 of the validation models

213

(i.e. models developed with 90% of the points) and on the median RMSE and Pearson

214

correlation coefficients of the hold-out samples.

215

After choosing our final fixed point model, we added the proportion of trucks (calculated

216

as the number of short and long haul trucks divided by the total number of vehicles) as a

217

predictor to see its impact on the adjusted R2 of the model, although it could not be used for

218

predictive purposes since the traffic data we are using only includes a total traffic volume.

219

Finally, we developed models based only on the road segments within 100m around the

220

fixed points. The aim was to compare the predictors included in these models with the predictors

221

included in the fixed point models.

222

2.6

223

To develop exposure surfaces, we divided the city of Toronto into grid cells of 100m by 100m

224

and computed the predictors for the mid point of each cell. We calculated the predicted pollutant

225

concentrations for each mid-point based on the final LUR models and associated each value with

226

the corresponding cell to generate surfaces. We slightly modified the area of Toronto considered

227

for the development of the surfaces (FIGURE S2). Indeed, the International airport of Toronto is

228

located West of the city; however, since we did not conduct measurements close to the airport

229

(due to the nature of our monitoring campaign being based on pedestrian and cycling

230

measurements), we only predicted levels up to the extent of our sampling coverage.

Exposure surfaces and comparison with panel study

231

The surfaces were not intended to predict short-term exposures; however, to explore their

232

differences, they were compared with data from a panel study also conducted in Toronto during 11 ACS Paragon Plus Environment


Page 12 of 30

233

the same time period as the cycling and stationary campaigns (from May to August 2016).

234

During this period, we collected data from a total of 62 visits by various participants, each visit

235

consisting in one day extending from 9am to 4pm. Several visits could take place during the

236

same day if more than one participant was recruited. The panel study was designed at the same

237

time as the stationary and cycling campaigns. The participants were equipped with DiSCMinis,

238

Microaeths and GPS during six hours each day and were asked to spend at least two hours

239

outdoors. For the current study, only the outdoor personal exposures were considered. We

240

determined the average outdoor UFP and BC exposure that each participant was exposed to

241

during the study (personal exposure). A participant's outdoor trajectory was then intersected with

242

the exposure surfaces to derive mobility-based exposures. The personal and mobility-based

243

exposures were compared to identify the extent of bias in short-term exposure estimates

244

generated by the various surfaces.

245

3

246

3.1

247

Hourly UFP and BC concentrations recorded by the DiSCMinis and Microaeths exhibited very

248

similar trends compared to the measurements of the reference station. However, the ranges could

249

vary a lot for UFP; the highest difference between two DiSCMini measurements was 21,000

250

particles/cm3 (noted part/cm3). The Pearson correlation coefficients between the DiSCMinis

251

ranged from 0.36 to 0.98 with an average of 0.82. The chosen reference DiSCMini had

252

correlation coefficients of 0.75 to 0.98 with the six other instruments. As for the Microaeths, the

253

differences between two instruments were low, especially during the second colocation when the

254

largest hourly averaged BC concentration difference between two instruments was less than 100

255

ng/m3. The correlation coefficients between the instruments ranged from 0.71 to 0.99 for an

RESULTS Instrument performance


Page 13 of 30


256

average of 0.90. For these reasons, we did not correct the BC measurements. More detail is

257

provided in the SI.

258

3.2

259

The average number of visits per fixed point was 5.1, i.e. an average of 102 minutes of sampling.

260

Concerning the cycling campaign, the mean number of visits to each road segment was 6, the

261

minimum and maximum were 2 and 15 respectively, corresponding to an average of 121 seconds

262

per segment (0.8 sec/m).

