Science of the Total Environment 409 (2011) 2419–2429

Contents lists available at ScienceDirect

Science of the Total Environment

j ourna l homepage: www.e lsev ie r.com/ locate /sc i totenv

STEMS-Air: A simple GIS-based air pollution dispersion model for city-wideexposure assessment

John Gulliver ⁎, David BriggsMRC-HPA Centre for Environment and Health, Department of Epidemiology & Biostatistics, Imperial College London, London, UK

⁎ Corresponding author at: Department of EpidemiCollege London, St Mary's Campus, Norfolk Place, W2 1P

E-mail address: j.gulliver@imperial.ac.uk (J. Gulliver

0048-9697/$ – see front matter. Crown Copyright © 20doi:10.1016/j.scitotenv.2011.03.004

a b s t r a c t

a r t i c l e i n f o

Article history:Received 12 July 2010Received in revised form 14 February 2011Accepted 1 March 2011Available online 31 March 2011

Keywords:ParticulatesPM10

ExposureDispersion modellingGISAir pollution

Current methods of air pollution modelling do not readily meet the needs of air pollution mapping for short-term (i.e. daily) exposure studies. The main limiting factor is that for those few models that couple with a GISthere are insufficient tools for directly mapping air pollution both at high spatial resolution and over largeareas (e.g. city wide). A simple GIS-based air pollution model (STEMS-Air) has been developed for PM10 tomeet these needs with the option to choose different exposure averaging periods (e.g. daily and annual).STEMS-Air uses the grid-based FOCALSUM function in ArcGIS in conjunction with a fine grid of emissionsources and basic information on meteorology to implement a simple Gaussian plume model of air pollutiondispersion. STEMS-Air was developed and validated in London, UK, using data on concentrations of PM10 fromroutinely available monitoring data. Results from the validation study show that STEMS-Air performs well inpredicting both daily (at four sites) and annual (at 30 sites) concentrations of PM10. For daily modelling,STEMS-Air achieved r2 values in the range 0.19–0.43 (pb0.001) based solely on traffic-related emissions andr2 values in the range 0.41–0.63 (pb0.001) when adding information on ‘background’ levels of PM10. Forannual modelling of PM10, the model returned r2 in the range 0.67–0.77 (Pb0.001) when compared withmonitored concentrations. The model can thus be used for rapid production of daily or annual city-wide airpollution maps either as a screening process in urban air quality planning andmanagement, or as the basis forhealth risk assessment and epidemiological studies.

Crown Copyright © 2011 Published by Elsevier B.V. All rights reserved.

1. Introduction

Traffic is a major source of elevated levels of air pollution in manyEuropean cities. In large urban areas, in particular, variations in trafficflows and speeds, meteorology, land use, and terrain lead to complexpatterns of air pollution (Wilson and Zawar-Reza, 2006). As such, evenrelatively dense networks of routine air pollutionmonitoring are oftennot sufficient to capture the spatial and temporal patterns of airpollution exposures. Reliance on air pollution monitoring alone canthus result in exposure misclassification. As an alternative means,modelling can be used to estimate concentrations of air pollutants atpotentially any number of locations in time and space (Bellander et al.,2001; Brauer et al., 2003; Briggs et al., 2000; Nyberg et al., 2000).Indeed, air pollution models coupled with geographical informationsystems (GIS) are increasingly used to meet the needs of exposureassessment in epidemiological studies (Morra et al., 2006, 2009;Jensen et al., 2009). The type of model applied in these studies largelydepends on the averaging period used in the exposure assessment and

ology & Biostatistics, ImperialG, UK. Tel.: +44 20 7594 5027.).

11 Published by Elsevier B.V. All rig

the availability of data to feed the models. Land use regressiontechniques, for example, are commonly used for modelling long-term(i.e. annual) exposures (Hoek et al., 2008) but are not well suited todealing with short-term (e.g. daily) modelling as they do not allow forthe influence ofmeteorology. Dispersionmodels on the other hand arewell suited to short-term exposure modelling and can theoretically beused to model for any averaging period if sufficient information isavailable on source emissions and meteorology. Nevertheless,although many different dispersion models have been developed,very few have the capability for detailed air pollution mapping overlarge study areas and are therefore adequate for exposure assessmentstudies. One of the few models to meet these needs is ADMS-Urban(Carruthers et al., 2000). This borrows functionality from ESRI's ArcGISto produce high resolution maps of air pollution which can be directlyused in GIS-based exposure assessment. Similar functionality has beenachievedwith the AERMOD (Kesarkar et al., 2007), DUSTRAN (Allwineet al., 2006) and OSPM (Berkowicz et al., 2008) models. Despite thesedevelopments, however, there remain three limiting factors inapplying these models to exposure studies: 1) they have hungrydata demands (e.g. detailed data on source emissions and boundarylayermeteorology), 2) they are often expensive (i.e. typically between£3 and £10k depending on the software and type of licence), and,mostimportantly, and 3) they do not simultaneously deal with large

hts reserved.

Fig. 1. STEMS-Air: model components and data flow.

Table 1Modified Pasquill stability classes.Davies and Singh, 1985.

Wind speed(m/s)

Daytime incoming solar radiation(W/m2)

Within 1 hbefore sunsetor aftersunrise

Night-timecloud amount(oktas)

Strong(N600)

Moderate(300–600)

Slight(b300)

Overcast 0–3 4–7 8

≤2.0 A A–B B C D F–G F D2.1–3.0 A–B B C C D F E D3.1–5.0 B B–C C C D E D D5.1–6.0 C C–D D D D D D D≤≥6.0 C D D D D D D D

2420 J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

numbers of emission sources. Indeed, even if data demands canbemet,tens of model runswill usually be required tomap air pollution at highspatial resolution (i.e. b50 m) over a large urban area.

