Mobile air monitoring data processing strategies and e ﬀ ects on spatial air pollution trends

The collection of real-time air quality measurements while in motion (i.e., mobile monitoring) is currently conducted worldwide to evaluate in situ emissions, local air quality trends, and air pollutant exposure. This measurement strategy pushes the limits of traditional data analysis with complex second-by-second multipollutant data varying as 5 a function of time and location. Data reduction and ﬁltering techniques are often applied to deduce trends, such as pollutant spatial gradients downwind of a highway. However, rarely do mobile monitoring studies report the sensitivity of their results to the chosen data processing approaches. The study being reported here utilized a large mobile monitoring dataset collected on a roadway network in central North Carolina to


Introduction
Air quality research has been revolutionized in recent years by the development and application of mobile platforms capable of resolving air pollutant concentrations in real time.These platforms -including instrumented cars, vans, bicycles, and handheld devices -have been enabled by advancements in air monitoring instrumentation, such as higher time resolution and greater portability, as well as improvements in location resolution using commercially available global positioning systems (GPSs).The mobile measurement strategy has been utilized for a diversity of applications, which can be loosely categorized as: (1) emissions characterization, (2) near-source assessment, and (3) general air quality surveying (Table 1).
Mobile monitoring is often chosen over other methods for its ability to efficiently obtain data at a high spatial resolution under a variety of different conditions.Emissions estimation can be conducted using a number of methods, including chassis dynamometer experiments, tunnel studies, and remote sensing, but mobile monitoring methods are often selected because they enable researchers to obtain in-use emissions estimates of individual vehicles under a variety of operating conditions (Park et al., 2011;Wang et al., 2009Wang et al., , 2011Wang et al., , 2012;;Westerdahl et al., 2009).

Conclusions References
Tables Figures

Back Close
Full A myriad of post-processing strategies have been applied to mobile monitoring data to answer specific research questions including algorithms fine-tuning temporal alignment, temporal or spatial smoothing, local background estimation, and detection of short-term emissions events (Table 1).Precise alignment of instruments with location data is universally important for mobile monitoring conducted while moving at relatively high speeds.If the vehicle is traveling at merely 10 m s −1 (22 mph), a two-second misalignment between an instrument and the GPS-derived location can introduce approximately 20 m of error.
After optimal alignment, extracting meaningful information from mobile monitoring data and developing intuitive visuals often requires further processing steps.For example, ultrafine particle data collected while driving on a roadway network would represent the temporal variation in local emissions and urban background, with spatial variation expected as the monitoring platform moves along a route, and evidence of short-term concentration spikes due to occasional emissions in very close proximity.Depending on the research question, different data processing steps may be applied to separate the component of interest from the other causes of variation (Fig. 1).If a researcher is interested in general spatial variation attributed to local sources, it may be desirable to minimize the influence of background variation and short-term concentration spikes caused by nearby exhaust events.
A variety of algorithms for emissions event detection and background standardization have been employed and published in the past several years (Table 2).Event detection is important both for studies that seek to estimate emission factors but also to minimize the impact of sporadic proximate exhaust when determining near-source spatial trends.Background estimation is important for extracting local-scale influences from regional background, and standardization of background is important for general air quality surveying studies which may have unique sampling routes measured over different days.Regional influences such as meteorology can cause background concentrations to vary both diurnally and daily.Without background standardization, the variable background may be erroneously attributed to spatial variation.Background Introduction