Descriptive statistics

263

The UFP and BC measurements at the fixed points were similar to the measurements by

264

bike (FIGURE S5), and the distributions were all approximately lognormal. The median UFP

265

concentrations were 20,867 and 19,756 part/cm3, and the median BC concentrations were 1,155

266

and 1,223 ng/m3 for the fixed points and cycling campaigns, respectively. Interquartile ranges

267

were comparable as well, but the extreme values were higher for the cycling records. The

268

maximum UFP concentration for a fixed point was 61,255 part/cm3, while it was 241,547

269

part/cm3 for a road segment (averages across all visits). Similarly, the maximum BC

270

concentration for a fixed point was 7,268 ng/m3, and 39,152 ng/m3 for a road segment.

271

We compared average UFP and BC concentrations at fixed points and on the portions of

272

road segments within buffers of 100m around the fixed points (FIGURES S6 and S7). The

273

median levels on road segments were in general higher than at fixed points. The median UFP

274

concentrations at fixed points and on road segments were 23,556 and 28,304 part/cm3, while BC

275

levels were 1,148 and 1,628 ng/m3. Pearson correlation coefficients between measurements at

276

fixed points and on road segments within a 100m buffer around them (45 fixed points and an

277

average of 212 m of road segments sampled around each point) were 0.07 and -0.11 for UFP and

278

BC, indicating no correlation. 13 ACS Paragon Plus Environment


Page 14 of 30

279

3.3

280

Using the meteorological variables to adjust for temporal variability proved to be more effective

281

than using the data from the fixed monitoring station, with higher adjusted R2 and lower RMSE.

282

We opted for the forward procedure for all models (more details in the SI).

LUR models

283

TABLE 1 presents the LUR models. For UFP, the two models (fixed vs. mobile

284

monitoring) present similar adjusted R2 (0.405 and 0.430 for the fixed point and mobile models).

285

Including the proportion of trucks in the fixed point model increased the adjusted R2 from 0.405

286

to 0.467. The adjusted R2 differed more distinctly between the BC fixed point and cycling

287

models (0.53 and 0.432 respectively). However, the proportion of trucks did not have an

288

important effect on the fixed point BC model as the adjusted R2 increased from 0.525 to 0.532.

289

There are few common variables between the fixed point and cycling models for the two

290

pollutants. Temperature and relative humidity are positively associated with UFP or BC, while

291

wind speed is negatively associated. Several times, the buffer size employed is different between

292

two models. For instance, traffic count is present in both BC models, but within a 750m buffer in

293

the fixed point model and within a 100m buffer in the cycling model. This is also the case for the

294

number of bus stops, the length of major roads, and the area of water. All coefficient signs are

295

consistent among the four models, except for the area of water. More small-scale traffic variables

296

(i.e. distances or predictors within buffers smaller than 500m) are included in the cycling models.

297

The results of the cross validations are available in the SI. The cycling models developed

298

with the road segments within 100m around fixed points were different from the fixed point

299

models (TABLE S16). This is not surprising as the measurements at fixed points and on the

300

surrounding road segments were not correlated.


Page 15 of 30


301

The models developed based on the fixed points were used to predict concentrations at

302

the road segments sampled during the cycling campaign to investigate transferability of the

303

variables. Inversely, the cycling models were also used to predict concentrations at the fixed

304

points. The R2 of the linear regression between the predicted and the measured concentrations

305

ranged between 0.0182 and 0.0888, indicating poor transferability of the variables and

306

coefficients.

307



308

Page 16 of 30

TABLE 1 UFP and BC LUR Models (Standardized Coefficients) for the Fixed Points and Cycling Data ln(UFP)

ln(BC)

Fixed points

Cycling

Fixed points

Cycling

Adjusted R²

0.405

0.430

0.525

0.434

Temperature

2.96E-01***

1.76E-01*

7.64E-02***

2.26E-01**

Relative Humidity Wind Speed

-1.94E-01*

-2.60E-01***

7.11E-02**

Number of bus stops within 50m 2.21E-01*

Number of bus stops within 100m Number of bus stops within 300m

1.89E-01* 2.09E-01*

Length of bus line within 100m Distance to the CBD

1.00E-01*** -2.37E-01***

3.88E-01** -4.17E-01 ***

Distance to the closest airport

-1.29E-01***

Distance to the closest major road -1.44E-01***

Distance to the shore Length of highway within 1000m

2.77E-01*

Number of intersections within 750m

3.33E-01** 8.84E-02***

Length of major roads within 200m Length of major roads within 500m

2.51E-01*

Open area within 1000m

-2.14E-01*

Population within 500m

1.77E-01.