Improved methods are therefore needed for estimation ofexposures at individual level and over large study populations,particularly for short-term (e.g. daily) exposure periods. This paperpresents a GIS-based dispersion model (STEMS-Air: ‘Space TimeExposure Modelling System — Air pollution’) that provides both ameans of simultaneously dealing with large numbers of emissionsources and mapping air pollution over large urban areas. The basicframework for STEMS was first introduced by Gulliver and Briggs(2005) which previously relied on third-party, conventional disper-sion models for estimating air pollution concentrations. STEMS-Air isintended to embed into the wider STEMS framework. The dispersionmodel in STEMS-Air is implemented via a ‘moving-window’ approachin ArcGIS with grid-based FOCALSUM routines. The grid-basedapproach allows for very fast processing of large numbers of emissionsources and by using a fine grid (e.g. 10–20 metre grid cells) ensuresno significant loss of detail in the underlying emission sourcegeography. Although this functionality has been used in a handful ofenvironmental applications (Gerrard et al., 2001; Loh et al., 1996) ithas seen remarkably little use in air pollution modelling. Collins(1998) probably represents the first example of using FOCALSUM formapping air pollution. In a study in Huddersfield, UK, a distance-weighting approach was used to sum the contributions of traffic-related air pollution sources within a fixed diameter ‘kernel’ aroundeach receptor (i.e. grid cell). The model was validated for NO2

(r2=0.67) using data from routine air pollution monitoring sites. Inone of the other few published examples, Loibl and Orthofer (2001)used a similar approach to model NO2 across Austria on a 250 metregrid. Againmodels were seen to comparewell with data onmonitoredconcentrations: 54% (72 sites) and 76% (101 sites) of predictions fellwithin ±15 μg/m3 and ±25 μg/m3 of monitored concentrations,respectively. To date, however, the use of FOCALSUM appears to belimited to simple distance-weighted approaches. Vienneau et al.(2009) used FOCALSUM in ArcGIS with a fixed set of ‘distanceweights’to account for the influence of emissions around each receptor (i.e. gridcell receptor) to model concentrations of NO2 across Europe on a 1 kmgrid. This approach did not however take account of the influence ofmeteorology on pollutant dispersion. In this paper the basic approachis extended to using a Gaussian dispersion model coupled withFOCALSUM routines. The paper describes themethods used to developthe model in a GIS and related data sets, presents the results from avalidation study using routinely available monitoring data, andmakesobservations on potential improvements to the approach for furtherdevelopment work. The model is intended for predicting concentra-tions averaged over periods of at least 1 day and is not aimed atreplicating the more advanced models that have been developed withthe capability of predicting hourly concentrations of differentpollutants.

2. Material and methods

2.1. Model development

STEMS-Air operates is two modes: short-term (daily) and long-term (annual) to fit with the requirements of many epidemiologicalstudies. The outputs from short-termmode can be averaged to longerperiods (e.g. weeks or months) outside the programme (e.g. in Excel).

STEMS-Air comprises four model components: 1) an emissionsmodel, 2) a meteorological pre-processor, 3) a GIS-based air pollutiondispersion model, and 4) a GIS-based exposure assessment tool. Fig. 1shows the overall structure of STEMS-Air with input and flow of data.There is also an optional component where data on ‘background’monitored concentrations can be added from sites (e.g. rural) that arenot significantly influenced by local source emissions but represent

levels of regional and long-range (i.e. trans-boundary) particulate airpollution Each of the mandatory model components and related datais described, separately, below.

2.1.1. Emissions modelThe emissions model applies standard DMRB emissions factors

(DEFRA, 2007) to traffic flows (volume and speed), by type of vehicle(i.e. light or heavy), to produce emission rates (g/m/s) of a givenpollutant, for each road link. To account for time-varying traffic flows,emission factor profiles can be applied by hour of the day, and days ofthe week, for specific road links or geographical areas. Where detailed‘local’ emissions profiles do not exist, a default set of emission factorsare provided that can apply to the whole road network. The (vector)road geography is subsequently converted into one or more fine grids(e.g. 10 or 20 metre grid cells) of emission rates (depending on theavailability of time- and space-varying emissions profiles) for eachpollutant.

2.1.2. Meteorological pre-processorAtmospheric conditions in the model are determined by applying

Pasquill (1961) stability categories. The modified Pasquill classifica-tion (Davies and Singh, 1985), as shown in Table 1, is essentially afunction of wind speed (generally measured at about 10 m aboveground) and incoming solar radiation. There are a number of otherschemes for assessing atmospheric conditions (Mohan and Siddiqui,1998) but they require measurement of temperature and/or windspeeds at two heights which are not routinely and freely available.

Incoming solar radiation is determined by the angle between thesun's rays and the Earth's surface, commonly referred to as the zenithangle (ZN), and the amount of cloud (oktas). Zenith angles varyfor any location on the Earth's surface by time of year and time of day

2421J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

(i.e. hour angle) and can be found by Eq. (1) (Mohan and Siddiqui,1998)

cosZN = sinϕ sinδ + cosϕ cosδ cosω ð1Þ

Where δ is the solar declination angle, ϕ is the latitude, andω is thehour angle.

Declination angle (δ) is in the range ±0.13π radians, or ±23.4°,and can be found by Eq. (2) (Campbell and Norman, 1998).

sinδ = 0:39785� sin½4:871 + 0:0172J + 0:03345�

× sin 6:2238 + 0:0175Jð �ð2Þ

Where J is the calendar day (i.e. Julian day) — with J=1 on 1stJanuary and J=365 on 31st December (or J=366 in a leap year).

The hour angle (ω) is in the range ±180°and can be found by thefollowing formula (3):

cosω = 0:0833π t−tsnð Þ ð3Þ

where t is the hour of the day and tsn is the time of the solar noon.Finally, the amount of cloud cover is used to attenuate the solar

radiation calculated under clear sky conditions.The meteorological pre-processor in STEMS-Air therefore uses

data on day of the year, hour of the day, wind speed (m/s), and cloudcover to calculate, for each time period (e.g. hour), the likely incomingsolar radiation (Watts/m2) (N.B. The meteorological inputs requiredare routinely monitored by national meteorological agencies (e.g.Meteo-France, UK Met. Office) at a height of about 10 m aboveground). The urban sites in this study have weather sensorspositioned at about 5 m above ground. The resulting values for solarradiation are used in conjunction with wind speed to determine thestability classes, as shown in Table 1. A ‘MET’ file is subsequentlyproduced and exported to the GIS containing day of the year, hour ofthe day, wind speed, wind direction and stability class. Users canalternatively import their own data, based on local measurement orother models, if appropriate.

