Introduction

The public health consequences of ambient particulate matter (PM) have been well-documented by more than three decades of epidemiologic, toxicologic and clinical studies. The global burden of disease attributable to ambient PM is at a historical high and is estimated to be greater than 5 million deaths per year, including an estimated 108,000 deaths in the United States [1]. In addition to increases in mortality, air pollution exposure is associated with adverse cardiovascular, respiratory and neurological outcomes and diseases. Cardiovascular effects identified through epidemiological studies include myocardial infarction, atherosclerosis, and heart rate variability [2]. Respiratory effects associated with exposure include pulmonary infections, cough, wheeze and chronic obstructive pulmonary disease (COPD). There is sufficient evidence of a causal association with asthma incidence and severity among children [3]. Recent evidence has shown that PM impacts the neurological system triggering the development of Alzheimer’s disease and autism in adults and children, respectively [4, 5]. Sensitive subpopulations, due to existing morbidities or marginalization by race, class or income, may be at increased levels of susceptibility compared to the general population [6–8].

Although air quality has improved in recent decades in the United States, there is substantial variation in exposure based on proximity to sources. Hot spots of exposure are present close to highways and busy roadways [9]. Urban air pollution is frequently dominated by traffic-related air pollution (TRAP) produced by the burning of diesel and gasoline fuel in trucks, buses and cars. Areas near highways and major roads have elevated levels of TRAP compared to areas that are further away from these mobile sources [10–12]. The distribution and down-range gradient of PM from traffic corridors in the near-road environment is not identical for all PM size fractions [11, 13]. Concentrations of ultrafine particles (UFP: < 0.1 μm in aerodynamic diameter) exhibit the steepest gradients. UFP are often estimated in exposure studies by particle number concentration (PNC), which is an effective proxy. We will therefore utilize the PNC measurement for the remainder of this paper. In addition, elevated PNC levels can extend further distances from roadways depending on local meteorological conditions, such as the height of the planetary boundary layer and temperature. In some circumstances, elevated PM concentrations have been found up to 2600 m downwind of highways [14, 15]. Solid carbonaceous particles (black carbon, BC) also exhibit steep gradients, while fine mode PM (PM with aerodynamic diameter less than 2.5 μm, PM_2.5) shows a very shallow gradient [10].

Scientific evidence has shown that PNC and BC are also associated with adverse health effects, like emergency department visits and elevated cardiovascular risk, independent of those of PM_2.5 and PM₁₀. Neither PNC nor BC are directly regulated by the United States Environmental Protection Agency, and therefore, risk due to exposure may not be controlled by existing standards for PM [16]. Approximately 45 million people in the United States live within 300 m of a major highway, which constitutes a sizable amount of the population being at risk [17].

It is imperative to investigate specific solutions that address elevated concentrations of pollutants in high exposure transportation corridors. Green infrastructure is one mitigation measure for pollution reduction. A recent report from the World Health Organization (WHO) advised further evaluation and implementation of green infrastructure as one avenue to improve health of urban residents [18]. Green infrastructure has a broad definition that may include various forms and functions. The type of green infrastructure discussed here includes existing or planted stands of vegetation alongside roadways [19–23]. Roadside vegetation has been shown to reduce ambient PM concentrations by increasing dispersion, interfering in pollutant transport and direct interception of particles [19, 20, 24, 25]. Other studies have measured increases in concentrations on road and in adjoining areas [20]. The efficacy of vegetation in reducing PM also varies across the year [16]. These apparent discrepancies stem from the fact that vegetation are living organisms that are in constant interplay with the environment around them. Vegetation absorbs carbon dioxide and releases oxygen along with other chemicals such as biogenic volatile organic carbons (BVOCs). A portion of these chemicals are the source of naturally created PM.

