Science of the Total Environment 409 (2011) 1145–1153

Contents lists available at ScienceDirect

Science of the Total Environment

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

An open-terrain line source model coupled with street-canyon effects to forecastcarbon monoxide at traffic roundabout

Suresh Pandian a, Sharad Gokhale b,⁎, Aloke Kumar Ghoshal c

a Center for the Environment, Indian Institute of Technology Guwahati, Indiab Department of Civil Engineering, Indian Institute of Technology Guwahati, Guwahati 781 039, Indiac Department of Chemical Engineering, Indian Institute of Technology Guwahati, India

⁎ Corresponding author. Tel.: +91 361 258 2419.E-mail address: sharadbg@iitg.ernet.in (S. Gokhale).

0048-9697/$ – see front matter © 2010 Elsevier B.V. Aldoi:10.1016/j.scitotenv.2010.12.003

a b s t r a c t

a r t i c l e i n f o

Article history:Received 14 April 2010Received in revised form 27 November 2010Accepted 1 December 2010Available online 8 January 2011

Keywords:Traffic emissionAir qualityRoundaboutStreet canyonTraffic flowDispersion model

A double-lane four-arm roundabout, where traffic movement is continuous in opposite directions and atdifferent speeds, produces a zone responsible for recirculation of emissions within a road section creatingcanyon-type effect. In this zone, an effect of thermally induced turbulence together with vehicle wakedominates over wind driven turbulence causing pollutant emission to flow within, resulting into more or lessequal amount of pollutants upwind and downwind particularly during low winds. Beyond this region,however, the effect of winds becomes stronger, causing downwind movement of pollutants. Pollutantdispersion caused by such phenomenon cannot be described accurately by open-terrain line source modelalone. This is demonstrated by estimating one-minute average carbon monoxide concentration by couplingan open-terrain line source model with a street canyon model which captures the combine effect to describethe dispersion at non-signalized roundabout. The results of the modeling matched well with themeasurements compared with the line source model alone and the prediction error reduced by about 50%.The study further demonstrated this with traffic emissions calculated by field and semi-empirical methods.

l rights reserved.

© 2010 Elsevier B.V. All rights reserved.

1. Introduction

The variability in traffic characteristics such as large trafficvolumes, intersections of roads, and accelerating and deceleratingfleet dynamics at urban street intersections or roundabouts influencespollutant concentrations within suchmicroenvironments (Lin and Ge,2006). The increased levels of pollutants in such microenvironmentsare attributed to. Further, a complex wind flow pattern and strongthermal surface conditions distort the pollutant fluxes, affecting theobserved air pollutant levels. At roundabouts, particularly those thatare non-signalized, vehicles spiral in and out, change lanes and varyspeeds in a circular pattern, resulting in complex emission anddispersion patterns. This generates a zone in which emissions are re-circulated within a road width leading to the canyon-type effectbetween the continuous moving vehicles. In this zone, an effect ofthermally induced turbulence dominates over atmospheric turbu-lence causing pollutant emission to flow within a small regionresulting into more or less equal amount of pollutants upwind anddownwind particularly during light wind conditions. Beyond thisregion, however, the effect of winds becomes stronger causingdownwind movement of pollutants. Typical characteristic of round-about of forcing vehicles to make a lateral displacement around

central-island prior to exit has a great effect on the speed ofapproaching vehicles (Polus et al., 2003). Such dynamics of fleetspeed of non-homogeneous traffic further affects the amount and fluxof emissions greatly and thereby the concentrations of pollutants.Such complex emission patterns and dispersion processes cannot bedescribed adequately by open-terrain line source models alone.

Most of the simple line source dispersion models, particularly theGaussian based, do not account for several of these factors which arevery typical in traffic intersections or roundabouts microenvironment.Moreover, they simulate downwind pollutant levels and as a result tendto over-predict very close to emission sourceswhere there is no specificflow pattern. In general, these models simulate the dispersion ofpollutants near a roadway by treating continually moving vehicles as apollutant emitting line source. A line source dispersion model treatstraffic emission as uniformly distributed across the roadway. TheCAL3QHC is one of the models which specifically accounts for severalpossible factors relevant to traffic intersections both moving and idlingvehicles (US EPA, 1995). However, non-signalized roundabout limitsthe usefulness of such models directly. This is because the traffic is incontinuous movement and vehicles are not expected to experienceidling mode. Further, Broderick et al. (2004) carried out a study at five-arm traffic roundabout in Galway city using CALINE4 in which the fivearms of the roundabout were treated as five line sources meeting at acentre point like a roadway. The results showed that the estimatedvalues for CO were far under predicted and did not capture the diurnalvariation. In such microenvironment, the combination of box model

1146 S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

approach, which captures a street canyon effects within a road sectionand line source dispersion model for away from road or emissionsource, could describe the dispersion phenomenon relatively better.

