Data were collected from six separate monitoring locations: four of which were in Toronto, Ontario (two situated near roadways and two in urban background environments), with the remaining two located in Vancouver, British Columbia (one situated near a roadway and another in the urban background). The location of each station, along with information regarding the major roadway next to which they were located (for the near-road sites), is summarized in Table 1. The two near-road stations in Toronto, NR-TOR-1 (43.7111, -79.5433) and NR-TOR-2 (43.6590, -79.3954), and their respective instrumentation setups have been utilized and reported by others and are described therein (Sabaliauskas et al., 2012; Sofowote et al., 2018; Wang et al., 2015). The NR-TOR-1 site was positioned 10 m from Highway 401, the busiest highway in North America in terms of annual average daily traffic (AADT) with over 400 000 vehicles per day distributed across eight eastbound and eight westbound lanes. The Southern Ontario Centre for Atmospheric Aerosol Research (SOCAAR) served as the second near-road site (NR-TOR-2), and was located 15 m from College Street in downtown Toronto, which experienced traffic volumes of 17 200 vehicles per day on average. The northernmost station in Toronto, BG-TOR-1, was located at Environment and Climate Change Canada (43.7806, -79.4675), 180 m from the nearest roadway, and the measurements from this station served as an urban background/baseline for NR-TOR-1, which was located 9.8 km to the southwest of it. The second background station, BG-TOR-2, was positioned on the southernmost point of the Toronto Islands on Lake Ontario (43.6122, -79.3887), and was 5.2 km south of NR-TOR-2. Since vehicular traffic on the Toronto Islands was limited to a small number of service vehicles, the BG-TOR-2 station was well removed from tailpipe emissions.

The near-road station in Vancouver, NR-VAN, was situated 6 m from Clark Drive (49.2603, -123.0778), a major roadway that experienced on average 33 100 vehicles per day across four southbound and three northbound lanes. Additionally, located 65 m south of the station was a major intersection, Clark Drive and 12th Avenue, at which there were two gas stations located on the northwest and northeast sides. The effect this intersection had on traffic patterns (stop-and-go especially) directly next to the station and its effect on measured TRAP concentrations are explored in this study. Lastly, the urban background station in Vancouver, BG-VAN, was located 2.2 km east of NR-VAN at Sunny Hill Children's Hospital (49.2529, -123.0492). This area was relatively removed from traffic emissions because it was located within a neighbourhood zoned predominately for single unit family dwellings.

A common suite of instrumentation was employed at all stations. Gas-phase pollutants measured include carbon dioxide (CO2; 840A, LI-COR Biosciences; attenuation of infrared radiation at wavelengths of 4.26 and 2.95 µm for H2O differentiation), carbon monoxide (CO; 48i, Thermo Scientific; attenuation of infrared radiation at a wavelength of 4.6 µm), ozone (O3; 49i, Thermo Scientific; attenuation of ultraviolet radiation at a wavelength of 254 nm), and nitrogen oxides (NOx; 42i, Thermo Scientific; infrared chemiluminescence). Particle-phase pollutant properties measured include mass concentration of particles less than 2.5 µm in diameter (PM2.5; SHARP 5030, Thermo Scientific; beta attenuation and light scattering); particle number concentration (UFP; 651, Teledyne API; water-based condensation particle counting); and black carbon (BC; AE33, Magee Scientific; filter-based attenuation of 880 nm wavelength light) mass concentration. Additionally, a meteorological sensor (WXT520, Vaisala; ultrasonic anemometer) recorded wind direction, wind speed, ambient temperature, pressure, and relative humidity at each station. Traffic intensities, velocities, and approximate vehicle lengths were measured continuously (SmartSensor HD, Wavetronix; dual beam radar) at the three near-road stations.

Gas-phase instruments were calibrated on-site every 3 months using cylinders of compressed gases at certified concentrations (Linde). One cylinder contained SO2, CO, and CO2, while the other contained NO; both contained N2 as an inert makeup gas. Dilution and mixing of the gases was accomplished using a dynamic gas calibrator (146i, Thermo Scientific; 6100, Environics) to produce zero checks and span concentrations that were similar to ambient ranges. Additionally, these dynamic gas-phase calibrators contained O3 generators based on ultraviolet (UV) radiation which were used to calibrate the 49i monitors as well as test the efficiencies of the molybdenum NO2 converters in the 42i monitors. SHARP 5030 instruments were zero checked using a HEPA filter, had their temperature and relative humidity sensors calibrated, and were span checked using mass standards supplied by Thermo Fisher Scientific twice annually. In addition to recommended monthly maintenance procedures for the API 651, each instrument underwent routine annual calibration by the manufacturer. Flow rates at each station were verified on a monthly basis, and a variable flow rate pump was attached to a stainless steel particle manifold, from which all particle-phase instruments were sampled, to ensure a constant flow rate of 16.7 Lmin-1 to satisfy the 2.5 µm cut-off conditions of the inlet cyclone.