Area of water within 50m

1.75E-01.

2.79E-01*** -1.80E-01***

Area of water within 750m

-2.79E-01*** 1.96E-01*

Commercial area within 500m 1.49E-01***

Traffic within 100m

1.57E-01*** 4.47E-01***

Traffic within 750m 1.05E-01***

Number of trees within 750m

6.93E-02***

Building footprint area within 500m 8.72E-02***

Maximum building height within 50m

-1.19E-01***

Park area within 500m

1.32E-01***

309

Length of roads within 500m Number of NPRI NOx chimneys within 200m Significance codes: 0 ‘***’, 0.001 ‘**’,0.01 ‘*’, 0.05 ‘.’

310

3.4

311

FIGURE 1 presents the four surfaces derived from the LUR models. They all look dissimilar,

312

which is confirmed by the weak correlations between them. The two BC surfaces have the

313

highest correlation, which is 0.49, but the two UFP surfaces only have a correlation coefficient of

7.29E-02***

Exposure surfaces


Page 17 of 30


314

0.232. The two cycling surfaces (for BC and UFP) and the two fixed point surfaces are

315

moderately correlated, with Pearson correlation coefficients of 0.63 and 0.49 respectively. Major

316

roads are more visible in the two cycling surfaces, in light of the variables included in the cycling

317

LUR models.

318

FIGURE 2 compares the predictions of the surfaces to the panel study data. The

319

participants spent on average 4 hours outdoors during each visit (TABLE S17), between 9am and

320

4pm. The median UFP outdoor exposures derived from the combination of GPS information and

321

exposure surfaces were different from the personal exposures: 26,344 (personal), 31,201 (cycling

322

model) and 19,057 part/cm3 (fixed point model), and similar results were obtained for BC, with

323

median exposures of 1,764 (personal), 1,799 (cycling) and 1,469 ng/m3 (fixed point). For BC, the

324

cycling surface provides closer results while the mobility-based exposure derived from the fixed

325

point surface is lower. For UFP, the cycling surface tends to overestimate the exposure while the

326

fixed point surface tends to underestimate them. In general, both mobility-based methods seem

327

to correctly capture central tendencies of exposure. However, the maximum UFP and BC levels

328

recorded via personal outdoor monitoring are higher than those derived from the surfaces,

329

because personal outdoor monitoring captures very punctual events such as a bus or a truck

330

passing by, which can drastically affect the range of observed concentrations. Additional

331

comparisons suggesting low correlations are available in the SI (FIGURES S20 and S21).

332



Page 18 of 30

(a) UFP surface derived from the LUR model based on the fixed points campaign

(b) UFP surface derived from the LUR model based on the cycling campaign


Page 19 of 30


(c) BC surface derived from the LUR model based on the fixed points campaign

333

(d) BC surface derived from the LUR model based on the cycling campaign FIGURE 1 UFP and BC exposure surfaces obtained from the LUR models based on fixed points and cycling.



Page 20 of 30

80,000 70,000

UFP (part/cm³)

60,000 50,000 40,000 30,000 20,000 10,000 0 Personal measurements

Cycling surface

Fixed points surface

(a) UFP (45 personal measurements) 6,000

5,000

BC (ng/m³)

4,000

3,000

2,000

1,000

0 Personal measurements

334 335 336 337

Cycling surface

Fixed points surface

(b) BC (62 personal measurements) FIGURE 2 Comparison of the UFP and BC personal exposures with mobility-based exposures predicted by the fixed point and cycling surfaces (the lower and upper whiskers correspond to the minimum and maximum, the box is drawn between the first and the third quartile with the middle line corresponding to the median; axes are truncated).