This study presents the results of a model, termed as SC-GFLSM,that combines the STREET urban pollution street canyon (SC) modelfor within a road section to account for thermal and vehicle inducedturbulences and the general finite line source model (GFLSM) foroutside the road where the influence of atmospheric turbulence ismore. The results of this coupled model are compared with the resultsof GFLSM. The study further demonstrates the impact of differentemission calculation methodologies used in air quality modeling suchas field and semi-empirical methods.

The GFLSM model is applicable for all orientation of winds to a linesource. Several studies presented the results of its application neartraffic intersection and roundabout for non-homogeneous trafficcondition. This has been, for example, demonstrated in the studiescarried out by Luhar and Patil (1989), Gokhale and Khare (2005) andGokhale and Patil (2010). The model uses empirically derivedmechanical turbulence, i.e. vehicle wake factor, as a function ofatmospheric stability (Gokhale and Khare, 2007) and therefore doesnot capture adequately the intermittent fluctuations of winds andthermally induced turbulence produced in short-time. This may be oneof the reasons forwhich it tends to over-predictwhenwinds are lowandwind directions parallel or nearly parallel to the line source as observedby Kono and Ito (1990), Benson (1992), Sivacoumar and Thanasekaran(1999) andKukkonen et al. (2001). The STREETmodel is applicable for astreet canyon because it calculates concentrations by combining aplume equation for direct impact of vehicle emitted pollutants with abox approach to enable computation of the additional impact due topollutants re-circulated within streets by vortex flow (Johnson et al.,1973; Yamartino and Wiegand, 1986).

The objective of the present study has been to estimate CO con-centration at a measurement point using coupled SC-GFLSM and

Fig. 1. Double lane four-arm rou

compare the results with those of GFLSM model. We demonstratedthis with three different emission calculation methodologies, the firstof which calculates the emission from COPERT-IV speed–emissionequations, the second one, from observed traffic density used withsemi-empirical method and the third one, on the basis of estimateddensity with semi-empirical method. The proposed SC-GFLSM modelin association with field based and semi-empirically estimatedemission rates, describes the non-uniformity of pollutant fluxesgenerated when vehicles approach, enter and leave the roundabout.This study has been carried out at a double-lane four-arm urban trafficroundabout to estimate one-minute average CO concentration for aperiod of 30 min. The entire traffic dynamics observed at the site for30 min was analyzed for capturing every 10 s characteristics of thefleet by video clips which were later averaged to 1 min for calculatingemission rates and CO concentrations. The characteristics included:vehicle category, the speed of individual vehicles, traffic density basedon measured space mean speed and semi-empirical method,approach time of vehicle, circulatory flow of vehicles, and leavingtime of vehicles.

2. Methodology

Fig. 1 shows a traffic roundabout with a measurement location forCO and meteorological station and the names of approaches. Thestudy involved data collection on CO concentrations, traffic char-acteristics andmeteorological parameters such as wind velocity, winddirection, ambient temperature, solar radiation and relative humidity.The meteorological data were measured at 12 m height and the COconcentrations and temperature were measured at 1.5 m from theground with the help of TSI CO monitor (IAQ-CALC 7545) for everyminute interval. Fig. 2 shows CO concentration data for a period of30 min along with the wind speed, direction and traffic count. Fig. 3shows the variation of temperature with time. For traffic data, videos

ndabout traffic intersection.

0

50

100

150

200

250

300

350

400

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

11:3

911

:40

11:4

111

:42

11:4

311

:44

11:4

511

:46

11:4

711

:48

11:4

911

:50

11:5

111

:52

11:5

311

:54

11:5

511

:56

11:5

711

:58

11:5

912

:00

12:0

112

:02

12:0

312

:04

12:0

512

:06

12:0

712

:08

Tra

ffic

cou

nt a

nd w

ind

dire

ctio

n

CO

and

win

d sp

eed

Time

CO concentration (ppm) Wind speed (m/s)

Wind direction (deg) Traffic count (veh)

Fig. 2. Time variation of one-minute average CO concentrations, wind speed, wind direction and traffic count.