2.1.3. Air pollution dispersion modelThe STEMS-Air pollution model is based on the standard solution

of a Gaussian model of dispersion at ground level (Williams, 1990;Venkatram and Horst, 2006). Contributions to the pollutant concen-tration from sources along the wind direction vector are found bysolving Eq. (4):

C x; y; zð Þ = QπUσz

exp −12

Hσz

� �2� �ð4Þ

Contributions to pollutant concentrations from emission sourcesthat lie at an angle (θN0) to the wind direction vector are found bysolving Eq. (5):

C x; y; zð Þ = QπUσyσz

exp −12

Hσz

� �2+

yσy

!2" #ð5Þ

where C is the pollutant concentration at location x, y, and z, Q is theemission source rate (g/m/s), U is the wind speed (m/s), σz is thestandard deviation of the vertical concentration in the plume, σy is thestandard deviation of the horizontal concentration in the plume, H isthe height of the release of the emissions (e.g. 1 m above ground forground-level sources), and Y is the distance along the normal fromemission source to the plume centreline.

The stability categories, determined in the meteorological pre-processor, along with wind speed and wind direction, are used as thebasis for calculating the vertical (σz) and horizontal (σy) dispersionparameters. These are applied using a formulation of the dispersion

curves from Briggs (1973) by Venkatram et al. (2005) and Venkatramand Horst (2006) for application of ‘short-range’ dispersion models inurban areas. For unstable conditions (A and B stability classes) thefollowing parameters are applied:

σz = 0:24x 1 + 0:001xð Þ−1=2

σy =0:32x

1 + 0:0004xð Þ−1=2

ð6Þ

where x(m) is downwind distance from the source.For neutral and stable conditions (C to G classes) the following

parameters are applied:

σz = 0:14x 1 + 0:0003xð Þ−1=2

σz = 0:16x 1 + 0:0004xð Þ−1=2 ð7Þ

where x(m) is downwind distance from the source.STEMS-Air implements the dispersions model with FOCALSUM

GRID routines programmed inArcGISwithArcMacro Language (AML).This approach has two main advantages over traditional approachesfor line sources. Firstly, it overcomes the problem of having to derivean analytical expression for the concentration that accounts for theline source as a whole (Venkatram and Horst, 2006; Luhar and Patil,1989). Secondly, grid-based routines are generally very fast GISoperations and so large numbers of sources and receptors cansimultaneously be modelled to allow rapid production of city-wideair pollution surfaces. This is of particular benefit when city-wideexposure assessment is needed for large numbers of individual days(e.g. during air pollution episodeswhichmay last for up to 10 days andre-occur several times in 1 year).

Many conventional dispersion models would require a minimumof between 25 and 100 model runs, depending on the number ofemission sources (e.g. road segments) and size of the study area, toundertake city-wide exposure assessment. This may become verytime-consuming work, which requires expertise to tessellate theemission sources into a manageable set of model runs. In mostmodels, simply applying a regular grid (e.g. 2 km resolution) over acity to tessellate the emission sources would result in a very largenumber of model runs (c. 150–200 model runs in London).Furthermore, many receptors (e.g. point locations of residentialaddresses for exposure estimates) would have to be duplicated acrossseveral model runs as they may lie on the boundary between the‘zones’ used to tessellate the emission sources.

ArcGIS offers a number of options in FOCALSUM that all automaticallycreate summary statistics (sum, mean, median, standard deviation) forthose cells falling within a range of different buffer sizes and windowshapes (e.g. circle, annulus, rectangle, wedge) around each centralprocessing cell [N.B. These are, in essence, pre-programmed DOCELLoperations]. FOCALSUM can also deal with irregular shapes by providinga ‘kernel file’ with individually referenced cells.

STEMS-Air uses the FOCALSUM routine with the ‘wedge’ option.Contributions to air pollution concentrations from traffic sources are,as standard, considered within a 300 metre buffer around eachreceptor. In most cases the contributions from sources outside thisdistance are sufficiently small to be treated as ‘background’ airpollution within an urban area. Indeed, an analysis of the contributionfrom traffic source emissions summed within different buffer sizes(i.e. 100 metre concentric rings up to 1000 m) at various receptorsites in London showed that the total contribution beyond 300 mwasb2% of predicted concentrations. In some situations, however, trafficlevels beyond 300 m are sufficiently high (e.g. motorways) that thebuffer distance around a receptor needs to be extended. The user canachieve this by adding an additional buffer (e.g. up to 500 m fromsource) in the programme.

Prior to model development, regression analysis was used todetermine the optimum internal angle of the inner and outer wedges.

2422 J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

A series of wedges at 5 degree intervals between 20 and 60° wereapplied to the emissions grid using FOCALSUM, Wedges of 45° wereseen to provide best performance in predicting monitoredconcentrations.

The procedure for implementing the programme in GRID isillustrated in Fig. 2 as follows:

1. For each receptor, create a FOCALSUM ‘wedge’ along the line of thewind direction with an internal angle of 45° and a radius of 300 m.Create two outer wedges adjacent to the inner wedge each withinternal angles of 45° and of radius 40 m. The purpose of the outerwedges is to account for the relatively small contributions likely toarise from neighbouring source cells that are positioned normal tothe wind direction.

2. Intersect a series of circular buffers at 20 metre intervals (i.e. tomatch the resolution of the emissions grid and thus retain thehighest level of precision in the dispersion modelling) with thewedges to create segments inside each wedge.

3. Using FOCALSUM, intersect each wedge and segment with thesource emissions geometry (i.e. gridded emissions).

4. Sum the emissions within each segment and then apply thedispersion calculation, using Eq. (4) for the inner wedge and Eq. (5)for the outer wedge. Finally sum the contribution from each wedgeto provide a modelled PM10 concentration at the receptor.

Ideally, separate calculations would be made between eachemission grid cell and receptor but this becomes a very protractedprocess to implement when large numbers of cells are used andparticularly when the orientation (i.e. due to wind direction) of thereference cells change in relation to the central processing cell. Theuse of FOCALSUM is therefore somewhat of a compromise, but thisapproach nonetheless remains relatively fast for dealing with largenumbers of emission sources.

2.1.4. Exposure assessmentExposure assessment can be performed for a series of points (e.g.

residential address locations) by entering co-ordinates into themodel, or for areas (e.g. census tracts) by their intersection withmaps of air pollution from STEMS-Air. In grid mapping mode there is

Fig. 2. STEMS-Air dispersion

no limitation to the spatial resolution of the grid but clearly processingtime will significantly increase with finer spatial resolution (e.g.10 m). Daily or annual estimates of exposure can be determined inpoint or grid mapping modes. Resulting data can be exported asstandard delimited text files for further analysis (e.g. epidemiologicalanalysis or health assessment).