The first method for determining local pollutant concentrations explored in this paper, henceforth referred to as method 1, is through the difference between concentrations measured at a near-road location, CNR, and at the nearest urban background location, CBG, for some concurrent observation, i. Concentrations associated with local influences determined using method 1, CL,1, rely on the assumption 1CNR[i]=CL,1[i]+CBG[i]. Average CL,1 values for each near-road location were then determined using Eq. (2): 2C‾L,1=1N∑i=1NCNRi-CBGi, again, CNR[i] and CBG[i] are near-road and urban background measurements, respectively, made over a concurrent time interval, i. As N, the number of observations used in calculating the temporal average, increases, the calculated average difference will encompass more of the variability from meteorological and traffic conditions, and therefore be more representative of an average site difference.

Through association with meteorology at a near-road measurement location, it is possible to assess traffic's influence on TRAP concentrations from the differences between downwind and upwind conditions. For example, Galvis et al. (2013) utilized average downwind and upwind concentrations of CO2, BC, and PM2.5 from a rail yard to calculate local pollutant concentrations for use in fuel-based emission factor calculations. A similar approach is used here to isolate concentrations emitted from a roadway, henceforth referred to as method 2. Defining ranges of wind directions as corresponding to downwind and upwind of the major street next to which a station is located, average local concentrations from method 2, CL,2, can be estimated using Eq. (3): 3C‾L,2=1N∑i=1NCDWi-1M∑i=1MCUWi, where CDW and CUW are near-road TRAP concentrations measured when winds originate from downwind and upwind of the major roadway, respectively. Note that the number of points used to compute the averages of these conditions, N and M, are not necessarily equivalent, and the times that comprise these two averages are mutually exclusive by definition. For example, if the prevailing wind at a site is downwind of the roadway, then downwind data will naturally occur more frequently than upwind. Figure S1 in the Supplement shows wind frequency data as measured at each near-road site throughout the monitoring campaign. Similar to method 1, as the averaging time for both conditions is increased, confidence in CL,2 will improve. It is also important to note that because these two meteorological scenarios encompass different time frames, it is possible for certain times of day, and other factors to be overrepresented in either average.

In all analyses in which meteorological data are utilized, stagnant periods (wind speed (WS) < 1.0 m s-1) were omitted. Local concentrations cannot be estimated as a function of time using this method, as downwind and upwind concentrations cannot be measured simultaneously with a single near-road station. Also, stagnant time periods, as well as time periods that are not within the downwind–upwind ranges, are omitted, thereby increasing the amount of time needed to attain a representative average. Lastly, an inherent assumption to this method is that upwind concentrations on either side of the roadway are similar. Depending on the site, however, this assumption may not be accurate.

Extracting information from one-dimensional ambient pollution time-series data (i.e. concentration as a function of time) for the purpose of source apportionment is appealing as it allows the possibility of obtaining local and background estimates without the need for more rigorous chemical analysis, computationally expensive multivariate analyses, or measurements made at multiple locations. Such algorithms make use of the underlying principle that signal frequency is inversely related to source distance. Regional or background sources (farther away from a receptor) produce slower-varying, lower-frequency signals, whereas local (nearby) sources, such as traffic, produce faster-varying, higher-frequency signals (Tchepel and Borrego, 2010).

The frequency at which data are acquired limits the highest frequencies separable by such a method. Daily averages, for example, are too lengthy to capture processes whose timescales are much shorter – a plume from a nearby on-road vehicle, for example, would have a characteristic time on the order of seconds to minutes. Therefore, in order to isolate these local temporal fluctuations, relatively high-time-resolution data are necessary. A technique recently developed by Wang et al. (2018) applied to hourly near-road measurements in order to determine above-background pollutant concentrations for use in calculating fleet-averaged emission factors is explored further in this paper.

Over the duration of the study period average CL,1 values were calculated using method 1, as described in Sect. 3.1, with resulting differences summarized in Tables 2–4. Note that no CO2 difference was calculated between Vancouver stations because CO2 was not measured at BG-VAN.