1147S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

were taken from the rooftop of a building covering the entry and exitfor vehicles for 30 min during the peak time. The videos wereanalyzed for various traffic characteristics such as speed of a vehicle,density, approach time of a vehicle, travel length, travel time, exit timeof vehicle, and individual vehicle count. This was done with the helpof two reference lines (similar to the speed trap) and the timeinstances at which the vehicles enter and leave the road section. Thevarious traffic characteristics were used to calculate the emissions bythree different methodologies. These methodologies determine therelationships between the various traffic flow characteristics andemission rates. The COPERT-IV methodology was used to calculateemission with traffic congestion level using semi-empirical relation-ships during the same time interval. Traffic density was estimatedusing measured as well as semi-empirically estimated space meanspeed as a means for defining traffic congestion. Air quality modelinganalysis included calculation of one-minute average CO concentration

30

31

32

33

34

35

36

37

11:3

911

:40

11:4

111

:42

11:4

311

:44

11:4

511

:46

11:4

711

:48

11:4

911

:50

11:5

111

:52

11:5

3

Tem

pera

ture

(° C

)

Time

Temp at 12m Temp at 1

Fig. 3. Temperature variation an

for a period of 30 min for the three emission scenarios by SC-GFLSMmodel. The results of the model have been evaluated against themeasurements using correlation statistics.

The strength of air quality modeling study is in the quality andresolution of input data. Since we measured all the data required forthe modeling at the location simultaneously, there was every-minutesynchronization among the traffic dynamics, its variability, the windflow vectors, thermodynamic properties and the CO concentrations. Inthis study, the traffic characteristics were analyzed at microscopiclevel from the video clips within the roundabout unlike the studiescarried out by Dirks et al. (2002, 2003) and Gokhale and Pandian(2007) in which the traffic parameters were estimated only fromtraffic count and did not provide detailed descriptions of character-istics from microscopic viewpoint. Traffic analysis showed a strongdynamics within the roundabout region. The modeling domainconsisting of the entire roundabout with four arms was divided into

-8

-7

-6

-5

-4

-3

-2

-1

0

11:5

411

:55

11:5

611

:57

11:5

811

:59

12:0

012

:01

12:0

212

:03

12:0

412

:05

12:0

612

:07

12:0

8

Ric

haR

icha

rdso

n N

umbe

r (R

b)

.5m Richardson Number

d Richardson number (Rb).

1148 S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

‘approach stretches,’ ‘circulating stretches’ and ‘exit stretches,’ and aredescribed below (Fig. 1):

1) Approach stretches begin before the roundabout junction wherethe profiles of mean speeds of all drives in the before situation andin the after situation diverge from each other (due to the speedreducing effect of traffic roundabout) and include passage throughto 10 m beyond. There are four approach stretches namely N(Airport road), E (IITG road), S (Guwahati city road), and W (NH-37 Shillong highway road).

2) Circulating stretches begin right after the deflection from theentering path where the vehicles forced to take lateral movementadjacent to the circle island. There are four circulating stretcheseach curve formed by joining NE (Airport-IITG road), ES (IITG-Guwahati city road), SW (Guwahati city-NH37 Shillong highwayroad) and NW (Airport-NH37 Shillong highway road).

3) Exit stretches begin right after the circulating stretches from thecurved path where the vehicles again regroup to take longitudinalmovement towards their destination. There are four exit stretcheseach for entering approaches namely N, E, S, and W.

Three different methodologies for calculating emissions aredescribed below:

1) The COPERT-IV methodology, which uses speed-dependent equa-tions for calculating emission rates was used, termed as E1.

2) Emissions E1 were further improved with the field measuredtraffic volume, density and driving speeds as per different vehiclecategory by a semi-empirical equation, termed as E2. Eq. (1) wasused to estimate density from the traffic volume and speed(Arasan and Dhivya, 2008). The relationship between D andvehicle emission rate (i.e. VER) has been obtained from a semi-empirical method given in Eq. (2) (Dirks et al., 2003). Table 1provides necessary constants required for Eq. (2) for estimatingemissions (E2). The traffic composition and traffic speed are foundto be highly varying in nature between each arms of theroundabout. The heterogeneity nature of the existing traffic isreflected from the variations of jam density. Moreover, the nonexistence of lane discipline and lane markings also significantlycontributes to the wide variations of jam density in differentstretches. The details of the constants required for semi-empiricalequations are described elsewhere (for instance, Dirks et al., 2003;Gokhale and Pandian, 2007). The density was calculated by Eq. (1).

D =TVs

ð1Þ

Where, T is traffic count in veh/h,D is density in veh/km, Vs is spacemean velocity of traffic fleet in km/h. The VER was calculated byEq. (2).

VER =ðaD2 + bD + cÞD

Dj−D+ VER0 ð2Þ

Table 1Traffic and emission component parameters for roundabout regions as per line sources.