2.2. Model application

Modelling was carried out in London, UK, because data on roadgeography and traffic flow/composition are readily available from theLondon Atmospheric Emissions Inventory (LAEI), and also there is arelatively dense network of air pollution monitoring sites for modelvalidation. PM10 was chosen as the reference pollutant because of thestrong body of epidemiological evidence linking particulates withnegative effects on health (Brunekreef and Holgate, 2002; Dockeryand Pope, 2006). STEMS-Air was implemented for both short-term(i.e. daily) and long-term (i.e. annual) air pollution modelling. Short-term modelling was undertaken for July to December 2007 as data onco-located hourly traffic composition and speeds, meteorological data,and air pollution monitoring were available during this period at anumber of sites. Long-term modelling was applied to 2001–2003 asthis period represented the best coverage of air pollution monitoringsites (n=30) with complete data records across three consecutiveyears.

2.2.1. Short-term modellingData from co-located traffic and air pollution monitoring at six

roadside/kerbside sites (from the London Air Quality Network) werekindly provided by King's College London and used as the primarybasis for testing and validation of STEMS-Air in relation to short-term(i.e. daily) modelling. The data covered 184 days between 1st July and31st December 2007 and comprised hourly traffic composition fordifferent vehicle categories (i.e. car, motorcycles, buses, heavy goodsetc.), counts of the number of vehicles in each category by vehiclespeed in 10 km/h intervals, monitored concentrations of PM10, andwind speed andwind direction frommeteorological data co-located atsome of the sites. Days related to unusual events (e.g. ‘bonfire week’ in

modelling procedure.

2423J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

November) and days with unusual traffic patterns (e.g. the Christmasperiod from mid-December) were subsequently removed to leave atotal of 149 days of data.

Data were screened for completeness and reduced to the followingfour sites that had good simultaneous coverage in traffic andmonitored concentrations of PM10: Brent 4 (BT4), Greenwhich 9(GR9), Hackney 6 (HK6), and Tower Hamlets 4 (TH4). Fig. 3 shows amap of the locations of these sites. All sites are located alongside busyurban roads and in relatively clear suburban settings, with theexception of HK6 which is in a street canyon with multi-storeybuildings. Also shown are the location and type (i.e. kerbside, roadside,urban background, suburban) of sites used in long-term modelling(see below).

Data on road geography and traffic (on roads except those adjacentto the traffic monitoring site) within a 300 metre buffer of eachmonitoring sites were taken from the LAEI. The LAEI contains data onannual average daily traffic (AADT) and composition (number oflight and heavy vehicles) for over 63,000 individual road links withinthe Greater London Area. Each link also is attributed with a single,average vehicle speed that applies to all vehicle categories. Roadgeography is based on data from Ordnance Survey Master-Map© andhas a ground precision of b1 mand includes individual carriageways androundabouts.

The emissions model was run with separate emission factors fordifferent vehicle speeds and categories (i.e. light/heavy) on all roadsfalling in a 300 metre circular buffer around each of the fourmonitoring sites. Emissions were summed to provide total emissionrates of PM10 on each road link. The time-varying traffic data fromeach monitoring site was then analysed to provide weightings (i.e.‘local emission profiles’ in Fig. 1) for each hour of the day and day oftheweek. As Fig. 4 shows, the emission profiles vary both between sitein the magnitude of weights and in hour-to-hour pattern, butgenerally the profiles naturally fall into three groups: weekdays,Saturdays, Sundays. The exception is at Hackney 6 which has a

Fig. 3. Air pollution monitoring site

different pattern of emissions during the evening period on Fridays.The profiles were assumed to be applicable to all other roads within300 m of the traffic monitoring site as information on time-varyingfactors for other roads in area surrounding each site was not available.There was no information available for time-varying vehicles speeds,so the long-term average speed at each site was applied to all hoursfor all roads surrounding each site. Emission weights were subse-quently applied to the baseline emissions data using the profiles inFig. 4 to provide time-varying emission rates on all source emissiongrids.

Co-located hourly data on wind speed and wind direction wereavailable with good coverage at all sites except for BT4 which wassubstituted with data from TH4 — the nearest similar site type.Data on cloud cover from Heathrow Airport, which is the nearestavailable site, were downloaded from the British Atmospheric DataCentre (www.badc.ac.uk). These data were linked to the data on windspeed and direction and entered into the meteorological pre-processor to provide hourly data on atmospheric conditions inSTEMS-Air. STEMS-Air was then run for each of the four monitoringsites to predict daily average concentrations of PM10 (μg/m3). Themodel was run at 1-hour time resolution to account for the effects ofvariability in meteorology on individual days (N.B. the concentrationbased on 24 hourly values is not the same as meteorological variablesaveraged over 1 day).

2.2.2. Long-term modellingData on monitored concentrations of annual average PM10

between 2001 and 2003, for a total of 53 locations (including sitecoordinates and site type), were downloaded from the website of TheLondon Air Quality Network (http://www.londonair.org). These yearscovered a period when relatively minor intervention took place inopening/closing of monitoring sites and thus provided a good basis foryear-to-year comparison between sites. A reduced list of sites weresubsequently produced (n=30) to include only those sites that

s used in the validation study.

Fig. 4. Time-varying emission weights for daily air pollution modelling sites.

Table 2Summary statistics for modelled daily traffic-related PM10 (μg/m3).

2424 J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

passed the following criteria: data were available for at least 75% ofdays within each year, data were denoted as fully ratified, and sitecoordinates provided on the website could be verified or adjustedusing satellite imagery from Google Earth. These monitoring siteswere used as the basis for modelling annual average PM10 fromSTEMS-Air.

Data on traffic flows and composition were taken from the LondonAtmospheric Emissions Inventory (LAEI) for each year correspondingto the above-mentioned air pollution monitoring data. Averageemission rates were then computed in the emissions model for allroads falling within 300 m of the 30 air pollution monitoring sites.