Very little is known about the impact of green infrastructure on the chemical composition of particulate matter or the physiological effects of exposure. It is possible that PM that has passed through vegetation may have more benign or more harmful impacts on the human body. Combining the knowledge we currently have there are some basic characteristics that separate a ‘good’ barrier from a ‘poor’ one; however, more research is necessary to optimize a vegetative barrier for a specific location or set of conditions [21, 26].

This study was designed to provide data that fills gaps regarding the impact of vegetative barriers on near-roadway PM concentrations as well as the effect of PM on human cells. We have explored multiple sampling sites in a single metropolitan area (Atlanta, Georgia, USA) to compare and contrast roadside barrier types and meteorological conditions. A novel aspect of this study was that we utilized generalized linear mixed models (GLMMs) to estimate the effect on PM concentration of different types of roadside barriers while controlling for covariates including meteorological and environmental parameters. GLMMs are a widely-used multivariate modeling approach including both fixed and random effects to produce effect estimates of grouped data [27].

The Atlanta metropolitan area has a total population of 4.7 million people. It contains four large controlled-access interstate highways (I-20, I-75, I-85 and I-285) and a major state-level controlled-access highway (GA-400). In 2019, the regional roadway network carried 151 million vehicle miles traveled (VMT), of which 50 million VMT where on controlled-access highways. The most-traveled highway segment is the co-listed I-75/85 through downtown Atlanta with 352,000 daily VMT. All traffic data is from the Georgia Department of Transportation [28]. We describe here the findings of continuous PM concentrations from the multiple sites, while toxicological impacts of exposure are given in a separate paper.

Methods

Site selection

We selected five sample sites in the Atlanta metropolitan area. Sites are shown in Fig 1 and a summary of site characteristics is given in Table 1. All sites were adjacent to a high-volume controlled-access highway. Given that all sampling was conducted on private property abutting the public right of way, no special permits were required from the Georgia Department of Transportation. All such highways in Georgia are separated from adjacent areas by a physical barrier to prevent pedestrian or unauthorized vehicle entry. These barriers are typically chain-link fences but in urban areas may instead be a solid sheet metal barrier for noise abatement. Selected sample sites include two with vegetative barriers, one with a solid barrier, and two with only a chain-link fence. Vegetative barriers consisted of multiple tree species, typically dominated by magnolias, water oaks, loblolly pines, and crepe myrtles, and vegetative barrier sites also include a chain-link fence. Sampling occurred during the summer season with all vegetation in full foliage. Given that chain-link fences have a negligible effect on particle concentrations, we included chain-link fence sites as a control group for comparison purposes. However, it should be noted that under summer conditions in Georgia, fast-growing vine and grass species can grow along chain-link fences in between road maintenance intervals and constitute a minimal amount of vegetative coverage. We used a dual location approach to make simultaneous comparisons of pollutant concentrations at the roadside and behind the barrier at each of the five sample sites. Each sample site therefore had a “roadside” location, which was as close to the roadside as permitted by Georgia Department of Transportation rules (7–19 m from the pavement edge), and a “comparison” location, which was behind the roadside barrier. We could not quantitatively measure lateral leaf area density; however, we used surface photos and satellite imagery to qualitatively assess vegetation density in five categories ranging from no vegetation other than grass to full vegetation to a 10 m height with no visible gaps.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Aerial view of sample sites.

All images are to the same scale and oriented with north being up.

https://doi.org/10.1371/journal.pone.0296885.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Sample site characteristics.