Parameter Approach Circle

N E S W NE ES

V0 (km/h) 17.6 35.3 21.3 23.2 16.7 29.4Dj (veh/km) 239.3 111.1 322.2 157.2 551.8 277.7VER0 (g/km/veh)

12.8 7.8 11.0 9.7 12.8 9.4

a −8.6E−05 −7.4E−04 −5.0E−05 −2.3E−04 −2.2E−05 −1.0E−b −3.8E−03 2.9E−02 −3.3E−03 −5.0E−05 3.0E−03 4.3E−0c 6.0 6.1 6.4 6.0 5.1 7.0

VER0 is the minimum vehicle emission rate at free-flow speed ingm/veh/km; a, b and c are empirical constants; VER, the vehicleemission rate after adjusting against traffic density in gm/veh.km andDj is the jam density in veh/km.

3) Emissions E1 were improved using estimated traffic density by asemi-empirical equation, termed as E3. In this case, traffic densitywas estimated by Eq. (3) considering traffic flow rate as the onlyknown input.

D =Dj

21−

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1− 4T

V0Dj

s !ð3Þ

V0 is free-flow speed (maximum achievable speed in the absenceof congestion (D=0) in km/h).

An application of SC-GFLSM model with these three differentemission methods with succeeding use of real-time and estimatedtraffic characteristics, say, from E1 to E2 and to E3 determine itsperformance level of predictions.

2.1. Air quality model

Let us understand the fundamental differences between streetcanyon and line source models and the causes of combining theseconcepts together at roundabout in the present study. The trafficroundabout, which consists of two lanes, were treated like a canyoncovered by the continuously moving vehicles on both sides in whichemission may be re-circulated. The total concentration occurring atthe receptor location was therefore calculated in two stages: the firststage, when the dispersion of emission is in close proximity of thesource from vehicular exhaust. In this case, the dispersion isinfluenced by the movement of vehicles and the dominant buoyancyflux. Hence, the concentration generated within this section wasdetermined by street canyon both leeward and windward sideequations. In the second stage, which includes the domain beyondthe canyon up to the receptor, the dispersion is mostly influenced bythe atmospheric turbulence and follows downwind movement.Hence, the concentration generated in this domain was calculatedby a line source model, GFLSM. Thus, the combination was termed asSC-GFLSM in which the boundary conditions of street-canyonequations become initial conditions for the second stage GFLSMmodel. The Eqs. (4)–(6) were used to calculate one-minute averageconcentration for a period of 30 min due to street canyon effects(Yamartino and Wiegand, 1986) in the first stage.

CLS =

KQL

ue

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffix2 + z2

p+ h0

� � ð4Þ

Where, CSL is the leeward side concentration (mg/m3); ue, themeanambient wind speed (m/s); K is dimensionless “best fit” constant(suggested to be 7); QL, the emission source strength (mg/m/s) from

Exit

SW NW N E S W

19.3 21.5 25.8 15.4 39.6 12.4485.2 335.8 191.7 329.8 134.9 244.611.4 10.3 8.6 12.2 7.7 15.4

04 −2.3E−05 −5.4E−05 −1.3E−04 −5.4E−05 −6.4E−04 −9.4E−053 −3.4E−04 1.8E−03 −7.8E−03 5.3E−03 4.6E−02 6.0E−03

5.7 5.6 6.5 4.3 5.5 4.4

1149S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

line source and h0 is initial height of pollutant dispersion (assumed tobe 2 m).

CWS =

KQL

W ueð ÞH−zH

� �ð5Þ

Where, CSW is windward side concentration (mg/m3); H is height ofthe canyon (taken as 4 m); andW is width of the canyon (taken as 7 mas equivalent to road width).

The average concentration of pollutant in the domain is given byEq. (6)

CSC =CLS + CW

S

2ð6Þ

In the second stage, with the concentrations, CSC calculated in thefirst stage, emission QSC was recalculated by dividing with theconstant,K. This emission at road section was then used to calculatethe concentration at the receptor location using GFLSM model. Thedetailed geometry for modeling is shown in Fig. 4. Thus, Eq. (7)describes the SC-GFLSM model.

CSC−GFLSM =QSC

2ffiffiffiffiffiffi2π

pσzue

exp −12

z−h0

σz

� �2� �+ exp −1

2z + h0σz

� �2� �

× erfsin θ

L2−y

� �− x−Wð Þ cos θffiffi2

pσy

2664

3775 + erf

sin θL2

+ y� �

+ x−Wð Þ cos θffiffi2

pσy

2664

3775

8>><>>:

9>>=>>;ð7Þ

Where, QSC is the emission rate per unit length within the roadsection; L, the finite length of road; θ, the angle made by road with thewind vector; z, the height of the receptor above the ground; H, theheight of line source; HP, the plume rise; h0=H+HP; σz, σy the

Fig. 4. Methodology for S

vertical and horizontal dispersion coefficients, respectively; u, themean ambient wind speed at source height H; u0, the wind speedcorrection due to traffic wake (Chock, 1978); ue = u sin θ + u0; erf,the error function and W is the lane width.