338


Page 21 of 30


339

4

340

In this study, we developed LUR models based on short-term mobile and stationary monitoring

341

campaigns. The mobile campaign was conducted on bicycle while the stationary campaign was

342

conducted based on short-term sidewalk measurements. The predictors captured in the LUR

343

models based on each sampling campaign are different and exposure surfaces resulting from the

344

LUR predictions are dissimilar. The adjusted R2 of the UFP LUR were 0.405 and 0.430 for the

345

fixed point and cycling models, and 0.525 and 0.434 for the BC fixed point and cycling models,

346

respectively. These values fall within the range of R2 presented in the literature. TABLES S20

347

and S21 summarise the characteristics of various UFP and BC short-term monitoring campaigns

348

and LUR models published in the recent literature. For UFP, values for the R2 range between

349

0.23 and 0.85 for LUR models based on fixed-site monitoring, while varying between 0.13 and

350

0.72 for LUR models based on mobile monitoring. For BC, they range between 0.28 and 0.86 for

351

fixed-site monitoring and 0.12 and 0.68 for mobile monitoring. Based on the literature, it seems

352

that fixed monitoring generally leads to higher LUR R2 than mobile monitoring for BC (TABLE

353

S21).

DISCUSSION

354

A review of recent LUR models developed for various cities (TABLES S20 and S21)

355

illustrates that the number of fixed sites (between less than 206 and 16011 for a single city) and

356

the duration of sampling (between 15 minutes5 and 90 minutes8) vary significantly across

357

studies. Moreover, the aggregation of measurements in the case of mobile campaigns varies

358

significantly, as the unit of analysis ranges from a road segment, extending between two

359

intersections, to segments of equal length. The number of aggregation locations varies between

360

about 10012 and more than 4,00023, depending on the extent of the area covered. The average

361

time spent at each location has been reported to be as low as 18 seconds9 and reaching more than



Page 22 of 30

362

10 minutes18. During the cycling campaign, we spent on average 121 seconds on each road

363

segment, which is within the range of time reported in the literature. During our stationary

364

campaign, each fixed point was visited on average 102 minutes, comparable with the range of 15

365

to 90 minutes typical of most fixed monitoring campaigns.

366

In our study, extreme UFP and BC values measured on bicycles were more enhanced

367

than those measured at fixed points, which is related to the time spent at each location. At fixed

368

points, the average concentrations were almost always calculated based on 100 minutes of

369

sampling, whereas the averages of road segments sampled while cycling were computed based

370

on a very short duration, and were therefore highly influenced by the surrounding environment

371

(e.g. bus, construction work) when sampling. Comparing more specifically the concentrations on

372

the road segments surrounding the fixed points provides the same conclusion (FIGURES S6 and

373

S7). The mean concentrations are higher on road segments, and the maximum values are also

374

high.

375

Our findings are consistent with the results of Kerckhoffs et al.9 who reported cycling

376

measurements on average 1.7 and 2.1 times higher than stationary measurements for UFP and

377

BC respectively. However, the correlation between stationary and mobile measurements reported

378

by Kerckhoffs et al.9 (0.48) was higher than the one observed in this study. In the study of

379

Kerckhoffs et al.9, short-term stationary monitoring was performed in between mobile

380

measurements. Therefore, the comparison was done between measurements conducted very

381

close in time and in space. In light of the low correlation between stationary and mobile

382

measurements observed in our study, we can conclude that the data collection protocol in itself

383

(i.e. cycling on road vs standing on sidewalk) has a large influence on the data recorded, beyond


Page 23 of 30


384

the averaging time. Few stationary measurements are typically performed in cars on the road side

385

while standing on the curbside is more commonly observed6,11.