https://doi.org/10.1371/journal.pone.0296885.t001

Results

Stratification by barrier type and by wind direction

We examined summary statistics for the change in pollutant levels for strata of sampling site and wind direction category. The summary statistics of the percent difference in PNC and BC data relating the comparison location to the roadside location stratified by sample site are shown in Fig 2. For PNC, the median percent difference including data from all sites and wind conditions was -3.0% with an interquartile range (IQR) of 18%. For BC, the median percent difference was -7.5% (IQR = 44%). In both cases, this indicates that the median concentration at the comparison location was slightly lower than at the roadside. When stratifying by site, we observed the median percent difference to be negative at all sites. The median percent difference was near zero for S2, but was 5–10% lower at all other sites, including sites with vegetative and solid fence barriers. This was an expected finding given that a key distinction for S2 is that there is only a chain-link fence at this site and both the roadside and comparison sampling locations were the same distance from the roadway (they were separated from each other by 20m parallel to the roadway). S4 also only has a chain-link fence but curiously shows a median reduction in PNC of 10%. Two features of this site are that the comparison location was at a slightly lower elevation (~1m) compared to the roadside location and that the chain-link fence in front of the comparison location was thinly covered by a fast-growing vine species that was present on some sample days but not others. Both of these features would be expected to have a modest negative effect on PNC at the comparison location. For BC, we observed a wider distribution of difference in concentration. The reduction in BC was higher at the sites with a vegetative barrier than at sites with a chain-link fence or solid barrier. The median change was 42% lower at S1, near zero for S2 (similar to PNC), 5% higher at S3, 8% lower at S4, and 11% lower at S5. The reduction was much higher at S1 than the other sites, and an important distinction for S1 is that it was adjacent to the roadway segment with the highest traffic volume of any sample site. The median difference was positive for S3, and a key distinction of this site is that it has the lowest traffic volume of all the sites, and furthermore, as it is adjacent to GA-400, which is predominantly a commuter highway with restrictions on commercial truck traffic, we observed lower BC concentrations at this site (1.4 μg·m^-3 compared to 3.0–4.6 μg·m^-3 for other sites).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

Summary statistics of the percent difference in PNC (left) and BC (right) data relating comparison locations to roadside locations stratified by sample site. Boxes are the first quartile, median and third quartile, and whiskers are the 5^th and 95^th percentiles.

https://doi.org/10.1371/journal.pone.0296885.g002

The summary statistics of this same data stratified by wind direction category are shown in Fig 3. Given that chain-link fences have a negligible effect on particle dispersion, Fig 3 does not include data from chain-link fence sites (S2 and S4). For PNC, the percent difference is negative for all categories, and the magnitude of the change is inversely proportional to the relative wind direction. In other words, when the comparison location is upwind of the highway, PNC is modestly lower at the comparison location, but when it is downwind of the highway, the reduction in concentration is more pronounced. For BC, the range of observed differences in concentration is substantially wider than for PNC. Under calm wind conditions, the median percent difference in BC concentration is slightly positive (2.9%). When the sample site is upwind of the highway, median BC concentration is 23% lower at the comparison site, whereas when the wind direction is either parallel to the highway or the sample site is downwind, the median percent difference is -39%. It should be noted that there is substantial overlap in the stratification categories in that the overwhelming majority (81%) of observations at site 3 (solid barrier) occurred under calm wind.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

Summary statistics of the percent difference in PNC (left) and BC (right) data showing the percent difference between the comparison locations and the roadside locations stratified by wind direction category. Sites with only a chain-link fence are not included (S2 and S4). Boxes are the first quartile, median and third quartile, and whiskers are the 5^th and 95^th percentiles.

https://doi.org/10.1371/journal.pone.0296885.g003

Multivariable modeling results

Table 2 shows the multivariable GLMM results. The dependent variable is the ratio of pollutant parameters at the comparison location to those at the roadside location. The backwards stepwise model selection process resulted in the final model including baseline quantity (i.e., pollutant parameter as measured at the roadside), ambient temperature, relative humidity, traffic counts, distance to the roadside sample location, distance to the comparison sample location, wind direction category, and categorical variables describing the barrier type. The barrier type categorical variables use the chain-link fence as the reference category, and the wind direction categorical variables use the calm wind conditions as the reference category. An alternative set of models used a qualitative categorical variable describing vegetation density in the place of the barrier type categorical variable. It was not possible for both variables to be included simultaneously due to their high correlation. The reference category for this variable was category 1, no vegetation other than grass.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Results of general linear mixed models (GLMM).