Besides, the GFLSMwas also applied alone in conventional mannerto the whole roundabout. The results of the so called SC-GFLSMmodelwere then compared with the GFLSM. The GFLSM has been tested andevaluated recently for Indian non-homogeneous traffic conditions forextensive emissions and meteorological data sets (Luhar and Patil,1989; Gokhale, 2005; Gokhale and Khare, 2005). Eq. (8) describes theGFLSM model, details of which are discussed elsewhere (Luhar andPatil, 1989).

CGFLSM =QL

2ffiffiffiffiffiffi2π

pσzue

exp −12

z−h0

σz

� �2� �+ exp −1

2z + h0σz

� �2� �

× erfsin θ

L2−y

� �−x cos θffiffi

2p

σy

2664

3775 + erf

sin θL2

+ y� �

+ x cos θffiffi2

pσy

2664

3775

8>><>>:

9>>=>>;ð8Þ

Important parameters of these models such as atmosphericstability, dispersion coefficients were estimated using standard theoryand formulae as described below:

Dispersion coefficients were estimated as the functions of stabilityconditions using Richardson number. The lateral and verticaldispersion coefficients σy and σz, respectively are written as:

σ2y = σ2

ya + σ2y0 ð9Þ

σ2z = σ2

za + σ2z0 ð10Þ

C-GFLSM modeling.

Table 2Stability criteria as per Richardson number.

Stability condition Richardson (Rb)

Extremely unstable and unstable Rbb−0.04Slightly unstable −0.03bRbb0Neutral Rb=0Slightly stable 0bRbb0.25Stable and extremely stable RbN0.25

1150 S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

The ‘a’ and ‘0’ refer to atmospheric turbulence and traffic gene-rated turbulence, respectively. The atmospheric turbulence has beenevaluated at an effective distance from a line source and at effectivesource height using Eqs. (11)–(15). The dispersion coefficients due totraffic generated turbulence have been calculated using the Eqs. (16)–(18).

σya =X sin λ

2:15 cos λð11Þ

λ depends on atmospheric stability conditions and is given by:

λ = 18:3−1:81 lnðx = 1000Þ= 57:3for unstable condition ð12Þ

λ = 14:3−1:8 lnðx= 1000Þ= 57:3for neutral condition ð13Þ

λ = 12:5−1:1 lnðx= 1000Þ= 57:3for stable condition: ð14Þ

Here, X is in meters and λ in radians. Then,

σza = a +b

sin θ× X

� �c

ð15Þ

The constants1

sin θ, a, b and c are based on stability and taken from

the GM experimental results (Chock, 1977). The other expressions aregiven by

σz0 = 3:57−0:53Uc ð16Þ

σy0 = 2σz0 ð17Þ

Uc = 1:85u0:164 cos2θ ð18Þ

The Richardson number uses wind speed, temperature differenceand the time of a day to determine atmospheric conditions.Considering the importance of convective as well as wind driventurbulence, Richardson number was used. This method simulates thechanges in temperature and wind with altitude classifying theatmospheric stability condition reasonably accurate at a shortertime scale for concentration calculations. The dominant convectiveturbulence creates unstable atmospheric condition. This conditionchanges to neutral when convection decreases and mechanical orwind driven turbulence dominates. This state of atmosphere mayfurther change when convective and mechanical turbulences arealmost absent creating highly stable condition with no verticalmixing. The Richardson number Rb is defined by Eq. (19) (Ueharaet al., 2000).

Rb =gH TH−TOð Þ

UHð Þ2 T + 273ð Þ� � ð19Þ

where g is the gravitational acceleration (m/s2), H, the height of themet station (m), TH, the temperature at the met station (°C), TO, thetemperature at the ground level, T, the reference (ambient)temperature, and UH is the wind velocity at the met station (m/s).The stability criteria are shown in Table 2 which defines theatmospheric stability condition within the region where canyoneffects are generated. Uehara et al. (2000) have reported thatatmospheric condition was weak when the Richardson number wasin the range of 0.4 and 0.8. The Rb values are shown in Fig. 3. Thevalues ranged from−0.14 to−6.77 with an average of−1.02. In oneof the studies on the temperature distribution in a street canyon,Nakamura and Oke (1988) observed the Rb values between −0.45and−0.17 on a clearmid-summer afternoon for UHN0.5. While the Rbvalues estimated in the present study are much lower (largernegative) than those reported by Nakamura and Oke (1988), theyresult in the same class of stability (i.e. extremely unstable).