386

Few variables are common to the fixed point and cycling models for both pollutants,

387

suggesting that measurements on roads and at fixed points do not capture the same

388

characteristics. The signs associated with some variables are worth noting. Although temperature

389

is generally negatively associated with UFP and BC concentrations14,24, it presents a positive

390

coefficient in all our models. We can explain the positive association between temperature and

391

UFP concentrations by nucleation events occurring when temperature is high, which is supported

392

by the peaks in UFP concentrations generally recorded at the beginning of the afternoon.

393

Distances to the shore and to the CBD appear in the cycling and fixed point UFP models with

394

negative and positive signs respectively. Several highways run along the shore, and two cycling

395

routes were designed in particular to capture this effect. On the other hand, trucks, which were

396

positively associated with UFP levels in the fixed point model, do not usually drive downtown

397

but mainly on highways located north of the city or along the shore and far from downtown,

398

explaining the positive association between UFP levels and distance to the CBD in the fixed

399

point model. Another interesting variable is the water area, which has a positive coefficient in the

400

UFP fixed point model but a negative coefficient in the UFP and BC cycling models, a

401

difference probably due to the distinct buffer sizes included in these models. Finally, in the

402

cycling model, the number of trees is positively associated with UFP concentrations. Several

403

studies have showed that vegetation could be positively or negatively associated with air

404

pollutant concentrations25–27. Indeed, the characteristics of the street (i.e. building height, street

405

width) combined with the type of vegetation (e.g. canopy density, crown diameter, trunk height)



Page 24 of 30

406

can have a significant influence on the pollutants levels, by either increasing the turbulence or

407

decreasing the wind speed.

408

Weichenthal et al.15 and Sabaliauskas et al.12 developed LUR models for UFP in Toronto

409

based on measurements conducted in 2015. Many predictors included in the model from

410

Weichenthal et al.15 were related to small-scale traffic characteristics (i.e. distance to the

411

highway, distance to the major road, distance to the bus line, number of intersections, length of

412

bus line), while distance to the airport was also an important predictor. Both our UFP models,

413

and more specifically the cycling model, also include small-scale traffic variables, and distance

414

to the airport is a predictor in the cycling model. The predictors included in Sabaliauskas et al.12

415

model were related to the land use (e.g. population, industrial and residential areas), also

416

captured in our UFP LUR models via predictors such as open area, population, or water area.

417

All four surfaces are dissimilar, which raises concerns regarding their use for

418

epidemiologic studies. Major roads are highlighted in both cycling surfaces, because air pollutant

419

levels were collected directly on roads and were therefore affected by small-scale traffic

420

variables. In contrast, the fixed point surfaces are capturing hotspots at a larger spatial scale,

421

which is probably related to the nature of the monitoring approach, i.e. stationary locations being

422

more spaced. The models based on short-term stationary and mobile approaches developed by

423

Kerckhoffs et al.9 were similar in terms of predictors, and the concentrations they estimated

424

using their two models at 12,682 residential addresses were also similar. These observations are

425

different from the outcomes of our study, which highlights sensitivity of exposure surfaces to the

426

sampling protocol. In the study of Kerckhoffs et al.9, stationary measurements were carried out

427

in between mobile measurements at the same distance from traffic since a car was used for both

428

protocols. This is not the case for our study, since our study is not only capturing the difference


Page 25 of 30


429

between mobile and stationary measurements but also the location of where mobile and

430

stationary sampling would normally occur: on the road vs. on the sidewalk. Together, these two

431

studies begin to unveil the complexity in the development of LUR surfaces for traffic-related air

432

pollution, highlighting the importance of testing their sensitivity prior to use in epidemiological

433

studies.