Conclusions References
Tables Figures

Back Close
Full standardization becomes particularly important for pollutants with high background relative to the total concentration.One method of background standardization is to keep the length of the sampling run short enough that temporal variation due to regional influences can be assumed to be minimal.Another approach is to complete several laps of the same route on each sampling day (e.g., Van Poppel et al., 2013), ensuring that each location measured along the route has the same representation of average background over the multiple sampling days.However, when a route completion exceeds the timeframe within which the regional background changes or comparisons are being made of multiple routes measured on different days, temporal variation of the background is an important consideration.
As a final data processing step, temporal or spatial smoothing is often applied either to reduce noise in the data or more effectively display trends (Westerdahl et al., 2005;Weijers et al., 2004;Pirjola et al., 2012).Applying a rolling median or mean can be used to maintain the high temporal resolution while reducing the amount of noise and influence of extreme outliers.Aggregating the data to a longer time window can also be used to reduce the influence of noise or to reduce the degree of autocorrelation among the measurements.Types of spatial smoothing include calculating median or mean values along fixed length intervals of the route or within a fixed radius of locations of interest.
Recently, efforts have been made to study the mobile monitoring approach.For ex-  The campaign included 12 routes within Wake, Durham, and Orange counties, North Carolina.The routes covered areas that had previously been classified using modeled traffic data as areas of traffic delay, high traffic volume, transit routes, high signal light density, and urban area.Mobile monitoring was conducted during morning rush hour (07:00-10:30 a.m.LT) on 24 weekdays between 23 August and 11 October 2012, with each run consisting of three or more laps and each route covered on two sampling days.The routes ranged from 5.2-18.1 km in length.Introduction

Conclusions References
Tables Figures

Back Close
Full

Data processing methods
Mobile monitoring data were processed and displayed using MATLAB (2012), ArcGIS (ESRI, 2011) and R version 2.15.1 (R Core Team, 2012) along with the R packages scatterplot3d (Ligges and Mächler, 2003), openair (Carslaw andRopkins, 2012), andmcgv (Wood, 2003).A noise-reduction algorithm was applied to black carbon concentrations to reduce the frequency of negative values (Hagler et al., 2011).Examples of emissions quantification, near-source air quality gradients, and general air quality surveying were selected to illustrate the implications of the following data processing steps: fine-tuning time alignment of multipollutant data, background standardization, emissions event detection, spatial smoothing and temporal smoothing.
Two methods of time alignment were compared: the engineering process and the cross-correlation method (Choi et al., 2012).The engineering process typically used in mobile monitoring campaigns consists of applying a concentration change (high efficiency particulate air filter for particle instruments, gas standard for gas instruments) at the inlet and noting the instrument response time.The cross-correlation method can be used as a fine-tuning step when multiple pollutants produced by a single source are measured and involves calculating the correlation coefficient between two co-emitted pollutants at various time lags.Because the pollutants are co-emitted, the best estimate of the difference in response times between the instruments can be assumed to correspond with the lag time that produces the maximum correlation coefficient (Choi et al., 2012).As an illustration of the impact precise time alignment can have on analysis results, emissions factors were calculated using the method described by Park et al. (2011) for a single emissions event identified using the test vehicle's webcam footage (Fig. 2).Fuel-based emissions factors for BC, CO, and ultrafine particles (diameter ≤ 100 nm, UFPs) were expressed in terms of pollutant mass or particle number emitted per kilogram of fuel burned by normalizing the integrated change in pollutant concentration to total carbon concentration change.Volatile organic compounds (VOCs) were not measured, but for the purpose of this illustration the fraction of carbon Introduction

Conclusions References
Tables Figures

Back Close
Full converted to VOCs during combustion was assumed to be negligible compared to the proportion emitted as CO 2 .The emissions factors (EF p ) were calculated using the following equation: Where ∆[P ] is the integrated change in pollutant concentration above roadway concentrations for the duration of the plume capture.Similarly, ∆[CO 2 ] and ∆[CO] are the integrated changes in the respective CO 2 and CO concentrations above the roadway concentrations.The target vehicle, identified on the monitoring vehicle's webcam and used to establish the aforementioned emission factor, was a light-duty gasoline vehicle, so 0.85 was used as the carbon weight fraction (Park et al., 2011).The roadway concentrations were calculated by fitting a straight line from the last measurement before the emissions event to the first measurement after the emissions event (Fig. 2).The roadway concentrations were subtracted from the measured concentrations, and the integrated change was calculated using a spline function in combination with integration to calculate a numerical integral.Emissions factors were calculated for both the 1 s measurements and 10 s averages (discrete windows), as well as with and without adjusting the temporal alignment using the cross-correlation method.
Three emissions event removal methods were compared: the running coefficient of variation (COV) method used by Hagler et al. (2012), the standard deviation of the background method used by Drewnick et al. ( 2012), and the rolling 25th percentile method used by Choi et al. (2012).Measurements of local exhaust tend to be both higher and more variable than measurements of well-mixed air.Both the COV method and the standard deviation of the background method rely on the high variability as well as the magnitude of measurements of local exhaust while the rolling 25th percentile relies solely on the magnitude.The running COV method (Hagler et al., 2012) was developed using UFP concentrations and consists of calculating the rolling 5 s standard deviation (2 s before and after the center data point) and dividing it by the Introduction