The background-subtracted differences were smallest at NR-TOR-2; for every TRAP measured, both NR-TOR-1 and NR-VAN saw greater CL,1 concentrations in comparison. This pattern is consistent with the lower traffic volumes at NR-TOR-2. Surprisingly, despite the drastic difference in traffic intensities between NR-VAN and NR-TOR-1, CL,1 values at both sites were remarkably similar for most TRAPs. This similarity was in part due to NR-VAN's closer proximity to the roadway (6 m) compared with NR-TOR-1 (10 m), in conjunction with the significant fraction of diesel vehicles passing along Clark Drive (Wang et al., 2018). While most CL,1 concentrations were similar between these two locations, UFPs at NR-TOR-1 were significantly greater (3.0×104 vs. 1.2×104 cm-3). However, this may in part be due to seasonal bias in UFP data availability (Table S1 in the Supplement) between NR-TOR-1 and BG-TOR-1 (note in particular the lack of concurrent data during summer months when ambient UFP concentrations are often lowest).

The NO2/NOx ratios for CL,1 at NR-TOR-2 were also markedly higher than the other near-road sites; these ratios at NR-VAN, NR-TOR-1, and NR-TOR-2 were, on average, 0.18, 0.29, and 0.61, respectively. A potential explanation for this is the relative residence times of vehicle plumes prior to detection at each site: because NR-VAN was positioned closest to the roadway, it is likely that vehicle plumes were fresher upon detection, whereas NR-TOR-2 sampled within an urban canyon where air tends to stagnate and recirculate. These results emphasize an important implication for near-road monitoring policies: while NO2 alone is often regulated because of associated health effects, measurements of only NO2 may not be a reliable metric for assessing near-road health impacts, as characteristics of the site may result in NO2 being a negligible fraction of total NOx.

The average differences for O3 were negative, indicating that ozone concentrations tend to be lower near major roads. Ozone is presumably being titrated due to the higher near-road concentrations of NO. Furthermore, O3 production in downtown Toronto and metropolitan Vancouver generally occurs in a regime limited by volatile organic compounds (VOCs), meaning that the additional NOx near roads does not enhance local ozone formation (Ainslie et al., 2013; Geddes et al., 2009).

While PM2.5 is generally considered to be a more regional and homogenous pollutant in urban environments, the observed values of CL,1 (1.48, 0.27, and 2.26 µg m-3 at NR-TOR-1, NR-TOR-2, and NR-VAN, respectively) were found to be significantly greater than zero, and may be indicative of both primary tailpipe and non-tailpipe (e.g. brake wear, road dust resuspension) emissions. A recent study by Jeong et al. (2019) characterized the sources and composition of PM2.5 at both NR-TOR-1 and NR-TOR-2 using an X-ray fluorescence continuous metals monitor. They found that while concentrations of aged organic aerosol, sulfate, and nitrate were similar between the two sites, contributions from sources such as traffic exhaust, brake wear, and road dust differed significantly, and were the primary factors responsible for differences in average PM2.5 concentrations. Another study by Sofowote et al. (2018) examined in more detail the reasons for elevated PM2.5 constituents at NR-TOR-1, with particular emphasis on BC, relative to another receptor 167 m from Highway 401.

As stated previously, NR-TOR-1 was the most ideal near-road monitoring location in this study for associating TRAP measurements with local meteorology, as it was positioned on flat terrain, and the major roadway which it was stationed next to was the only significant source of TRAPs in the immediate area. Thus, the direction of wind at this site had a significant impact on measured pollutant concentrations (Fig. 1). Using the methods described in Sect. 3.2, hourly TRAP concentrations were aggregated based on wind direction, and were classified as being downwind, upwind, or neither. Downwind and upwind averages were calculated across the entirety of the study period and their differences, CL,2, are also summarized in Tables 2–4. Additional information regarding the number of downwind–upwind hours and confidence intervals are provided in the Supplement (Sect. S2). Note that downwind and upwind conditions were generally not uniform with respect to time of day (Fig. S2); however, it was found that even if downwind and upwind data occurred uniformly with respect to time of day the impact it would have on average CL,2 values is minimal for most pollutants (Tables S5 and S6).

The CL,2 values reported in Table 2 for NR-TOR-1 correspond relatively well with, but are higher than, respective CL,1 values. This is true for most pollutants, with the exception of O3 and PM2.5. The reason local concentrations generated via method 2 (CL,2) are generally greater than those generated via method 1 (CL,1) is believed to be due to the following: when a site is directly downwind from a road it will generally experience the greatest TRAP concentrations, as is this case in which there is the smallest distance for dilution between the road and the site. In contrast, CL,1 values were averaged across all meteorological scenarios. The fundamental differences between methods 1 and 2 is explored further in Sect. S3 in the Supplement.