Data on hourly meteorological parameters (wind speed, winddirection, cloud cover) were downloaded for Heathrow Airportfrom BADC for the period 1st January 2001 to 31st December 2003.These data were then entered into the meteorological pre-processorto obtain data on atmospheric stability for each hour of each day.The resulting data were then summarised as frequency data: thenumber of occasions (e.g. hours during the year) on which the samecombination of meteorological variables were recorded (e.g. numberof hours during the year with frequency of wind direction of 30°,wind speed of 2 m/s, and cloud cover equals 8 oktas). This methodreduces the hourly data by about 8-fold to speed up the modelruns with no loss of information as the predictions for each line ofmet data are averaged to provide an annual mean concentration inlong-term modelling. STEMS-Air was then run to model annualmean concentrations of PM10 for the 30 monitoring sites shown inFig. 3.

Site Adjusted r2 RMSE FB N (days)

BT4 — Brent 4 0.43 17.55 −0.84 140GR9 — Greenwich 9 0.43 18.01 −1.37 54HK6 — Hackney 6 0.19 17.85 −1.06 143TH4 — Tower Hamlets 4 0.40 29.86 −1.31 144

3. Results

Model validation was undertaken by comparing predicted PM10

concentrations with monitored PM10 concentrations at the sites

shown in Fig. 3. Model performance was assessed on the basis of thevalidation analysis by computing adjusted r2, the root mean squareerror (RMSE) and fractional bias (FB) — defined for each site as:

FB =Cp−Co

� �0:5� Cp + Co

� �

where Co is the mean observed concentration and Cp is the meanpredicted concentration.

3.1. Short-term modelling

Comparisons between monitored and predicted concentrations ofdaily average PM10 were made for the four sites shown in Fig. 3.Summary statistics for model performance are shown in Table 2 andscatterplots between monitored and predicted concentrations at eachsite are shown in Fig. 5. Modelling was done for nearly all of the daysavailable for comparison with monitored concentrations except in thecase of GR9 where only 54 days of monitored concentrations wereavailable.

Fig. 5. Scatter-plots of modelled (local traffic) versus monitored concentrations of PM10 (μg/m3), July to December 2007.

2425J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

As Table 2 shows, model performance is similar for BT4, GR9 andTH4 with adjusted r2 values of 0.43, 0.43, and 0.40 respectively(pb0.001), with the site at HK6 having a lower r2 of 0.19 (pb0.001).The lower r2 at HK6 can perhaps be explained by this site beingcharacterised by a continuous street canyon, the effects of which arenot considered in STEMS-Air. Values of RMSE and FB are relativelyhigh for all sites, as would be expected when only including the local,traffic component. Values of FB lie in the range−0.84 to−1.37— a FBvalue of −1.00 represents under-prediction by a factor of 3. For bothmeasures, results are weakest at TH4 with under-prediction by abouta factor of 4 and RMSE of 29.86 μg/m3.

The results are based on comparisons made against monitoredPM10 which includes all sources (e.g. traffic, secondary particulates,wind-blown soil anddust etc.), so in termsof the explained variation intraffic-related PM10 the model is likely performing relatively well.Indeed, examination of the ratio of PM10 concentrations betweenthose at rural, urban background and roadside sites (by the authors)showed that, on average, local traffic-related emissions account for nomore than 10–30% of monitored concentrations at urban backgroundlocations, and 30–50% at urban roadside locations: the remainder isattributable to regional pollution, present at the remote rural sites. Inorder to consider the influence of regional background PM10 here,therefore, daily mean concentrations of PM10 from rural monitoringdata were added to the modelled predictions from STEMS-Air. RuralPM10 concentrations were taken as the average of data from the onlytwo rural monitoring stations in the south east of England: Harwell(70 km due west of central London), and Rochester (60 km east-south-east of central London).

The results of combining modelled concentrations from the localmodel of traffic and ‘background’ concentrations are shown in Table 3and Fig. 6. As can be seen, model performance variably improves (r2:0.41–0.61, pb0.001) at all sites as a result of adding the data on ruralmonitored concentrations. The largest improvement in terms of r2 is atHK6: r2=0.41. Values of RMSE and FB are, as would be expected,lowered at all sites. RMSE is still high at TH4 (12.75 μg/m3) comparedto the other sites (7.48 μg/m3–8.66 μg/m3) but greatly reduced withthe addition of daily mean ‘background’ PM10 concentrations.Including rural concentrations of PM10 in STEMS-Air results insignificant improvement in model performance.

3.2. Long-term modelling

Comparisons between predicted and monitored concentrations ofannual average PM10 were made for the 30 sites with data availablebetween 2001 and 2003. Annual mean concentrations of ‘background’air pollution were obtained from Harwell for each year and wereadded to the modelled concentrations — there were insufficient datato consider the rural site at Rochester.

Table 4 shows the summary statistics for performance of STEMS-Airwith the addition of concentrations of ‘background’ PM10.

As can be seen, models perform well in each year (r2: 0.67–0.77)with the best result in terms of r2, RMSE and FB for 2001 (0.77, 3.32,−0.08, respectively). Results are weakest for 2003, especially in terms ofRMSE: 10.75 μg/m3. It should be noted, however, that concentrationswere generally inflated in 2003 compared to the other years due to anumber of air pollution episodes, with four periods of unusually high

Table 3Summary statistics for modelled local, traffic-related and regional ‘background’ PM10

(μg/m3).

Site Adjusted r2 RMSE FB N (days)

BT4 — Brent 4 0.61 7.54 0.16 140GR9 — Greenwich 9 0.58 8.66 0.22 54HK6 — Hackney 6 0.41 7.48 0.16 143TH4 — Tower Hamlets 4 0.53 12.75 −0.19 144

Table 4Performance statistics for predictions of annual average of PM10 (μg/m3), 2001–2003.

Year Adjusted r2 RMSE FB N (sites)

2001 0.77 3.32 −0.08 302002 0.68 4.98 −0.20 302003 0.67 10.75 −0.37 30

2426 J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

PM10 levels in south east England. August 2003, for example, saw aperiod of 10 days when daily mean concentrations of PM10 exceeded50 μg/m3 (Stedman, 2004).

As demonstrated by the negative values of FB, the model under-predicts in all years, but only very slightly in 2001 (FB=−0.08).Furthermore, there is over-prediction for some,mostly background sitesin 2001. This is due to the rural concentrations being higher than levelsrecorded at background sites in London, which likely come from higherproportions of non-anthropogenic sources at the rural site.