Conclusions References
Tables Figures

Back Close
Full mean concentration of the sampling run.The 99th percentile of the calculated COV is used as a threshold (in Hagler et al., 2012, the threshold COV for UFPs was 2) and any data points with a COV above this threshold are flagged for removal along with the data points 2 s before and after.In the standard deviation of the background method, the standard deviation of measurements (UFP or CO 2 ) below the median is calculated (σ b ).Any measurement more than 3σ b greater than the previous measurement is flagged.Subsequently, all measurements with concentration C i that meets the following criteria are flagged: Where C uf is the concentration of the last unflagged measurement and n is the number of measurements between C uf and C i .The rolling 25th percentile method involves calculating the 25th percentile of various time windows.Choi et al. (2012) used 53 s (26 s before and after the center data point) when the distance from a freeway was farther than 1 km, 31 s (15 s before and after) for distances between 300 m and 1 km, and 3 s (1 s before and after) within 300 m of a freeway.Because the majority of the data used in this comparison was between 300 m and 1 km, a 31 s window was used for the entire dataset to simplify the calculation.One final method of reducing the effect of local emissions is to use outlier resistant statistics such as the median instead of the mean.
A single run conducted on 11 October 2012, was chosen to compare these methods because of the large number of laps conducted (11) and favorable wind conditions (from the highway towards the transect).The route included a section of highway (AADT 109 000), a transect running at an angle to the highway with moderate traffic (AADT 32 000) and a low traffic road.As an illustration, gradients of CO, UFP, BC, and NO 2 along the highway transect were used to compare the effect of the emissions event removal methods on the 50 m mean concentrations and with the 50 m median concentration of the unfiltered data.Introduction

Conclusions References
Tables Figures

Back Close
Full A new method of estimating the time-varying contribution of the regional background was developed and compared with the use of a low percentile (Bukowiecki et al., 2002).This method could not be compared with averaging measurements by fixed monitoring sites (Arku et al., 2008) or designating a background area (Hagler et al., 2012;Van Poppel et al., 2013) because no stationary monitoring was conducted concurrently with the mobile monitoring and the same background zone was not covered on each sampling run.Zwack et al. (2011a, b) also used a time-varying background estimation, but instead of estimating background concentrations separately, a smooth function of time over each sampling run was included as a term in the linear regression used to determine concentration differences.
Ultimately, the results of spatial and temporal smoothing were compared.The average speed of the monitoring vehicle on the route used to compare temporal and spatial smoothing was approximately 10 m s −1 .The smoothing intervals chosen for comparison were a 5 s and 10 s window, and 10 m, 50 m, and 100 m segments.The same route chosen to compare emissions event detection was used to compare smoothing methods.For this comparison, data from sampling runs conducted on 11 October 2012 and 21 September 2012 were used and Spearman correlation coefficients were calculated for CO, BC, UFPs, NO 2 , PM 2.5 and CO 2 .

Results and discussion
The results described in this paper focus on the data from a few of the routes and the implications of various data processing steps.The complexity of the pre-processing and analysis of mobile monitoring data precludes a detailed assessment of all study results in this paper.Introduction