The dependent variable in each model is the ratio of the pollutant parameter measured at the comparison location to that measured at the roadside location. Categorical variables for wind direction use calm wind conditions as the reference category and barrier type categorical variables use the chain-link fence barrier as the reference. The first row in each box is the parameter estimate while the second and third rows are the lower and upper 95% confidence intervals (respectively).

https://doi.org/10.1371/journal.pone.0296885.t002

The parameter estimates for PNC were not statistically significant for either barrier type when controlling for other co-variates. The estimated change in the ratio of PNC at the comparison site to the roadside site was slightly positive (β = 0.080) for the vegetative barrier using a chain-link fence as the reference, though the 95% confidence intervals include zero (p = 0.234). For solid barriers, the estimated change in the PNC ratio was slightly negative (β = -0.072) though also not significant (p = 0.325). The PNC model estimates for wind direction are consistent with numerous previous findings of PNC being reduced with distance downwind from highways [20, 30]. In other words, using calm wind conditions as the reference, the parameter estimates for wind direction categorical variables indicate the greatest statistically significant reduction in PNC ratio when the sample site is downwind of the highway (β = -0.19, p = 0.00075), a smaller reduction when the wind is parallel to the highway (β = -0.13, p = 0.0063), and a non-significant difference when the sample sites are upwind of the highway (β = -0.10, p = 0.081). In addition, the PNC model parameter estimate for the distance from the pavement edge to the roadside sample location is positive and significant (β = 0.030, p = 0.00066), indicating that as distance from the roadway increases, the ratio of comparison location to roadside location PNC also increases. This is consistent with previous findings that PNC decrease down-range of a highway is non-linear [30].

For the ultrafine mode size distribution parameters of GMD and σ_G, the parameter estimates for both barrier types were positive, though only those for the solid barrier were statistically significant. Given that smaller sized particles have higher Brownian diffusion rates and thus higher dry deposition velocities, it was expected that the reduction in concentration for ultrafine particles would be inversely correlated with particle size. The overall reduction in PNC as well as a shift in the distribution toward larger sizes is consistent with this expectation that changes in PNC are the result of disproportionate deposition of smaller sized particles on roadside surfaces including vegetation.

For the change in the ratio of BC concentration, the parameter estimates were statistically significant for both barrier types, though in opposite directions. The estimated change was negative for the vegetative barrier (β = -0.23, p = 0.0010), indicating the reduction in BC concentration was higher for vegetative barriers than for chain-link fences when controlling for other co-variates. On the other hand, the estimated change for solid barriers was positive (β = 0.21, p = 0.034). For wind direction categories, the model estimates for BC are similar as for PNC with the exception that wind parallel to the highway exhibited somewhat less of a reduction in BC concentration than wind perpendicular to the highway. Similar to that found for PNC, the BC model parameter estimate for distance to the roadway was positive and statistically-significant.

We conducted sensitivity analysis for the vegetative barrier variable using wind direction categories other than calm wind conditions. For the wind from the highway category, the parameter estimate for the vegetative barrier was -0.091 (p = 0.62) for PNC and -0.94 (p = 0.0049) for BC. For the wind parallel to the highway category, the parameter estimate for the vegetative barrier was 0.022 (p = 0.63) for PNC and 0.049 (p = 0.58) for BC. For the wind toward the highway category, the parameter estimate for the vegetative barrier was 0.51 (p = 0.0091) for PNC and -0.12 (p = 0.55) for BC. Given that calm wind conditions were predominant at the solid barrier site while there were few calm wind observations at the chain-link fence sites, sensitivity analysis for wind direction category did not yield meaningful parameter estimates for the solid barrier type due to a lack of statistical power.