434

Our comparison with data from outdoor personal exposures demonstrated that using two

435

different surfaces can lead to different values for daily exposure. Our results suggest that the

436

mobile monitoring campaign tends to overestimate short-term exposures to UFP, while using the

437

fixed point surface tends to underestimate the short-term exposures to both UFP and BC. For

438

BC, the cycling surface was able to capture the distribution of outdoor personal exposures. For

439

UFP, despite over-predicting the mean exposure, the cycling surface was able to better capture

440

the distribution. Mobile sampling enables the coverage of wide areas, and we took advantage of

441

this when designing our campaign as we sampled 270 km of unique roads. However, as the

442

measurements are done on a small time scale, they are more susceptible to be influenced by

443

short-term events. When comparing short-term exposures, capturing these short-term peaks is

444

important. It is worth noting that in our study, personal exposures were collected between 9am

445

and 4pm, which could have an impact on the comparison with exposures estimated based on the

446

surfaces as the monitoring campaigns took place between 7am and 7pm.

447

The development of exposure surfaces using data collected during short-term sampling

448

campaigns have gained momentum in recent years, primarily in light of advances in air pollution

449

monitoring technology. Small portable instruments have enabled the development of data

450

collection protocols that involve mobile sampling by foot, by bike, or by cars fitted with various

451

sensing devices. The development of inexpensive air pollution sensors will further enable these



Page 26 of 30

452

protocols and might even lead to ad-hoc data collection means through participatory air pollution

453

sensing. The question of robustness of LUR surfaces will become of utmost importance since the

454

streams of data collected are expected to become more diverse, including various locations in an

455

urban area: on-road, on the sidewalks, or at residential locations. There is a need to further

456

investigate the effect of the data collection protocol on exposure surfaces and evaluate the

457

sensitivity of LUR models.

458 459

ACKNOWLEDGMENTS

460

The authors thank Ryan Kulka, Xinli Tu, Marie-Sophie Wint, Kuruparam Satkunanathan,

461

Tszchun Chow, Tian Kang, James Mario and Teresa Li for their great help in collecting data.

462

SUPPORTING INFORMATION

463

The SI contains information about 1) the design of the campaigns, 2) the LUR predictors, 3) the

464

correction of the monitors, 4) the data collected, 5) the choice of the models, 6) the comparison

465

of the exposures, 7) the characteristics of various UFP and BC monitoring campaigns and LUR

466

models.

467

REFERENCES

468

(1)

Aguilera, I.; Dratva, J.; Caviezel, S.; Burdet, L.; De Groot, E.; Ducret-Stich, R. E.;

469

Eeftens, M.; Keidel, D.; Meier, R.; Perez, L.; et al. Particulate matter and subclinical

470

atherosclerosis: Associations between different particle sizes and sources with carotid

471

intima-media thickness in the SAPALDIA study. Environ. Health Perspect. 2016, 124

472

(11), 1700–1706.

473

(2)

Clark, N. A.; Demers, P. A.; Karr, C. J.; Koehoorn, M.; Lencar, C.; Tamburic, L.; Brauer, 26 ACS Paragon Plus Environment

Page 27 of 30


474

M. Effect of early life exposure to air pollution on development of childhood asthma.

475

Environ. Health Perspect. 2010, 118 (2), 284.

476

(3)

Sahsuvaroglu, T.; Jerrett, M.; Sears, M. R.; McConnell, R.; Finkelstein, N.; Arain, A.;

477

Newbold, B.; Burnett, R. Spatial analysis of air pollution and childhood asthma in

478

Hamilton, Canada: comparing exposure methods in sensitive subgroups. Environ. Heal.

479

2009, 8 (1), 14.

480

(4)

Crouse, D. L.; Peters, P. A.; Hystad, P.; Brook, J. R.; van Donkelaar, A.; Martin, R. V.;

481

Villeneuve, P. J.; Jerrett, M.; Goldberg, M. S.; Arden Pope, C.; et al. Ambient PM2.5, O3,

482

and NO2 exposures and associations with mortality over 16 years of follow-up in the

483

canadian census health and environment cohort (CanCHEC). Environ. Health Perspect.