Conclusions References
Tables Figures

Back Close
Full

Data alignment
For the current instrument setup, the time between a concentration change at the inlet and visual inspection of instrument response ranged from 0 s to 5 s for both real-time gas and particle instruments (Fig. 3).Using the engineering method, the response time of the QCL (CO) and APS (particle count in fine to coarse range) was less than 1 s, the CAPS (NO 2 ) and aethalometer (BC) was 4 s, and the Li-COR (CO 2 ) and EEPS (UFP) was 5 s.
After applying the lags determined using the engineering method, the crosscorrelation method was used to fine-tune the alignment.CO was chosen as the reference measurement because the quantum cascade laser was the most sensitive instrument with the fastest response time.Because the primary source of CO and BC in the study area was vehicle exhaust, it was assumed that the maximum correlation would occur when the measurements were perfectly aligned.Using the cross correlation method, the measured BC concentration was found to lag the CO concentration by 3 s (Fig. 4).The other particle instruments were also found to lag the CO measurement by 3 s.The CO 2 measurement was found to lead the CO measurement by 4 s.The only pollutant measured that was not strongly correlated with CO at a specific lag time was NO 2 ; however, NO 2 was strongly correlated with UFPs at a lag of 0 s, so the lag used for UFPs (3 s) was also applied to NO 2 .
Emissions factors for a specific vehicle and specific driving condition were calculated for CO, UFPs, and BC using the data aligned by the engineering tests and the data with the lag times adjusted to maximize correlations.A 10 s average was also calculated to compare the effect of the time-base on the emissions factor calculation and temporal alignment.Using 10 s measurements resulted in higher estimates of emissions and greater differences between the data adjusted using the cross-correlation method and the un-adjusted data (Table 3).Using the 1 s data, the alignment adjustment changed the emissions factors by 0.5 %, 7 %, and 25 % for UFPs, CO, and BC respectively; while for the 10 s data alignment adjustment changed the emissions factors by 26 %, Introduction

Conclusions References
Tables Figures

Back Close
Full 12 %, and 61 %, respectively.In this example, the correct values are not known, but this illustration demonstrates the relative effect of a shift in alignment by a few seconds on the resulting emissions factors.

Comparison of methods of filtering emissions events
Three emissions event filtering methods and their effect on the shape of the pollutant gradient were compared using the data collected on 11 October 2012: the running COV method used by Hagler et al. (2012), the standard deviation of the background method used by Drewnick et al. ( 2012), and the rolling 25th percentile used by Choi et al. (2012).Using the COV method, several emissions events were identified on the transect over the course of the run (Fig. 5).
For near-source air monitoring studies, a common analysis is to consider concentrations as a function of distance from the source of interest (e.g., edge of road) (Karner et al., 2010).Similar to previous studies, elevated concentrations of mobile source pollutants were observed on the highway (boxplots in Fig. 6), and measured concentrations decreased with increased distance from the highway (Fig. 6).However, the mean 50 m concentrations along the transect are clearly affected by emissions events, as is evidenced by the mean concentrations of UFP, BC, NO 2 and CO at 250 m (Fig. 6).Using any of the emissions event filters or the median values substantially reduces the influence of these events.The 25th percentile filter results in the lowest estimates of concentrations along the transect because it affects all of the measurements, not just those influenced by local exhaust.The 25th percentile filter also results in the smoothest estimate of the gradient along the transect (Fig. 6).
Another important consideration is that different exhaust plumes contain different pollutant mixtures.For example, the plume that was encountered at 250 m caused spikes in all four exhaust indicators, while the plume encountered at 800 m caused increases in CO and UFPs but not in BC or NO 2 (Fig. 6).The measurements used as indicators of local exhaust must be chosen carefully to adequately remove the spikes Introduction

Conclusions References
Tables Figures

Back Close
Full while retaining the majority of the data.For this run, by using both CO and UFP as indicators, the spikes in NO 2 and BC were successfully removed.