Unlike NR-TOR-1, NR-TOR-2 was not an ideal site for applying method 2 in a straightforward manner, as it measured air samples within an urban canyon where micrometeorology was complicated by vortices, stagnation, and recirculation effects. Using the downwind and upwind sector definitions in Sect. 3.2.2, CL,2 values were calculated at NR-TOR-2 and are summarized in Table 3. This methodology of contrasting downwind and upwind pollutant averages at NR-TOR-2 was unable to produce meaningful differences and the resulting disagreement with the differences between near-road and urban background measurements (CL,1) is evident. Associating ground-level TRAP concentrations with city-scale meteorology at this site was complicated by surrounding urban architecture and the presence of an intersection approximately 50 m SW of the receptor. In actuality, the difference calculated for this site was between that of leeward and windward in-canyon concentrations, and this difference was not as substantial as the NR-TOR-2 and BG-TOR-2 average site difference. For these reasons, associating near-road pollutant concentrations with meteorological data was not an effective way of differentiating between local and regional influences on pollutant concentrations at this particular near-road site. In general, in order to attain this differentiation for measurements made in urban canyons, more complicated meteorological models are necessary; hence, simple downwind–upwind differences are not universally applicable to near-road monitoring data, especially for locations in heavily urbanized landscapes.

Lastly, the siting of NR-VAN was somewhere between NR-TOR-1 and NR-TOR-2 in terms of complexity in associating TRAP concentrations with meteorology. The presence of densely spaced residential buildings within the immediate vicinity of the measurement station resulted in surface roughness having an effect on winds carrying TRAPs from major roadways farther away. Despite this, the differences between average downwind and upwind TRAP concentrations at NR-VAN were similar to, albeit larger, than the NR-VAN–BG-VAN differences in Table 4, a result similar to that for NR-TOR-1. The fact that consistent results were seen for NR-VAN and NR-TOR-1 but not NR-TOR-2 underlines the importance of a station's location, surrounding obstructions to winds, and location of traffic sources and that associating near-road TRAP concentrations with meteorological variability should be done with caution, taking into account the subtleties of each site's environment. The apparent stronger influence of the intersection rather than traffic directly next to NR-VAN (i.e. winds originating from 90∘; see Fig. 3), despite Clark Drive being 6 m vs. the intersection being 65 m away, may seem paradoxical. We speculate that the acceleration of southbound traffic along Clark Drive at this intersection was the main source of emissions, while coasting past the site, particularly when slowing down for the stoplight, would have contributed much less.

Method 3, as described in Sect. 3.3.1, was applied to hourly pollutant concentrations, and the algorithm input parameters used were α=4 and W=8 h. From the output, CL,3 was determined as a function of time, and then averaged across the entirety of the measurement campaign; the resultant averages are summarized in Tables 2–4 for each near-road site.

A benefit to this method was that it was able to estimate local and background CO2 concentrations at NR-VAN, where CO2 measurements were made only in the near-road environment and not at the background site. This emphasizes a key advantage to approaches such as these: traffic-related signal can be isolated from near-road measurements alone, without the need for background or even meteorological measurements. Furthermore, this differentiation was performed on an hourly basis, thereby retaining information in the time domain, which was not possible with method 2.

Across all near-road locations, average CL,3 concentrations were quite similar to respective average CL,1 values, implying that method 3, which uses only near-road data, is a robust means of estimating urban background and local traffic-related pollutant concentrations. This was true even for NR-TOR-2, where micrometeorology complicated analysis using method 2. Fine particulate matter was an exception to this, however. Regarding PM2.5, because its signal was largely dominated by regional-scale sources and dynamics, temporal fluctuations in roadside PM2.5 concentrations generally varied more slowly than those of primary pollutants such as NO or BC, for example. Furthermore, this variability is generally meteorologically driven and occurs homogeneously over large areas (tens of kilometres); we posit that these variabilities associated with meteorology were falsely attributed to local signal, causing local PM2.5 concentrations ascertained through this method to be much higher than respective CL,1 concentrations. Lastly, for ambient concentrations < 80 µg m-3, the hourly precision of the SHARP 5030 is ±2 µg m-3. So, the average site differences between near-road and background sites, which are all around 2 µg m-3 or less, are likely too small for method 3 to isolate as the signal-to-noise ratio on an hourly basis is quite small.