484

2015, 123 (11), 1180–1186.

485

(5)

Rivera, M.; Basagaña, X.; Aguilera, I.; Agis, D.; Bouso, L.; Foraster, M.; Medina-Ramón,

486

M.; Pey, J.; Künzli, N.; Hoek, G. Spatial distribution of ultrafine particles in urban

487

settings: A land use regression model. Atmos. Environ. 2012, 54, 657–666.

488

(6)

Saraswat, A.; Apte, J. S.; Kandlikar, M.; Brauer, M.; Henderson, S. B.; Marshall, J. D.

489

Spatiotemporal Land Use Regression Models of Fine, Ultrafine, and Black Carbon

490

Particulate Matter in New Delhi, India. Environ. Sci. Technol. 2013, 47 (22), 12903–

491

12911.

492

(7)

Ghassoun, Y.; Ruths, M.; Löwner, M. O.; Weber, S. Intra-urban variation of ultrafine

493

particles as evaluated by process related land use and pollutant driven regression

494

modelling. Sci. Total Environ. 2015, 536, 150–160.

495 496

(8)

Montagne, D. R.; Hoek, G.; Klompmaker, J. O.; Wang, M.; Meliefste, K.; Brunekreef, B. Land Use Regression Models for Ultrafine Particles and Black Carbon Based on Short-



Page 28 of 30

497

Term Monitoring Predict Past Spatial Variation. Environ. Sci. Technol. 2015, 49 (14),

498

8712–8720.

499

(9)

Kerckhoffs, J.; Hoek, G.; Messier, K. P.; Brunekreef, B.; Meliefste, K.; Klompmaker, J.

500

O.; Vermeulen, R. Comparison of Ultrafine Particle and Black Carbon Concentration

501

Predictions from a Mobile and Short-Term Stationary Land-Use Regression Model.

502

Environ. Sci. Technol. 2016, 50 (23), 12894–12902.

503

(10)

Wolf, K.; Cyrys, J.; Harciníková, T.; Gu, J.; Kusch, T.; Hampel, R.; Schneider, A.; Peters,

504

A. Land use regression modeling of ultrafine particles, ozone, nitrogen oxides and

505

markers of particulate matter pollution in Augsburg, Germany. Sci. Total Environ. 2017,

506

579, 1531–1540.

507

(11)

van Nunen, E.; Vermeulen, R.; Tsai, M.-Y.; Probst-Hensch, N.; Ineichen, A.; Davey, M.;

508

Imboden, M.; Ducret-Stich, R.; Naccarati, A.; Raffaele, D.; et al. Land Use Regression

509

Models for Ultrafine Particles in Six European Areas. Environ. Sci. Technol. 2017, 51 (6),

510

3336–3345.

511

(12)

Sabaliauskas, K.; Jeong, C. H.; Yao, X.; Reali, C.; Sun, T.; Evans, G. J. Development of a

512

land-use regression model for ultrafine particles in Toronto, Canada. Atmos. Environ.

513

2015, 110, 84–92.

514

(13)

Patton, A. P.; Zamore, W.; Naumova, E. N.; Levy, J. I.; Brugge, D.; Durant, J. L.

515

Transferability and generalizability of regression models of ultrafine particles in urban

516

neighborhoods in the boston area. Environ. Sci. Technol. 2015, 49 (10), 6051–6060.

517

(14)

Farrell, W.; Weichenthal, S.; Goldberg, M.; Valois, M. F.; Shekarrizfard, M.;

518

Hatzopoulou, M. Near roadway air pollution across a spatially extensive road and cycling

519

network. Environ. Pollut. 2016, 212, 498–507.


Page 29 of 30

520


(15)

Weichenthal, S.; Van Ryswyk, K.; Goldstein, A.; Shekarrizfard, M.; Hatzopoulou, M.

521

Characterizing the spatial distribution of ambient ultrafine particles in Toronto, Canada: A

522

land use regression model. Environ. Pollut. 2016, 208, 241–248.