To improve model performance, linear regression was used tocalibrate the model predictions to the monitored concentrations foreach year. Summary statistics for the calibrated models are show inTable 5. Fractional bias is not included as zero mean bias is implied bya linear regression. Fig. 7 shows scatterplots of modelled versusmonitored concentrations for the same data with sites classified aseither roadside (i.e. roadside/kerbside) or background (i.e. urban

Fig. 6. Scatter-plots of modelled (local traffic+rural) versus moni

background/suburban) based on the description of site types from thenational air quality archive (www.airquality.co.uk).

As Table 5 indeed shows, the calibrated model reduces RMSE tobetween about one third and one half of the levels found from the un-calibrated model (Table 4). Also noteworthy is that the constant (i.e.background) from the regression is broadly similar to that of the averageconcentration from the rural monitoring sites in 2001 (16.76 μg/m3)and 2002 (18.78 μg/m3), but in 2003 the regression yields a backgroundvalue of 25.61 μg/m3 compared to the rural concentration of20.31 μg/m3 at the rural site. Only three of the London sites hadmonitored values lower than the constant from the regression for 2003.The coefficients (i.e. slope) of the regression are 2.92, 3.92, and 5.79 for2001, 2002 and 2003, respectively. Given the values of the constant inthe regression, and notwithstanding that some sources are not includedin the modelling (e.g. wind-blown soil and dust, and vehicle tyre andbrake wear), these values imply that STEMS-Air systematically under-predicts annual concentrations of traffic-related PM10. It must be bornein mind, however, that meteorological parameters were only available

tored concentrations of PM10 (μg/m3), July to December 2007.

Table 5Calibration models for predictions of annual average PM10 (μg/m3), 2001–2003.

Year a(x) Constant SEE RMSE N (sites)

2001 2.92 18.12 1.85 1.90 302002 3.92 18.04 2.19 2.34 302003 5.79 25.61 3.31 3.47 30

2427J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

from Heathrow Airport which is relatively exposed compared to thelocations of the air pollution monitoring sites in London. The typicallyhigher levels of wind speed recorded at Heathrow Airport result insmaller modelled concentrations from STEMS-Air. Moreover, thismagnitude of under-prediction was not seen at any of the sites in theshort-term modelling where meteorological parameters were morerepresentative of ‘urban’ conditions. To look at this area in more detail,the short-term modelling was done for a second time by substitutingwind speed at each of the four sites (BT4, GR9, HK6 and TH4)withwindspeed from the meteorological site at Heathrow. Over the 6-monthperiod from July to December 2007 average wind speeds at Heathrowwere approximately three times higher than those found at the urbansites and resulting predictions of PM10 were on average 67.5% lowerthan seenwhen using localwind speed from the four sites. This suggeststhat the coefficients of the regression shown in Table 5might be reduced

Fig. 7. Scatterplots of modelled versus monitored an

by about three-fold if using wind speed from less exposed urbanlocations.

Fig. 7 shows that STEMS-Air reasonably well explains themagnitude of variation in the traffic contributions to concentrationsof PM10. There are some anomalies to this which may be due to theexclusion of other, local (i.e. domestic, industrial) sources in themodel. As can be seen in Fig. 7, one site, Marylebone (kerbside), hasconsistently higher concentrations than all other sites. Clearly this siteis influential on the performance of themodel in terms of r2. To look atthe effect of this site on overall model performance, values of r2 wereobtained when excluding Marylebone from the analysis. As expected,values of r2 are reduced but the results were still seen to be good in allyears (r2: 0.56–0.67, pb0.001).

For comparative purposes, modelling was also undertaken withADMS-Urban (Carruthers et al., 2000), a proprietary dispersionmodel,for the same validation data described above, replicating as much aspossible the inputs (i.e. emission rates and factors, meteorology) usedin STEMS-Air. ADMS-Urban was thus offered information on trafficemissions (as vectors) within a 300-metre circular buffer of eachmonitoring site, and data on background concentrations from thesame rural monitoring stations. Tables 6 and 7 show results of thevalidation analysis for ADMS-Urban for short-term modelling andlong-termmodelling, respectively. ADMS-Urban did not produce data,

nual average PM10 concentrations, 2001–2003.

Table 6Summary statistics for modelled local, traffic-related and regional ‘background’ PM10

(μg/m3) from ADMS-Urban.

Site Adjusted r2 RMSE FB N (days)

BT4 — Brent 4 0.46 6.94 −0.06 138GR9 — Greenwich 9 0.54 8.37 0.15 54HK6 — Hackney 6 0.30 6.60 0.07 143TH4 — Tower Hamlets 4 0.34 15.73 −0.31 142

2428 J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

due to incomplete information on meteorology, for two of the daysmodelled at BT4 and TH4 (see Table 6).

As shown in Table 6, in terms of r2 the performance of ADMS-Urban(r2: 0.30–0.54) is weaker than seen for STEMS-Air (r2: 0.41–0.61),though nearly the same for GR9. Conversely, however, ADMS-Urbanperformsmarginally better than STEMS-Air in terms of RMSE and FB atthree out of four sites. Correlations between modelled and monitoreddata and othermeasures of performance should, however, improve forADMS-Urban if more detailed data on meteorology (e.g. temperature,rainfall, and air pressure), information on street canyons, and modelsof terrain are offered. For long-termmodelling, as shown in Table 7, interms of r2 ADMS-Urban (r2: 0.71–0.77) performs slightly better thanSTEMS-Air (r2: 0.67–0.77) for two (2002 and 2003) out of the 3 years.Values of RMSE and FB are broadly the same for the two models.

4. Discussion

The validation results presented suggest that STEMS-Air can beapplied to both short-term and long-term modelling of PM10.Attempts to make comparisons between these results and thosefrom other studies are difficult, however, simply because particlevalidation studies are not available for many models (Holmes andMorawska, 2006). There are a few notable examples that haveevaluated the performance for dispersion models in predicting dailyaverage PM10. In a study in Christchurch, New Zealand, Wilson andZawar-Reza (2006), for example, compared predictions of dailyaverage PM10 from TAPM (The Air Pollution Model) with monitoredPM10 concentrations at 12 monitoring sites over 59 days during thewinter of 2003–2004. Results were reported in terms of RMSE andcorrelation (r). Generally higher values of RMSE were seen than thosereported here, with 10 out of the 12 sites yielding values N30 μg/m3,though the range of monitored concentrations were similar to thisstudy. Correlations are not directly comparable but were broadlyweaker (mostly in the range r=0.4–0.5) than those seen in this study.Gokhale and Raokhande (2008) compared results from three differentmodels (CALINE3, CAL3QHC, and M-GFLSM) against daily meanconcentrations of PM10 and PM2.5 at a single traffic site in the city ofGuwahati, India, over 4 months during the winter of 2005/2006. Thestudy reported fractional bias between ±0.01 and ±0.5 for the threemodels. Again these results are broadly comparable with thosepresented here when the ‘background’ air pollution model is applied(FB: −0.19 to 0.22).