2.2. Statistical evaluation of model performance

The results of SC-GFLSM and GFLSM were evaluated against themeasured CO concentrations with the help of correlation statisticssuch as index of agreement (d), root mean square error (RMSE) andfractional bias (FB). The d value ranges between 0 and 1with the valueof 1 indicates the model to be error-free. The optimum value of RMSE(expressed in the concentration unit) is zero. The FB ranges as−2≤FB≤2, where −ve indicates under-prediction and +ve overprediction, while FB zero means a perfect agreement betweenestimated and measured values. The d, RMSE and FB are given byEqs. (20)–(22).

d = 1− ∑ CP−COð Þ2

∑ jCP−CO j + jCO−CO jn o2 ð20Þ

RMSE =

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi∑n

i=1CP−COð Þ2

n

vuuutð21Þ

FB = 2 ×CP−CO

CP + CO

ð22Þ

Where, Cp and Cp are the estimated and mean estimated concen-trations; Co and Co , the observed and mean observed concentrationsand n is the number of observations.

3. Results and discussion

For themodeling analysis, the roundabout was divided into twelveline sources constituting a traffic movement flow on all the stretchesas shown in Figs. 1 and 4. Two models GFLSM and SC-GFLSM wereapplied to estimate CO concentration contributed from each linesource to the receptor point for the three emission inputs (E1, E2 andE3).

Assuming that there are no sources other than twelve line sourcesaffecting the concentration at the receptor, we added the concentra-tion calculated from all the line sources (i.e. ALL) to compare with theconcentration measured. Fig. 5a, b and c show the time variation ofconcentrations by both models against the measured concentrations.The statistics for ALL are given in Table 3. The results showed that thecombined model reduced the degree of prediction error produced inthe estimation of the line source model by a large extent. For example,the d was 0.28 for GFLSM which improved up to 0.65 for SC-GFLSMand similar improvements were observed by RMSE and FB statistics. Itmeans that the combined model estimated the concentrations muchmore satisfactorily than single line source model.

However, it was perceived that due to constant variability in windflow pattern, mostly in wind direction, and instantaneous changes inturbulence and traffic dynamics all of the line sources may notcontribute fully to the concentration at the receptor. The nature ofmoving vehicles in various directions with different speeds makes itfurther difficult to understand the pattern of pollutant dispersion evenat such a shorter time average. That is why we identified line sourcesto determine the best possible combinations, which might be

0.0

0.5

1.0

1.5

2.0

2.5

0 5 10 15 20 25 30

CO

con

cent

rati

on (

ppm

)

Time (minutes)

GFLSM SC-GFLSM Obs

0.0

0.5

1.0

1.5

2.0

2.5

3.0

0 5 10 15 20 25 30

CO

con

cent

rati

on (

ppm

)

Time (minutes)

GFLSM SC-GFLSM Obs

0.0

0.5

1.0

1.5

2.0

2.5

0 5 10 15 20 25 30

CO

con

cent

rati

on (

ppm

)

Time (minutes)

GFLSM SC-GFLSM Obs

a

b

c

Fig. 5. a: Modeled GFLSM and SC-GFLSM comparison with measured CO concentrationfor ALL with E1. b: Modeled GFLSM and SC-GFLSM comparison with measured COconcentration for ALL with E2. c: Modeled GFLSM and SC-GFLSM comparison withmeasured CO concentration for ALL with E3.

Table 3Correlation statistics for ALL and S, E.

Evaluation E1 E2 E3

Statistics GFLSM SC-GFLSM GFLSM SC-GFLSM GFLSM SC-GFLSM

AllMean 1.44 0.49 1.47 0.54 1.35 0.49Std dev 0.57 0.23 0.58 0.27 0.48 0.22CoV 0.39 0.46 0.40 0.50 0.35 0.45d 0.28 0.65 0.28 0.63 0.28 0.60RMSE 1.25 0.31 1.27 0.35 1.14 0.32FB 1.34 0.53 1.35 0.61 1.30 0.53

S–EMean 0.66 0.26 0.72 0.31 0.64 0.28Std dev 0.28 0.13 0.34 0.20 0.26 0.15CoV 0.42 0.51 0.48 0.63 0.40 0.55d 0.54 0.64 0.53 0.71 0.52 0.61RMSE 0.46 0.21 0.52 0.22 0.44 0.23FB 0.79 −0.09 0.86 0.08 0.76 −0.03

1151S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

contributing at the receptor, by adding concentrations from those linesources together. The concentrations from approach and exitstretches were combined together and named as per their positions(N, E, S and W) where N indicates the concentration from enteringand leaving vehicles from the North approach (for example, N— entry

and exit together from north) and so on. Similarly, the concentrationsfrom four circulating stretches were combined together named asCurve (for example, adding CNE, CNW, CES, CSW together) and thecontribution from all the twelve line sources were named as ALL.Similarly N, Wmeans adding N andW together while N,W, and Curvemeans adding N, W and Curve together and so on. These nomen-clatures have been shown in Fig. 1.