The results of alternative models using a qualitative vegetation density categorical variable were similar to model results using the barrier type categorical variable. For BC, the largest reduction in concentration was for denser vegetation (β = -0.44, p = 0.042 comparing category 5 to category 1 and β = -0.47, p = 0.030 comparing category 4 to 1). For PNC, the parameter estimates for vegetation density were smaller and not statistically-significant. Given the high correlation between vegetation density and barrier type, it was not possible to evaluate both variables in the same model. However, we have chosen to focus on barrier type models because these models more plausibly include a distinction between solid barriers and chain-link fences whereas vegetation density models do not.

Discussion

In this study, both vegetative and solid barriers were associated with changes in particle concentrations when compared to locations with only a chain-link fence. We found associations between wind direction and particle concentrations, and we found that the effect of the barriers was highest when they were downwind of the highway. The observed particle reductions were more pronounced for BC than for PNC. Some of this difference is likely due to differences in sources between types of particles. BC is entirely combustion-sourced, and primarily influenced by roadway emissions during this study. PNC on the other hand has both anthropogenic and natural sources and is influenced by roadway emissions and other anthropogenic sources as well as natural biogenic emissions. For PNC, the mean concentration at the comparison site was 8.5% lower than the roadside when downwind of the highway and was 2.6% lower when upwind. For BC, the mean concentration at the comparison site was 23% lower when downwind and 10% lower when upwind. We conducted monitoring campaigns in the summer with full foliage and warm temperatures, and our results are only representative of this time range. This time period was selected to evaluate effects of vegetation with the greatest density of foliage. Other seasons may differ due to changes in the amount of foliage as well as ambient temperatures.

A majority of studies through 2017 reported reductions in pollutant concentrations between 15% and 60% as noted in a review Abhijith et al. [20]. For those studies quantifying BC and PNC, several found reductions on the order of 10–65%. Black carbon concentrations were reduced up to 12.4% under downwind wind conditions in a study of a mixed species tree stand [31]. In that study there was a reduction in BC of 7.8% under parallel wind conditions and 5.9% when winds were low [31]. Abhijith et al. [24] measured BC concentrations under six configurations of vegetative barriers. The configuration most similar to ours (instruments placed alongside and behind trees and hedges) reported a 4% reduction in BC and a 2% reduction in PNC, while a reduction of 63% for BC and 30% for PNC was measured comparing in front of and behind a row of trees and hedges when the wind direction was parallel to the roadway. Black carbon concentrations were reduced by 12.4% in a sampling study of a mixed species tree stand when the site was downwind of the road.

A notable strength of this study was our use of GLMMs to estimate the effect of roadside barriers on particulate concentrations while controlling for covariates including wind direction, traffic volume, and distance to the highway. To our knowledge, this was the first field study to use GLMMs for this purpose. Importantly, this approach afforded us the ability to estimate the effect of covariates while controlling for barrier types. The effect estimate for wind direction was lowest for calm wind conditions and was highest when the comparison location was downwind of the highway when controlling for other covariates. This is consistent with both our own results stratified by wind direction category (Fig 3) as well as with previous findings by Hagler et al. [32], one of the few roadside barrier studies reporting results stratified by wind direction.

Sensitivity analysis reinforced these findings for vegetative barrier sites. For both PNC and BC, the reduction in concentration was most pronounced when the wind was blowing from the highway. This trend may be due to that when the barrier is upwind of the highway, the contribution of non-highway sources was more pronounced, which intuitively obscures the effect of the barrier. When wind direction was parallel to the highway, both highway and non-highway sources impacted the barrier site, and our results were less consistent under these conditions.