The choice of time window parameter, when comparing results obtained from method 1, is both site-specific and pollutant-dependent. For example, shorter time windows will produce results that are in better agreement with stations that are closer in proximity. Further, the role of secondary chemistry will affect agreement between method 1 and method 3. Variability in CL,3 is shown in Table S9, where average CL,3 values are reported for W=6 and W=14. When comparing average CL,3 values to average CL,1 values as a function of W, it appears as though some pollutants produce better agreement for smaller W values (e.g. CO2 and PM2.5), whereas others agree better for larger values of W (e.g. UFPs). This is likely due to the relative homogeneity of PM2.5 and CO2 and heterogeneity of UFP concentrations in urban environments. Generally, however, it appears that the values α=4 and W=8 h are an appropriate middle ground for the pollutants considered in this study, and likely represent an urban background spatial scale of between 5 and 10 km.

Although application of method 3 was less suitable for some pollutants (i.e. PM2.5), it appears to behave in an accurate and robust manner for most others. Comparing CL,1 and CL,3 values in Tables 2–4, it appears that method 3 produced similar results when compared with method 1, with the added benefit of retaining information in the time domain and not requiring a second site. It is worth emphasizing that method 3 was an independently developed method for background-subtracting near-road data without the need for concurrent background measurements. The parameters α=4 and W=8 h were originally chosen to be generalizable for near-road measurements, and to differentiate similar local/regional scales. While a direct comparison with method 1 to assess the accuracy of method 3 is tempting, method 1 is not without its own limitations (i.e. differences in distance between near-road and background stations, difficulty in removing background stations from local sources). Thus, while this comparison is useful for understanding the spatial scales of different pollutants, background-subtraction parameters should not necessarily be chosen based on this alone.

Three techniques were applied to the near-road monitoring locations in this study to extract information regarding local TRAP concentrations: (1) average differences between near-road and urban background locations, (2) downwind–upwind differences in near-road measurements, and (3) average concentrations inferred through time-series analysis of near-road data. Generally, methods 1 and 3 agreed well with one another, whereas method 2 produced values that were high in comparison with the other two methods at NR-TOR-1 and NR-VAN, and generated results that were close to zero at NR-TOR-2. A comparison of the three methodologies is summarized graphically in the Supplement (Figs. S4–S6). The close agreement of methods 1 and 3, which describe the average concentrations attributed to local traffic, is encouraging, suggesting a background is inferable from near-road data alone using method 3. Method 2 was able to isolate traffic-related pollutant signal for NR-VAN and NR-TOR-1, but was not feasible for NR-TOR-2, thus highlighting a drawback of relying exclusively on wind direction data for source apportionment efforts. It is believed that method 2, while useful for isolating traffic-related pollution, is less relevant for epidemiological purposes as it only considers certain meteorological scenarios.

Subtraction of background concentrations allows the influences of local traffic on near-road TRAP concentrations to be assessed. The benefits in terms of improved understanding were examined and illustrated by applying the local concentrations thereby derived in two ways. The degree to which traffic influences TRAP concentrations beside a road can vary day to day depending on the prevailing meteorology. Using the local signal allowed the magnitude of this source of variability to be assessed in a manner that is consistent across most TRAPs and across all near-road sites. In contrast, the contribution of traffic to the total concentration will differ across pollutants. For example, some pollutants such as NO may be predominantly from traffic while others such as CO2 will be dominated by the background. Separating the local and background concentrations allowed assessment of how the portion from local traffic varied between sites and across the pollutants. Effectively, the background subtraction methodology provided estimates that illustrate how much concentrations beside a road would drop if all the traffic on that road were to be removed, as concentrations would converge to that of the urban background in that case.

The time-series-based estimates of the background concentrations were also applied to estimate the portion of the pollutant concentrations that were due to local traffic. For example approximately half of total BC concentrations were estimated to be due to local sources at NR-TOR-1 with lower and higher percent contributions at NR-TOR-2 and NR-VAN, respectively (Fig. 7). The contribution of local sources varied across the pollutants; NO had the highest local contribution at the near-road sites while CO2 had the lowest (Fig. 8). Further, this methodology was able to replicate trends in weekday–weekend background pollution variability – shown in Fig. 7 is BC, for example, with others in the Supplement (Figs. S7–S12). Local components of air pollution showed far greater differences between weekdays and weekends at each near-road monitoring location, emphasizing the effect of different on-road traffic conditions between these two sets of days. Generally, TRAP concentrations measured at urban background sites were slightly higher on weekdays compared to weekends, and this change in regional pollution was captured in the background contributions extracted from the near-road data. It should be expected that average concentrations measured at BG-TOR-1 should match the background elements of NR-TOR-1 reasonably well, with a similar argument to be made for BG-TOR-2 and NR-TOR-2; however, these urban background concentrations are likely not perfectly homogeneous throughout the city. The spatial difference between BG-TOR-1 in north Toronto and BG-TOR-2 in south Toronto was 20 km, and the difference in average pollutant levels between the two reflects this.