Comparison of background standardization methods
One method of accounting for background variation is to calculate a single value for each sampling run to use to normalize the concentrations.This value can be a fixed concentration such as the 1st or 10th percentile of the measurements (Bukowiecki et al., 2002) or can be added into a model as a random effect.However, in the present study over the course of a two hour sampling period, the baseline of the CO time series decreased from 400 ppb to 200 ppb (Fig. 7).Depending on the research question and the pollutant of interest, using a single value to normalize the data may introduce unnecessary error.A rolling minimum did not appear to be a good alternative.A 60 s rolling minimum is a better descriptor of variation in local concentrations than variation in urban background (Fig. 7b).A 300 s rolling minimum results in a more drastic stairstep pattern which is not descriptive of the change in background over time which generally changes very gradually.
In many studies, the sampling routes include a designated area that is considered background (Hagler et al., 2012;Van Poppel et al., 2013).The measurements made in this area over time can be separated out and used to normalize the rest of the measurements.This method may require additional analysis and for some routes, it may be difficult to determine which areas should be considered background.To estimate a temporally variable background over the course of a sampling run without specific designation of a background area, a 5-15 min minimum can be used (Fig. 7c).In this case, a rolling 11 s median was used to smooth the time series and then the minimum was calculated for each 5 min or 15 min discrete interval.A smooth curve was fit to the minima using a thin plate regression spline (Wood, 2003).Using a fixed temporal window relies on the assumption that the measurement platform will pass through an area of "clean" air within each interval.The route used for this analysis was so short that this Introduction

Conclusions References
Tables Figures

Back Close
Full assumption is not a problem.However, this assumption may not be appropriate with longer routes that do not frequently encounter areas representative of background.
To calculate the temporal variation of the background without making assumptions about specific time intervals, a flexible window baseline algorithm was developed.This algorithm relies on the fact that background variation occurs over a larger time scale than spatial variation.The algorithm essentially looks for local minima along the time series, but restricts the baseline to change only gradually over time.The full algorithm description is available in Appendix A.
Background standardization will have the greatest effect on measurements of pollutants that have a high regional background concentration relative to the concentrations emitted by the source of interest.Of the mobile source pollutants measured in this study, PM 2.5 , PM 10 , NO 2 , and CO 2 all fall into the category of co-emitted pollutants with high regional background concentrations (≥ 50 % of the mean measured concentrations, Table 4).In contrast, CO, BC, and UFPs can all be classified as coemitted pollutants with low regional background concentrations (Table 4).To compare the variation in background concentrations, the mean background value for each run was calculated and the between-run standard deviation (SD) was determined from the resulting 24 mean background values.Additionally, the within-run SD was calculated by first calculating the SD for each run and then taking the range of those values.The large differences in within-run SD are likely due to variations in the stability of meteorological conditions.For this reason, the range is given instead of the mean (Table 4).For CO and NO 2 , the between-run SD was greater than all of the within-run SDs (Table 4), indicating that the daily variation of these pollutants was greater than the hourly variation.For the rest of the pollutants measured, the between-run SD fell within the range of the within-run SD.Kimbrough et al. (2013) also found that the background contribution of NO 2 to the total concentration was higher than the background contribution of CO and BC, with measured upwind concentrations approximately 69 %, 63 %, and 44 % of downwind concentrations for NO 2 , CO, and BC, respectively.The background contributions measured by Kimbrough et al. (2013)  for the current study, likely because the downwind measurements were collected 20 m from the road, while many of the measurements in the current study were collected on the highway or on roads with high traffic volume causing the total concentrations to be higher and the fraction attributable to regional background to be lower.Upwind concentrations of UFPs measured by Hagler et al. (2009) were roughly 30 % of the nearest downwind site and about 50 % of the levels observed at 100 m from the road.Before background standardization, the regional background variation obscured the spatial variation in PM 2.5 (Fig. 8a).Before background removal, PM 2.5 concentrations measured on Route B on a highway with an AADT of 109 000 were below the 50th percentile when compared with all of the measurements made over the course of the field campaign, while measurements collected on Route A on a road with an AADT of 18 000 and Route C on a road with an AADT of 17 000 were all above the 50th percentile (Fig. 8a).After the influence of the regional background was removed, the spatial trends become much more evident.The majority of the measurements collected on highways (AADT ≥ 100 000) fall in the higher percentiles, and measurements made on roads with less traffic fall in the lower percentiles of the dataset (Fig. 8b).