Although sample sites for this study were adjacent to high-volume highways with traffic counts of at least 7000 vehicles per hour, these sites were also impacted by emissions from other sources. Motor vehicle emissions are among the largest contributors to air pollution in the Atlanta-region, but other important sources include emissions from biomass burning, secondary organic aerosols (SOA), and sulfate aerosol from regional coal-fired power plants [33–35]. During the time periods sampled for this study, the most important non-roadway source affecting the difference in particle concentrations between the roadside and comparison sites was SOA. Despite its potential to contribute as a non-highway particle source, SOA does not contain elemental carbon and therefore has little influence on BC observations; however, it could potentially have some impact on PNC. Previous studies have found the size distribution of SOA to be approximately 0.4–1.0 μm [36, 37]. Although this size range is larger than that measured by the TSI Model 3910 NanoScan, the tail of the distribution on the small end may extend into the ultrafine mode and lead to modest increases in PNC.

Another likely non-highway source of particles is motor vehicle emissions from other roadways in addition to the adjacent highway. All sites were near low traffic volume secondary streets, and site 1 was approximately 50 m from a bus stop on a secondary street that was serviced by a diesel-powered bus at intermittent intervals. Site 2 was 480 m from a large arterial (Buford Highway, approximately 2000 vehicles/hour during the sampling time period) and 900 m from both a heavy rail transit line (MARTA, Doraville station) and a freight rail switching yard (Norfolk Southern, Chamblee yard). Although these distances are large compared to the distance to the highway, they are not atypical for urban settings, and it is likely these sources had a non-zero influence on particle concentrations at site 2. In any case, non-highway particle sources have the effect of obscuring the impact of the barriers.

All sampling sites were located as close to the highway as possible; however, it is worth highlighting that their proximity to the traffic sources did vary among the sites. In addition, we attempted to keep the distance between the roadside and comparison sites consistent at all locations, but we were constrained in some cases by property lines such that the two sampling sites were closer to each other at some locations. At site 2 (chain-link fence), the roadside and comparison sites were the same distance from the pavement edge, and at site 3 (solid barrier), the comparison site was only 4 m further from the pavement than the roadside site. Given that particle concentrations exhibit exponential (as opposed to linear) decay with distance to the roadway [30], the ratio of particle concentration at the comparison site to the roadside site would be expected to differ based on both distance to the highway and distance between sites. In order to control for differences in sampler proximity to both the roadside and comparison sites as well as to assess the independent effects of other covariates, we included explicit terms for distance in our analytical approach using GLMMs. Despite these efforts to control for distance, we acknowledge differential proximity to the roadways as a potential, residual source of uncertainty that may explain some of the differences in our reported observations. In addition, it is important to note that the lengths of our barriers (the distance parallel to the highway) were relatively short. Previous studies have noted the impact of eddy effects along the edges of barriers whereby roadway particles are able to wrap around behind the barrier. This may have occurred during our monitoring and yielding higher concentrations behind the barrier at the locations we monitored.

By coincidence, we performed one season of sampling in the summer of 2019 before the COVID-19 pandemic occurred and the second season in summer 2020, during the initial phase of the pandemic that involved extensive stay-at-home policies in the U.S., including Atlanta. Traffic volumes in this region were more than 10% lower in 2020 than in 2019 which resulted in a larger dynamic range for the observed traffic volumes than would have otherwise occurred. Interestingly, the model estimates for the effect of traffic volume were largely null, consistent with the effect of barriers being proportionally similar for high and low volume roadways.

We observed mixed results at site 3 (which included a solid barrier) in both 2019 and 2020. For BC, the direction of the change in concentration relative to the roadside was opposite to that observed at the vegetative barrier sites. There are a number of features of this site that could potentially contribute to these findings; however, we were unable to evaluate which, if any, of these features contribute to our unexpected findings. First, the vast majority (81%) of observations at site 3 occurred under calm wind conditions, for which it was not possible to determine wind direction. Consequently, the solid barrier categorical variable was highly correlated with the calm wind conditions categorical variable, and it was not possible to infer the causal influence of the solid barrier. Second, this site was adjacent to GA-400, which is a controlled-access freeway utilized disproportionately by light passenger vehicles with comparatively fewer diesel commercial trucks, implying lower BC emissions from the highway in this location. This is consistent with our observation that the median roadside BC concentration at this site was less than half that of any other site. This could increase the impact of non-highway BC sources in the vicinity of this site (including the adjacent secondary road), which would have the effect of increasing the ratio of comparison to roadside BC at this site. Third, the sampling location at this site was along a 2-lane secondary road and in between a 6m solid metal wall and a row of 20m pine trees. The location of the comparison site in a virtual trough between a wall and a row of trees could reduce the ability of highway emissions to reach the near ground-level instrumentation, which would again increase the impact of non-highway sources.