Although dispersion modelling remains the only viable basis forshort-term modelling of air pollution exposures, for long-term (i.e.annual) modelling there are a range of available techniques (Jerrettet al., 2005). Of the more advanced techniques, land use regression(LUR) modelling has been used widely in Europe and North America

Table 7Performance statistics for predictions of annual average of PM10 (μg/m3) from ADMS-Urban, 2001–2003.

Year Adjusted r2 RMSE FB N (sites)

2001 0.77 5.04 −0.20 302002 0.71 3.81 −0.10 302003 0.73 11.06 −0.38 30

(Hoek et al., 2008) for modelling annual mean concentrations ofparticulates and NO2. In general, LUR models have been seen toperformwell with validation studies yielding r2 values typically in therange 0.5 to 0.7 (Hoek et al., 2008). The results for LUR models arebroadly similar to those obtained here, but STEMS-Air, apart from inall likelihood still taking longer to produce modelled exposures, hasthree distinct advantages over LUR models. Firstly, STEMS-Air istransferable between study areas whereas LUR models, which tend tobe built with local source emissions data (e.g. traffic flows, roadgeography, land use) and are calibrated to local air pollutionmonitoring data, may not transfer well (Hoek et al., 2008). Secondly,STEMS-Air can directly provide modelled concentrations that reflectthe spatial effects of yearly patterns in meteorology, unlike LURmodels which have to rely on some temporally static emissions data(e.g. land use) and can only be universally calibrated to monitoringdata for each year. Thirdly, STEMS-Air is directly based on emissions,so can be used to estimate future concentrations (e.g. under differentpolicy scenarios), or to carry out hind-casts for past (different)emission situations. Given these advantages, and with its simpleimplementation and capacity for modelling city-wide air pollution,STEMS-Air thus represents a useful alternative to LUR modelling inlong-term exposure assessment studies.

Although STEMS-Air was seen to generally perform well, it needsto be recognised that it does not provide a substitute for formaldispersion modelling where detailed data are needed for a smallnumber of locations and sources. Its role is essentially as a screeningmodel or mapping tool, when high resolution and time-varyingestimates of air pollution are needed across large study areas orpopulations. A number of potential improvements could neverthelessbe made to address some of the inherent weaknesses in the currentmodel formulation. Firstly, the current version of the model does notconsider local, non-traffic sources (i.e. domestic, industrial). In the UK,estimated annual emissions from these sources are routinely availableon a 1 km grid (e.g. National Atmospheric Emissions Inventory), withexplicit point locations for large industrial sources (e.g. stacks). Forarea sources (e.g. 1 km grids) it would be possible add a ‘boxmodel’ toSTEMS-Air similar to those included in other dispersion models suchas ADMS; for point sources emissions, an adaptation of the generalGaussian dispersion model cited above could be made to account forsource height. A second improvement would be the addition of asimple street canyon model, such as the one described by EuropeanEnvironment Agency at http://www.eea.europa.eu/publications/TEC11a/page014.html. A canyon model would require data on, or asurrogate for, building heights which have traditionally been difficultto obtain/derive. From 2009, however, Light Detection and RangingData (LIDAR) (Cici et al., 2009) from which building heights can bederived will be freely available (for research purposes) in the UK forlarge conurbations. LIDAR could theoretically be used to provide dataon streets with canyons of continuous and significant height. The datacould directly be added into STEMS-Air via a second grid to act as areference to where a canyon model should be applied. Themeteorological sub-model used in STEMS-Air is also relatively simple,and does not make use of the more advanced algorithms that havebecome available in recent years (e.g. to describe boundary layerconditions). Enhancement of this component would be possible, butat the expense of increasing data demands. Finally, further workshould include modelling and validation of range of pollutants whichmight be required for air quality management and exposureassessment, including key urban pollutants such as NO2 and CO.

5. Conclusions

A GIS-based air pollution dispersion model has been developed andapplied in London for daily and annual modelling of PM10. In broadterms, the model was seen to produce good predictions of monitoredconcentrations for different time periods and in at different locations. A

2429J. Gulliver, D. Briggs / Science of the Total Environment 409 (2011) 2419–2429

number of improvements to the model and directions for future workhave been cited. The result of the study is a simple dispersionmodel andexposure assessment tool that can be implemented for screening andmapping purposes by lay users. Incorporation of the ArcGIS FOCALSUMroutine, at the heart of the model, allows relatively quick production ofhigh resolution, city-wide air pollution maps. The maps and exposureassessment tool could be applied in epidemiological studies and healthrisk assessment looking at linksbetween air pollution andhumanhealthover daily or longer-term (e.g. annual) periods.

Acknowledgements

The research presented here was initially developed as part of theEU 5th Framework Programme on Health Effects and Risks ofTransport Systems (HEARTS), led by the World Health Organisation(Rome), with further development taking place as part of the EU 7thFramework Programme on GENeric European Sustainable Informa-tion Space for environment (GENESIS). The authors wish to thankKing's College London for providing data on traffic flows and speedsand meteorology from six roadside air pollution monitoring sites inLondon.

References

Allwine KJ, Rutz FC, Shaw WJ, Rishel JP, Fritz BG, Chapman EG, Hoopes BL, Seiple TE.DUSTRAN 1.0 user's guide: a GIS-based atmospheric dust dispersion modellingsystem. Technical report PNNL-16055. Richland, Washington: Pacific NorthwestNational Laboratory; 2006.

Bellander T, Berglind N, Gustavsson P, Jonson T, Nyberg F, Pershagen G, Järup L. Usinggeographic information systems to assess individual historical exposure to airpollution from traffic and house heating in Stockholm. Environmental HealthPerspectives 2001;109(6):633–9.