The results showed that the concentration from the combinationS and E (i.e. S, E) matched well and even better than from ALL. It isclear that the line sources S, E alone contributed to CO concentrationat the receptor. This indicates that the complicated dispersionpattern at the roundabout overpowers the forces of prevailing windspreventing the movement of pollutants in the downwind. Thisfurther verifies the assumption of canyon effects at the roundabout.The comparisons of one-minute average CO concentrations estimat-ed by models with the measured concentrations are shown in Fig. 6a,b and c for emissions E1, E2 and E3, respectively. It was observed thatboth models more or less captured the pattern but the one that wassimulated by SC-GFLSM model was much closer to the measuredconcentration pattern. For the second case of emission, that is E2, theestimated concentrations matched well with the measured and thusreproduced the trend relatively better than other cases of emissions.There has also been a drastic reduction in the fractional biasness asevident from the values for S, E (see Table 3). This emphasizes theimportance of field observed pertinent traffic characteristics at ashorter time scale.

Both models over-estimated the concentration as observed fromthe diurnal variations for a period of 30 min. However the relativedegree of over-prediction by SC-GFLSMwasmuch less compared withGFLSM model. It has been observed that the SC-GFLSM model has FBvalues closer to zero indicating good correlation between estimatedand measured one-minute average CO values. In comparison withGFLSM model the difference in FB values was large. It is noticeablehere that application of GFLSM alone resulted in high over-prediction,for example, for S, E in the range of 0.76 to 0.86 while it was only−0.03 to 0.08 after combined with the street-canyon effects. It hasbeen further observed that d values for SC-GFLSM model wererelatively higher, that is, improved up to 0.71.

Statistics of the SC-GFLSM model indicate that the second case ofemission, E2, measured and estimated matched well compared to theother methods of emission, E1 and E3. A marginal difference wasobserved between E2 and E3 whichmight have been because they aremore or less similar cases with traffic characteristics measured in E2and estimated in E3. This indicates that for more accurate predictionsit may be necessary to obtain vehicle and traffic characteristics by fieldmeasurements and or semi-empirical methods incorporating funda-mental field variables. This is because the traffic dynamics particularly

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

0 5 10 15 20 25 30

CO

con

cent

rati

on (

ppm

)

Time (minutes)

GFLSM SC-GFLSM Obs

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

0 5 10 15 20 25 30

CO

con

cent

rati

on (

ppm

)

Time (minutes)

GFLSM SC-GFLSM Obs

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

0 5 10 15 20 25 30

CO

con

cent

rati

on (

ppm

)

Time (minutes)

GFLSM SC-GFLSM Obs

a

b

c

Fig. 6. a: Modeled GFLSM and SC-GFLSM comparison with measured CO concentrationfor S, E with E1. b: Modeled GFLSM and SC-GFLSM comparison with measured COconcentration for S, E with E2. c: Modeled GFLSM and SC-GFLSM comparison withmeasured CO concentration for S, E with E3.

1152 S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

traffic flow, speed, heterogeneity, vehicle composition and drivingmodes vary place to place. RMSE values were close to optimum valuefrom 0.21 to 0.23 for all the three emission inputs.

The variation in wind direction tends to increase with decreasingwind speed, causing increased plume meandering. Substantialmeandering can cause measured concentrations at a certain point to

be much lower than those computed assuming a homogeneous windflow (Benson, 1992) as in the case of GFLSM model. It is also difficultto evaluate atmospheric conditions, the extent of vertical mixing andatmospheric dispersion processes at very low or calm winds. Theeffects of these limitations seem to have been reduced aftercombining street canyon and line source models and with trafficand atmospheric dynamics at shorter-time scale. This is of coursewithin the road sections where continuous moving vehicles of aboutaverage 4 m height form a street-canyon and pollutants undergofaster mixing in vertical direction than advection in the direction ofwind. The higher negative values of Richardson number support thisperception.