Temporal and spatial smoothing methods
The influence of temporal and spatial smoothing on the calculation of the concentration gradient along the highway transect analyzed in Sect.4.2 was compared.The data shown were collected during two sampling runs conducted on 21 September 2012 and 11 October 2012 and were filtered using the running COV method.When compared to the raw data (Fig. 9a), spatial smoothing alone clarifies the trend (Fig. 9b).In contrast, although temporal smoothing results in fewer data points, the trend is still obscured (Fig. 9c).Furthermore, while spatial smoothing alone results in a fairly smooth gradient and the degree of spatial smoothing does not have a significant effect on the fitted curve (Fig. 9b), aggregating the data to a larger time scale before applying spatial smoothing introduces additional noise (Fig. 9d).This noise is due to the error introduced into the estimation of location by using a longer time scale.The slight increase Introduction

Conclusions References
Tables Figures

Back Close
Full in concentrations at 500 m is due to a busy intersection at this location.The model NO 2 = m × log(distance) + b was fit for each smoothing case because previous studies have found that pollutant concentrations tend to decrease exponentially with distance from a major source Karner et al. (2010).Spatial and temporal smoothing also causes pollutant concentrations to become more correlated as measured by the Spearman correlation coefficients.The average speed of the car on this route was approximately 10 m s −1 .The Spearman correlation coefficients were calculated for BC, CO, CO 2 , NO 2 , PM 2.5 and UFPs after applying the COV filter and after calculating 5 s and 10 s averages (discrete windows) and dividing the route into 5 m, 50 m, and 100 m segments and calculating the average of the measurements in each segment.Spatial smoothing resulted in much stronger correlations compared to temporal smoothing (Table 5).NO 2 and CO 2 were slightly negatively correlated before smoothing.After temporal smoothing the correlation coefficient remained negative, but after spatial smoothing the correlation coefficient rose to 0.80.After 10 m averaging, all of the pollutants were correlated with coefficients greater than 0.7.After 50 m averaging all of the correlation coefficients were greater than 0.8, but increasing the averaging interval to 100 m did not change any coefficients by more than 0.02.

Conclusions
The recent increase in the number of studies that employ mobile monitoring and the variety of applications demonstrate both the utility and versatility of mobile monitoring.
As air monitoring instumentation continues to advance toward greater portability, higher time resolution, greater capacity for operating autonomously, and lower costs, it is likely that these types of studies will become even more ubiquitous (Snyder et al., 2013).
The greater temporal and geographic coverage of air pollution measurements can in turn lead to better protection of health and the environment.However, as was shown in this study, this new wealth of data requires the implementation of innovative data Introduction

Conclusions References
Tables Figures

Back Close
Full proccessing techniques to extract meaningful information and develop intuitive visuals.This study investigated the sensitivity of final analysis results to the data processing steps chosen.
A variety of research questions and the corresponding data processing strategies were discussed, and a framework for deciding which strategies to apply was presented.
Precise time alignment of instruments with each other and location data is particularly important for mobile platforms traveling at high speeds and emissions factor calculations.Removal of local emissions events can substantially change a near-source gradient, but the various methods of emissions event removal that were compared resulted in similar results.Background standardization was particularly important for pollutants with a high background concentration relative to the total concentration, and estimated background concentrations were shown to vary with time.Spatial averaging (50 m) resulted in smoother concentration gradients and stronger correlations than temporal averaging (5 s).
The results demonstrate the vast amount of information contained in datasets collected using mobile monitoring and the myriad of research questions that can be answered using these data, as well as the sensitivity of the conclusions to the data processing approach utilized.

Appendix A Flexible window baseline algorithm
In this study, a new algorithm was developed to determine the background concentration as a function of time using a time series of concentrations (C t ) collected using a mobile monitoring vehicle, without using a fixed time interval.The algorithm consists of three phases: first, potential baseline points in the time series are identified, then the number of baseline points is thinned so that there is no more than one baseline point in each 5 min window, and finally a smooth function is fitted to the remaining baseline Introduction

Conclusions References
Tables Figures

Back Close
Full  with the threshold, θ, equal to the twice the standard deviation of the lowest 10 % of the measured concentrations.The 15 min window is then shifted to [t b,2 , t b,2 + 15 min] and the algorithm repeats.If C t > C t b,1 +θ for all t = [t b,1 , t b,1 +15 min], then it is assumed that the mobile platform did not pass through an area of background in that window and the window is shifted forward another 15 min so that t = [t b,1 + 15 min, t b,1 + 30 min] and the algorithm repeats.At the completion of this first pass, extraneous baseline points are removed by taking the minimum of the baseline points in each 5 min discrete window.
A smooth thin plate regression spline is fitted to the remaining points (Fig. A1).Introduction

