Using Mobile Monitoring to Develop Hourly Empirical Models of

Subscriber access provided by UNIV OF TEXAS DALLAS

Environmental Modeling

Using Mobile Monitoring to Develop Hourly Empirical Models of Particulate Air Pollution in a Rural Appalachian Community Steve Hankey, Peter Sforza, and Matt Pierson Environ. Sci. Technol., Just Accepted Manuscript • Publication Date (Web): 15 Mar 2019 Downloaded from http://pubs.acs.org on March 15, 2019

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.

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 29

Environmental Science & Technology

Using Mobile Monitoring to Develop Hourly Empirical Models of Particulate Air Pollution in a Rural Appalachian Community

Steve Hankey,1* Peter Sforza,2 Matt Pierson2 1School

of Public and International Affairs, Virginia Tech, 140 Otey Street, Blacksburg, VA

24061 2Center

for Geospatial Information Technology, Virginia Tech, 620 Drillfield Drive, Blacksburg, VA 24061 *Corresponding Author: email: [email protected]; phone: 540.231.7508. Word count: 4,874 + 300×4 small tables/figures + 600×2 large tables/figures = 7,274 Submitted: March 13, 2019

KEYWORDS traffic-related air pollution; ultrafine particles; particle number concentration; land use regression

1 ACS Paragon Plus Environment


1

ABSTRACT

2

Most empirical air quality models (e.g., land use regression) focus on urban areas.

3

Mobile monitoring for model development offers the opportunity to explore smaller, rural

4

communities – an understudied population. We use mobile monitoring to systematically sample

5

all daylight hours (7am-7pm) to develop empirical models capable of estimating hourly

6

concentrations in Blacksburg, VA – a small town in rural Appalachia (population: 182,635). We

7

collected ~120 hours of mobile monitoring data for Particle Number (PN) and Black Carbon

8

(BC). We developed (1) daytime (12-hour average) models that approximate long-term

9

concentrations and (2) spatiotemporal models for estimating hourly concentrations. Model

10

performance for the daytime models is consistent with previous fixed-site and short-term

11

sampling studies; adjusted R2 (10-fold CV R2) was 0.80 (0.69) for the PN model and 0.67 (0.58)

12

for the BC model. The spatiotemporal models had comparable performance (10-fold CV R2 for

13

the PN [BC] models: 0.42 [0.25]) to previous mobile monitoring studies that isolate specific time

14

periods. Temporal and spatial model coefficients had similar magnitudes in the spatiotemporal

15

models suggesting both factors are important for exposure. We observed similar spatial patterns

16

in Blacksburg (e.g., roadway gradients) as in other studies in urban areas suggesting similar

17

exposure disparities exist in small, rural communities.

18

TOC Art

19 2 ACS Paragon Plus Environment

Page 2 of 29

Page 3 of 29

20 21


INTRODUCTION Empirical models of air quality have traditionally employed fixed-site monitoring to

22

estimate annual-average concentrations for model building 1-2. Air quality researchers have

23

primarily focused on improving the spatial precision of empirical models of air quality (e.g., land

24

use regression [LUR]) to estimate long-term average concentrations 3-5 for use in health effects

25

studies 6-7. Recent work has added temporal resolution by modeling monthly 8-9 or daily 10

26

average concentrations. Emerging work has explored use of mobile monitoring to develop

27

models with hourly resolution 11. Most of these empirical models focus on urban areas. The US

28

Census estimates that ~29 million people (or ~9% of the US population) live in 199 Metropolitan

29

Statistical Areas (MSA) of < 250,000 people (~50% of the total MSAs) 12. An understudied topic

30

is development of empirical models in these relatively smaller communities where little air

31

quality monitoring data is available. More work is needed to assess whether trends from previous

32

empirical modeling studies in urban areas are consistent with or differ for smaller communities.

33

An increasing number of studies use mobile monitoring to characterize specific traffic

34

corridors 13-15, develop and assess empirical models 16-21, and apply those models for

35

epidemiological studies 22. Use of mobile monitoring may be particularly useful for pollutants

36

with high spatial and temporal variability 23. Some studies have compared empirical models

37

developed using mobile monitoring vs. fixed-site monitoring 24-27 or assessed merging mobile

38

monitoring with fixed-site monitoring to develop spatiotemporal models 28; however, results of

39

these comparisons vary by study and the amount of mobile and fixed-site data collected. Since

40

mobile monitoring is typically conducted across many days (and various times of day), previous

41

studies have used a variety of methods to adjust mobile monitoring data for background



42

concentrations 11, 16, 27, 29. An understudied topic is how the choice of background adjustment

43

method impacts model performance and the spatial patterns of modeled concentration estimates.

44

Accurately assessing human exposure to air pollution depends on correctly estimating the

45

spatiotemporal patterns of pollutant concentrations to match people’s time-activity patterns 30-33.