523

(16)

Weichenthal, S.; Ryswyk, K. Van; Goldstein, A.; Bagg, S.; Shekkarizfard, M.;

524

Hatzopoulou, M. A land use regression model for ambient ultrafine particles in Montreal,

525

Canada: A comparison of linear regression and a machine learning approach. Environ.

526

Res. 2016, 146, 65–72.

527

(17)

Hankey, S.; Marshall, J. D. Land Use Regression Models of On-Road Particulate Air

528

Pollution (Particle Number, Black Carbon, PM2.5, Particle Size) Using Mobile

529

Monitoring. Environ. Sci. Technol. 2015, 49 (15), 9194–9202.

530

(18)

Larson, T.; Henderson, S. B.; Brauer, M. Mobile monitoring of particle light absorption

531

coefficient in an urban area as a basis for land use regression. Environ. Sci. Technol. 2009,

532

43 (13), 4672–4678.

533

(19)

Li, H. Z.; Dallmann, T. R.; Gu, P.; Presto, A. A. Application of mobile sampling to

534

investigate spatial variation in fine particle composition. Atmos. Environ. 2016, 142, 71–

535

82.

536

(20)

Klompmaker, J. O.; Montagne, D. R.; Meliefste, K.; Hoek, G.; Brunekreef, B. Spatial

537

variation of ultrafine particles and black carbon in two cities: Results from a short-term

538

measurement campaign. Sci. Total Environ. 2015, 508, 266–275.

539

(21)

Hagler, G. S. W.; Yelverton, T. L. B.; Vedantham, R.; Hansen, A. D. A.; Turner, J. R.

540

Post-processing method to reduce noise while preserving high time resolution in

541

aethalometer real-time black carbon data. Aerosol Air Qual. Res. 2011, 11 (5), 539–546.

542

(22)

Eeftens, M.; Beelen, R.; De Hoogh, K.; Bellander, T.; Cesaroni, G.; Cirach, M.; Declercq,



Page 30 of 30

543

C.; Dedele, A.; Dons, E.; De Nazelle, A.; et al. Development of land use regression

544

models for PM2.5, PM 2.5 absorbance, PM10 and PMcoarse in 20 European study areas;

545

Results of the ESCAPE project. Environ. Sci. Technol. 2012, 46 (20), 11195–11205.

546

(23)

Farrell, W. J.; Deville Cavellin, L.; Weichenthal, S.; Goldberg, M.; Hatzopoulou, M.

547

Capturing the urban canyon effect on particle number concentrations across a large road

548

network using spatial analysis tools. Build. Environ. 2015, 92, 328–334.

549

(24)

Hatzopoulou, M.; Weichenthal, S.; Dugum, H.; Pickett, G.; Miranda-Moreno, L.; Kulka,

550

R.; Andersen, R.; Goldberg, M. The impact of traffic volume, composition, and road

551

geometry on personal air pollution exposures among cyclists in Montreal, Canada. J.

552

Expo. Sci. Environ. Epidemiol. 2013, 23 (1), 46–51.

553

(25)

554 555

Buccolieri, R.; Gromke, C.; Di Sabatino, S.; Ruck, B. Aerodynamic effects of trees on pollutant concentration in street canyons. Sci. Total Environ. 2009, 407 (19), 5247–5256.

(26)

Gromke, C.; Ruck, B. Influence of trees on the dispersion of pollutants in an urban street

556

canyon-Experimental investigation of the flow and concentration field. Atmos. Environ.

557

2007, 41 (16), 3287–3302.

558 559

(27)

Jin, S.; Guo, J.; Wheeler, S.; Kan, L.; Che, S. Evaluation of impacts of trees on PM 2 . 5 dispersion in urban streets. Atmos. Environ. 2014, 99, 277–287.

560


development and comparison of air pollution exposure surfaces

Recommend Documents