Berkowicz R, Ketzel M, Jensen SS, Hvidberg M, Raaschou-Nielsen. Evaluation andapplication of OSPM for traffic pollution assessment for a large number of streetlocations. Environmental Modelling & Software 2008;23:296–303.

Brauer M, Hoek G, Meliefste K, Fischer P, Gehring U, et al. Estimating long-term averageparticulate air pollution concentrations: application of traffic indicators andgeographic information systems. Epidemiology 2003;14(2):228–39.

Briggs GA. Diffusion estimation for small emissions. ERL, ARL USAEC Report ATDL-106.Oak Ridge, TN: US Atomic Energy Commission; 1973.

Briggs DJ, de Hoogh C, Gulliver J, Wills J, Elliott P, Kingham S, et al. A regression-basedmethod for mapping traffic-related air pollution: application and testing in fourcontrasting urban environments. Science of the Total Environment 2000;253(1–3):151–67.

Brunekreef B, Holgate ST. Air pollution and health. The Lancet 2002;360:1233–42.Campbell GS, Norman JM. Introduction to environmental biophysics. 2nd ed. Springer-

Verlag: New York; 1998. p. 167–83.Carruthers DJ, Edmunds HA, Lester AE, McHugh CA, Singles RA. Use and validation of

ADMS-Urban in contrasting urban and industrial environments. InternationalJournal of Environmental Pollution 2000;14:1–6.

Cici A, Smith-Voysey S, Jarvis C, Tansey K. Integrating building footprints and LiDARelevation data to classify roof structures and visualise buildings. Computers,Environment and Urban Systems 2009;33(4):285–92.

Collins, S., 1998. A GIS approach to modelling traffic related air pollution. PhD thesis.The University of Huddersfield, UK.

Davies ME, Singh S. Thorney Island: its geography and meteorology. Journal ofHazardous Materials 1985;11:91-124.

DEFRA. Design manual for roads and bridges. UK: HMSO; 2007.Dockery DW, Pope CA. Health effects of fine particulate air pollution: lines that connect.

Journal of the Air & Waste Management Association 2006;56:709–42.Gerrard R, Stine P, Church R, Gilpin M. Habitat evaluation using GIS: a case study

applied to the San Joaquin kit fox. Landscape and Urban Planning 2001;52:239–55.Gokhale S, Raokhande N. Performance evaluation of air quality models for predicting

PM10 and PM2.5 concentrations at urban traffic intersection during winter period.Science of the Total Environment 2008;394:9-24.

Gulliver J, Briggs DJ. Time–space modelling of journey-time exposure to traffic-relatedair pollution using GIS. Environmental Research 2005;97(1):10–25.

Hoek G, Beelen R, de Hoogh K, Vienneau D, Gulliver J, Fischer P, Briggs D. A review ofland-use regression models to assess spatial variation of outdoor air pollution.Atmospheric Environment 2008;42:7561–78.

Holmes NA, Morawska L. A review of dispersion modelling and its application to thedispersion of particles: an overview of different dispersion models available.Atmospheric Environment 2006;40:5902–28.

Jensen SS, Larson T, Deepti KC, Kaufman JD. Modeling traffic air pollution in streetcanyons in New York City for intra-urban exposure assessment in the US Multi-Ethnic Study of atherosclerosis and air pollution. Atmospheric Environment2009;43:4544–56.

Jerrett M, Arain A, Kanaroglou P, Beckerman B, Potoglou D, Sahsuvaroglu T, Morrison J,Giovis C. A review and evaluation of intraurban air pollution exposure models.Journal of Exposure Analysis and Environmental Epidemiology 2005;15:185–204.

Kesarkar AP, Dalvi M, Kaginalkar A, Ojha A. Coupling of the weather research andforecasting model with AERMOD for pollutant dispersion modeling. A case studyfor PM10 dispersion over Pune, India. Atmospheric Environment 2007;41(9):1976–88.

Loh DK, Van Stipdonk SEP, Holtfrerich DR, Hsieh Yi-Te C. Spatially constrainedreasoning for the determination of wildlife foraging areas. Computers andElectronics in Agriculture 1996;15:323–34.

Loibl W, Orthofer R. From national emission totals to regional ambient air qualityinformation for Austria. Advances in Environmental Research 2001;5:395–404.

Luhar KA, Patil RS. A general finite line source model for vehicular pollution prediction.Atmospheric Environment 1989;23:555–62.

Mohan M, Siddiqui TA. Analysis of various schemes for the estimation of atmosphericstability classification. Atmospheric Environment 1998;32(21):3775–81.

Morra P, Bagli S, Spadoni G. The analysis of human health risk with a detailed procedureoperating in a GIS environment. Environment International 2006;32:444–54.

Morra P, Lisi R, Spadoni G, Maschio G. The assessment of human health impact causedby industrial and civil activities in the Pace Valley of Messina. Science of the TotalEnvironment 2009;407:3712–20.

Nyberg F, Gustavsson P, Järup L, Bellander T, Berglind N, Jakobsson R, et al. Urban airpollution and lung cancer in Stockholm. Epidemiology 2000;11:587–95.

Pasquill F. The estimation of the dispersion of windborne material. MeteorologyMagazine 1961;90(1063):33–40.

Stedman JR. The predicted number of air pollution related deaths in the UK during theAugust 2003 heatwave. Atmospheric Environment 2004;38(8):1087–90.

Venkatram A, Horst TW. Approximating dispersion from a finite line source.Atmospheric Environment 2006;40:2401–8.

Venkatram A, Isakov V, Pankratz D, Yuan J. Relating plume spread to meteorology inurban areas. Atmospheric Environment 2005;39:371–80.

Vienneau D, de Hoogh K, Briggs D. A GIS-based method for modelling air pollutionexposures across Europe. Science of the Total Environment 2009;408:255–66.

Williams ML. Atmospheric dispersal of pollutants and the modelling of air pollution(chapter 10). In: Harrison RM, editor. Pollution, causes, effects and control. RoyalSociety of Chemistry, 2nd editionCambridge, U.K: Thomas Graham House; 1990.p. 201–19.

Wilson JG, Zawar-Reza P. Intraurban-scale dispersion modelling of particulate matterconcentrations: applications for exposure estimates in cohort studies. AtmosphericEnvironment 2006;40:1053–63.