46

In this study, we present results of spatiotemporal empirical models of particulate air pollution

47

for the rural Appalachian community of Blacksburg, VA based on mobile monitoring of Particle

48

Number (PN) and Black Carbon (BC) concentrations – pollutants that have high spatial

49

variability in urban environments 34-38 and are associated with health disparities at small spatial

50

scales 39-42. We develop a number of empirical models to test how inclusion of temporal

51

variables (e.g., meteorology, time of day) and type of background concentration correction

52

impact model performance. We focus on comparing how model-derived short-term (hourly)

53

concentration estimates compare to long-term (daily) estimates to assess how choice of spatial or

54

spatiotemporal modeling approach impacts exposure assessment.

55

MATERIALS AND METHODS

56

Our work employs mobile monitoring during the daylight hours (7am-7pm) of the

57

summer and fall of year-2016 for the purpose of developing empirical models of particulate air

58

pollution with hourly time resolution. Our study area is a small rural town in the Blue Ridge

59

Mountains (Blacksburg, VA; MSA population: 182,635).

60

Mobile monitoring data

61

Mobile monitoring platform

62

We developed a custom mobile monitoring platform to collect spatially and temporally

63

resolved measurements of PN and BC (Figure 1). The platform consists of (1) a CPC 3007 (TSI,

64

Inc.) to measure PN, (2) a micro-aethalometer (AE51; AethLabs) to measure BC, (3) a


Page 4 of 29

Page 5 of 29


65

Raspberry Pi and GPS antenna, and (4) an external battery. The platform logs and geo-locates

66

measurements every second. The Raspberry Pi includes a WiFi antenna to transfer all monitoring

67

data to a server when a WiFi signal is present. All post-processing Python scripts are run on a

68

server prior to making the data available for download on a website (details on the procedures

69

included in the post-processing script is below) which allowed for viewing time series and

70

mapped data for each monitoring run. (The Python script is available upon request from the

71

corresponding author.) The platform is mounted to a bicycle rack on an electric assist bicycle for

72

mobile monitoring; an elevated air intake was used to sample at approximately the breathing

73

height of a cyclist or pedestrian. The monitoring platform is in Figure 1; further details are in

74

Figures S1-S3.

75 76 77

[Insert Figure 1 about here] Mobile monitoring campaign We selected two mobile monitoring routes with the goal of spanning a variety of

78

environments that may impact PN and BC concentrations (Figure S4). Each route was ~20

79

kilometers in length and could be completed in ~45-60 minutes. We chose the sampling routes to

80

sample various road types (major and minor roads, off-street trails), land uses (residential,

81

commercial, natural environments), and emission sources (heavy- and light-duty vehicle

82

volumes).

83

Core goals of our work are (1) to compare short- and long-term concentration estimates

84

of empirical models developed using mobile monitoring data, (2) test how inclusion of various

85

meteorological variables and background concentration adjustments impact model performance,

86

and (3) explore spatial trends in a small, rural Appalachian town. We stratified mobile

87

monitoring by hour of day. Specifically, we completed each sampling route a minimum of 5



88

times per hour of day from 7am-7pm (Figure S5) resulting in ~10 hours (2 routes × 5 sampling

89

runs × 1 hour per sampling run) of mobile monitoring data per hour of the day for model

90

building. In total, we collected ~120 hours (12 hours of day × 10 hours of monitoring data per

91

hour) of data during the mobile monitoring campaign.

92

Central site measurements and monitor calibration

93

We monitored PN (CPC 3783; TSI, Inc.) and BC (AE51; Aethlabs) at a central site for

94

the entire monitoring campaign (7/20/2016 – 10/23/2016). We added a particle size selector

95

(TSI, Inc.) to the CPC 3783 (cutoff size: 7 nm) to match the 10 nm cutoff size of the CPC 3007.

96

The central site was located where all mobile monitoring routes began and finished (Figure S6);

97

measurements were logged in 1-minute intervals. We used the central site measurements to

98

correct all mobile measurements for variation in background concentrations (see below). We co-

99

located all monitoring equipment at a near-road location for 22 hours (among 4 days) before and

100

14 hours (among 3 days) after the mobile monitoring campaign. We used data from the co-

101

located monitoring to correct all measurements to a reference CPC and micro-aethalometer.

102

Figures S7-S12 and Table S1 give further details on the results of the co-location analysis.

103

Processing measurement data for spatial modeling

104

Below we briefly describe data processing for the mobile monitoring data (further detail

105

is in the SI).

106

Measurement corrections and data processing

107

Monitor co-location corrections. Based on the co-location measurements, we developed

108

instrument-specific correction factors to adjust all mobile monitoring data to a reference CPC

109

and micro-aethalometer. The corrections resulted in an average correction factor of 1.09 for the


Page 6 of 29

Page 7 of 29


110

CPCs and 0.98 for the micro-aethalometers. We applied a time delay correction as observed in

111

our previous mobile monitoring study with a similar monitoring platform 16, 43 (Table S2).