4. Conclusions

Line source models may forecast pollutant concentrations wellfrom an open roadway or a highway but not at roundabout where thedynamics of traffic behavior and meteorology together causepollutants circulation within road sections. This might be the reasonthat line source model, which simulates pollutant fluxes downwind,alone estimated measured concentrations almost three times higher.But after combining with street-canyon effects, the degree of overprediction reduced to great degree with respect to means and theprediction errors reduced by about 50%. The results were consistentfor all the emission methodologies. From the three emission methods,the results with E2 and E3 were closer to measured concentrationsindicating that field observations of traffic density and its semi-empirical relationship with vehicle emission rates described thedispersion of traffic emissions more adequately. This study envisionsthat such combination of two models may evaluate air quality moreaccurately even from multi-lane roadways. However, further verifi-cation of the analysis may be necessary with measurements at severallocations and for different averaging periods.

References

Arasan VT, Dhivya G. Measuring heterogeneous traffic density. World Congress onScience, Engineering and Technology, December 17–19, 2008. Bangkok, Thailand;2008.

Benson PE. A review of the development and application of the CALINE3 and 4 models.Atmos Environ 1992;26B(3):379–90.

Broderick B, Budd U, Misstear B, Ceburnis D, Jennings SG. Modelling CO concentrationsunder free-flowing and congested traffic conditions in Ireland. 18–22. June 1–4(2004), Garmisch-Partenkirchen, Germany; 2004.

Chock DP. General Motors sulfate dispersion experiment: assessment of the EPAHIWAY model. J Air Pollut Control Assoc 1977;27:39–45.

Chock DP. A simple line-source model for dispersion near roadways. Atmos Environ1978;12:823–9.

Dirks KN, Johns MD, Hay JE, Sturman AP. A simple semi-empirical model for predictingmissing carbon monoxide concentrations. Atmos Environ 2002;36:5953–9.

Dirks KN, Johns MD, Hay JE, Sturman AP. A semi-empirical model for predicting theeffect of changes in traffic flow patterns on carbon monoxide concentrations.Atmos Environ 2003;37:2719–24.

Gokhale S. Prediction of carbon monoxide levels at urban hotspots using hybridmodelling technique. The National Level BrainstormingWorkshop on Indian UrbanAir Quality 2005 24–25 October, 2005 at NEERI, Nagpur, India; 2005.

Gokhale S, Khare M. A hybrid model for the prediction of carbon monoxide fromvehicular exhausts in urban environments. Atmos Environ 2005;39(22):4025–40.

Gokhale S, Khare M. Vehicle wake factor for heterogeneous traffic in urbanenvironments. Int J Environ Pollut 2007;30(1):97-105.

Gokhale S, Pandian S. A semi-empirical box modeling approach for predicting thecarbon monoxide concentrations at an urban traffic intersection. Atmos Environ2007;41:7940–50.

Gokhale S, Patil R. Uncertainty modeling of PM10 and PM2.5 at a non-signalized trafficroundabout. Atmos Pollut Res 2010;1(2):59–70.

Johnson WB, Ludwig FL, Dabbert WF, Allen RJ. An urban diffusion simulation model forcarbon monoxide. JAPCA 1973;23:490–8.

Kono H, Ito S. A comparison of concentration estimates by the OMG volume-sourcedispersion model with three line source dispersion models. Atmos Environ1990;24B:253–60.

Kukkonen J, Walden J, Karppinen A, Lusa K. Evaluation of the dispersion model CAR-FMIagainst data from a measurement campaign near a major road. Atmos Environ2001;35:949–60.

Lin J, Ge YE. Impacts of traffic heterogeneity on roadside air pollution concentration.Transp Res D 2006;11:166–70.

1153S. Pandian et al. / Science of the Total Environment 409 (2011) 1145–1153

Luhar AK, Patil RS. A general finite line source mode for vehicular pollution prediction.Atmos Environ 1989;23:555–62.

Nakamura Y, Oke TR. Wind, temperature and stability condition in an East–Westoriented urban canyon. Atmos Environ 1988;22:2691–700.

Polus A, Lazar SS, Livneh M. Critical gap as a function of waiting time in determiningroundabout capacity. J Transp Eng 2003;129(5):504–9.

Sivacoumar R, Thanasekaran K. Line source model for vehicular pollution predictionnear roadways and model evaluation through statistical analysis. Environ Pollut1999;104:389–95.

Uehara K, Murakami S, Oikawa S, Wakamatsu S. Wind tunnel experiments on howthermal stratification affects flow in and above urban street canyons. AtmosEnviron 2000;34:1553–62.

US EPA. CAL3QHC version 2: a modeling methodology for predicting pollutantconcentrations near roadway intersections, US Environmental Protection Agency.Research Triangle Park, NC: Office of Air Quality Planning and Standards; 1995.

Yamartino RJ, Wiegand G. Development and evaluation of simple models for flow,turbulence and pollutant concentration fields within an urban street canyon.Atmos Environ 1986;20:2137–56.