Conclusions References
Tables Figures

Back Close
Full  Full    (c-e).Green lines represent emissions events identified using the modified COV filter (Hagler et al., 2012).
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | ample, Van Poppel et al. (2013) evaluated how many sampling route repeats were required to develop a representative data set.However, a rigorous examination of mobile monitoring data processing steps and the implications for the derived results is needed.This study utilizes a robust multipollutant mobile monitoring data set collected on a roadway network in North Carolina, USA to evaluate common data-processing methods, including time alignment optimization, emissions event detection, background standardization, and spatial and temporal smoothing.Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | are higher than those calculated Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Disclaimer.This document has been reviewed in accordance with the US Environmental Protection Agency policy and approved for publication.Mention of trade names or commercial products does not constitute endorsement or recommendation for use.The views expressed in this journal article are those of the authors and do not necessarily reflect the views or policies or the US Environmental Protection Agency.Discussion Paper | Discussion Paper | Discussion Paper | Westerdahl, D., Wang, X., Pan, X., and Zhang, K. M.: Characterization of on-road vehicle emission factors and microenvironmental air quality in Beijing, China, Atmos.Environ., 43, 697-705, 2009.10445 Wood, S. N.: Thin-plate regression splines, J. Roy.Stat.Soc.B, 65, 95-114, 2003.10449, 10455 5 Zwack, L. M., Paciorek, C. J., Spengler, J. D., and Levy, J. I.: Characterizing local traffic contributions to particulate air pollution in street canyons using mobile monitoring techniques, Atmos.Environ., 45, 2507-2514, 2011a.10445, 10452, 10466, 10467 Zwack, L. M., Paciorek, C. J., Spengler, J. D., and Levy, J. I.: Modeling spatial patterns of trafficrelated air pollutants in complex urban terrain, Environ.Health Persp., 119, 852-859, 2011b.Discussion Paper | Discussion Paper | Discussion Paper | zone Hagler et al. (2012); Van Poppel et al. (2013) Average of fixed monitoring sites Arku et al. (2008); Dionisio et al. (2010) 1 min or 5 min 5th percentile Bukowiecki et al. (2002) Assume all of the measurements lower than the most frequent measurement are background Kolb et al. (2004) Include a smooth function of time over each sampling shift as a term in a linear regression Zwack et al. (2011b, a) Emissions Event Detection Calculate standard deviation of measurements below the median (σ b ).Flag any measurement more than 3σ b greater than the previous measurement.Flag all measurements > 3σ b + √ n × σ b from the last non-flagged measurement, where n is the number of points since the last non-flagged measurement (UFPs a and CO b .(2004) Rolling 25th percentile (UFPs a , NO c , PB−PAH d , CO f , PM g 2.5 ) Choi et al. (2012) Video records were checked at times when pollution concentrations peaked at greater than twice the observed background concentrations Discussion Paper | Discussion Paper | Discussion Paper |

aFig. 2 .Fig. 5 .
Fig. 2. Emissions factor calculation using peak from local emissions event.Dashed lines represent 10 s averages.Purple lines represent emissions event.Black lines represent roadway concentrations.
points.Given a time series C t that starts at t 0 , the time of the first baseline point t b,1 and the concentration of the first baseline point C t b,1 are determined by calculating the minimum of the first 15 min of data so that(C t b,1 ) = minimum(C t )Then a 15 min window [t b,1 , t b,1 + 15 min] is used to locate the next baseline point (t b,2 ), which (if available) is the first occurrence where C t ≤ C t b,1 .If C t > C t b,1 for all t = [t b,1 , t b,1 + 15 min], then t b,2 is the first occurrence where:

Table 1 .
Mobile monitoring example applications.

Table 4 .
Summary comparison of pollutant background concentration and variation.