112

Micro-aethalometer corrections. We corrected all BC measurements for two known

113

artifacts of sampling: (1) particle loading on the micro-aethalometer filter and (2) spurious data

114

associated with mechanical shock. We followed Kirchstetter and Novakov 44 to adjust for

115

particle loading and Apte et al. 45 to identify and remove spurious data from mechanical shock.

116

This procedure resulted in removal of 10% of the mobile monitoring data. We collected BC

117

measurements on a 1 second basis but temporally smoothed the data on a 60s basis due to

118

inherent noise associated with the micro-aethalometer.

119

CPC corrections. We corrected PN measurements from the CPC 3007 for undercounting

120

at high concentrations (>100,000 pt/cc). We used the correction equation developed by

121

Westerdahl et al. 46. Our near-road, co-located measurements demonstrated a similar relationship

122

as reported by Westerdahl et al. 46 (Figure S12); we chose to apply the previously published

123

correction equation since those measurements were collected in a variety of locations rather than

124

a single near-road location.

125

Background adjustments. To test the impact of day-to-day variability in background

126

concentrations we developed models using three scenarios (1) mobile monitoring data

127

unadjusted for background concentrations, (2) an additive background adjustment similar to

128

Hankey and Marshall 16, and (3) a multiplicative background adjustment as described in Apte et

129

al. 29.

130

For the additive adjustment, we first applied a 60-minute temporal smoothing to the

131

central site data. Then, we subtracted the background concentration estimate from each

132

corresponding mobile monitoring observation (leaving only the local, on-road concentration



133

component of each measurement). Upon completion of the entire monitoring campaign we

134

calculated an average hourly background concentration and added that value to each on-road

135

concentration estimate for the corresponding hour for use in model building. This background

136

adjustment aims to adjust concentration estimates to reflect the average temporal background

137

concentration trends during our study period which may better reflect typical patterns rather than

138

trends specific to days where mobile monitoring data was collected.

139

For the multiplicative adjustment, we developed correction factors with more temporal

140

resolution than the additive adjustment. First, we calculated the median concentration for each

141

mobile monitoring day for all daylight hours (7am-7pm) at the central site. We also calculated

142

the median central site concentration for each hour of day. The correction factor for each mobile

143

monitoring observation was calculated as the ratio of the daily concentration to the hourly

144

concentration. This correction has superior temporal resolution to the additive background

145

adjustment; however, it is likely best suited for estimating daily concentrations since the

146

correction adjusts all mobile monitoring observations to the daily median value. Distributions of

147

unadjusted and background adjusted concentrations are shown in Figures S13-S15.

148

Spatial aggregation

149

We spatially aggregated the mobile monitoring data as the core input to the empirical

150

models described below. We first snapped all mobile measurements to aggregation points

151

created at 100 meter intervals along the monitoring routes. Then, we tabulated median

152

concentrations at each aggregation location to minimize the impact of outlier values and to

153

follow on our previous work 16. We performed this aggregation for all mobile monitoring data

154

during all daylight hours (7am-7pm; ~931 observations per aggregation location) and for each


Page 8 of 29

Page 9 of 29


155

individual hour of day (~77 observations per aggregation location). This procedure resulted in

156

423 locations with concentration estimates for model development (Table S3; Figure S16).

157

Database of covariates for spatial modeling

158

We tabulated a variety of candidate covariates that measure land use, the transportation

159

network, and the natural environment (Table 1). All variables were tabulated at 15 buffer sizes

160

(25 – 3,000 meters) and were included in the model selection process. All underlying data was

161

from publicly available sources with the exception of traffic volumes. To estimate traffic

162

volumes across the network we used existing traffic counts to develop statistical models to

163

estimate annual-average daily traffic for the entire network; we developed separate models for

164

heavy- and light-duty vehicles (Tables S4-S5). We also included hour of day (as a dummy

165

variable) and 5 meteorological parameters (temperature, relative humidity, planetary boundary

166

layer, precipitation, wind speed) in the spatiotemporal models.

167

[Insert Table 1 about here]

168

Modeling approach

169

We developed three sets of models to compare short- and long-term concentration

170

estimates from empirical models developed using mobile monitoring: (1) daytime models which

171

aggregate all mobile monitoring data to approximate traditional, fixed-site empirical models

172

(e.g., LUR), (2) single-hour models which develop a single model for each hour of the day, and

173

(3) spatiotemporal models that include time of day and meteorological parameters as candidate

174

variables for selection in a single model to estimate hourly concentrations.

175

We employed the stepwise regression approach originally developed by Su et al. 47 and

176

applied to mobile monitoring data by Hankey and Marshall 16 using MATLAB R2017b. Briefly,

177

model selection proceeds by adding variables most correlated with the dependent variables (PN



178

and BC concentrations), and subsequently most correlated with model residuals. The model

179

building process is terminated when a variable is either not significant in the model (p

Using Mobile Monitoring to Develop Hourly Empirical Models of

Recommend Documents