This analysis included design limitations that may be addressed in future projects examining roadside vegetative barriers. Although this study included more monitoring sites than the majority of studies of this type, the variation between sites made comparisons challenging. We selected 5 sites in the Atlanta region based on several criteria including the presence of barriers of various types and the ability to safely operate equipment for 4–6 hours on both sides of the roadway barrier. With additional resources, a larger number of sites could be sampled to provide additional statistical power for assessing the impact of barriers and other covariates. Second, the conditions at the sample sites were controlled by various state and local agencies such that trimming or cutting of vegetative barriers occurred between sampling visits. Site 1 had more dense vegetation in year 1 than in year 2, and site 3 (the solid barrier site) had vegetation present in year 1 but not in year 2. These changing conditions may have had an effect on our results. Third, the species of vegetation differed between sample sites, and we were unable to quantify vegetation density. As other roadside barrier studies have observed or predicted using air quality models, vegetation density has a measurable impact on particle concentrations down-range from roadways [25, 32, 38–40]. Fourth, the distance from the roadway to the sample location was different for the solid barrier site than for the sites with vegetative barriers. Furthermore, the solid barrier site frequently experienced calm wind conditions, and this made it difficult to disentangle the effect of distance to roadway and wind direction category from the effect of the solid barrier. Beyond these noted limitations, this study provides much needed information on the impact of near-road particulate matter concentrations. Although this study was conducted in the southeastern United States, the findings are applicable to other regions with similar temperature as well as a mixture of both deciduous and evergreen vegetation.

Conclusion

We measured the difference in PNC and BC concentration between roadside sites and comparison sites along major highways in Atlanta, Georgia. We found that wind direction and the presence of either vegetative or solid roadside barriers affected particle concentrations away from the highway, and we observed median particulate concentrations to be lower down-range of vegetative barriers. This effect was more pronounced for BC than for PNC. We used GLMMs to estimate the effect of barriers while controlling for other covariates including wind direction, distance from the sample sites to the roadway, traffic volume on the adjacent highway segment as well as temperature and relative humidity. Modeling results were consistent with vegetative barriers reducing BC concentration away from the highway, and sensitivity analysis suggested this effect was most pronounced when the barrier was downwind of the highway. The effect of vegetative barriers on PNC were not statistically significant but were suggestive of decreasing PNC when the barrier was located downwind from the highway. Modeling results for the effect of the solid barrier were not significant, and due to high correlation between this barrier type and calm wind conditions, were difficult to disentangle from the effect of wind direction. Our results demonstrate several important study design factors to consider in future roadside barrier studies, namely the distance from the roadside to the samplers as well as the distance between roadside and comparison sample sites. These results may be comparable to other urban areas in temperate climates with similar vegetation, traffic volume and vehicle fuel mix. Our findings support the use of roadside barriers to mitigate exposure to particulate TRAP, especially vegetative barriers with higher total surface area than solid barriers to facilitate aerosol dry deposition.

The complete data set for this study is included in the online supporting information (S1 Data) as well as the R code used for dta analysis (S2 Data).

Supporting information

Acknowledgments

The authors wish to acknowledge our students for their integral part in data collection and organization, including Raheem Smith, Tareq Alhonaiti, Sophia Le, Haris Bejdic, Mohammad Islam, and Klea Troka. We also thank Donghai Liang for assistance with statistical modeling.