Science of the Total Environment 538 (2015) 458–467

Contents lists available at ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitotenv

A method to estimate spatiotemporal air quality in an urban

traffic corridor
Nongthombam Premananda Singh, Sharad Gokhale ⁎
Department of Civil Engineering, Indian Institute of Technology Guwahati, Guwahati 781 039, India

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• The method predicts spatiotemporal air

quality in an urban traffic corridor.
• The prediction method is based on one
monitoring station in the traffic corridor.
• The method predicts spatial probability
of exceedances of CO reasonably well.

a r t i c l e i n f o a b s t r a c t

Article history: Air quality exposure assessment using personal exposure sampling or direct measurement of spatiotemporal
Received 6 April 2015 air pollutant concentrations has difficulty and limitations. Most statistical methods used for estimating
Received in revised form 21 July 2015 spatiotemporal air quality do not account for the source characteristics (e.g. emissions). In this study, a prediction
Accepted 12 August 2015
method, based on the lognormal probability distribution of hourly-average-spatial concentrations of carbon
Available online 27 August 2015
monoxide (CO) obtained by a CALINE4 model, has been developed and validated in an urban traffic corridor.
Editor: D. Barcelo The data on CO concentrations were collected at three locations and traffic and meteorology within the urban
traffic corridor.1 The method has been developed with the data of one location and validated at other two
Keywords: locations. The method estimated the CO concentrations reasonably well (correlation coefficient, r ≥ 0.96).
Urban air quality Later, the method has been applied to estimate the probability of occurrence [P(C ≥ Cstd] of the spatial CO
Human exposure concentrations in the corridor. The results have been promising and, therefore, may be useful to quantifying
Spatiotemporal spatiotemporal air quality within an urban area.
Concentration © 2015 Elsevier B.V. All rights reserved.
CO
Traffic emission

1. Introduction
⁎ Corresponding author.
E-mail address: sharadbg@iitg.ernet.in (S. Gokhale).
1 Road traffic is the main cause of poor air quality in urban areas (Colvile
Traffic corridor encompasses the entire passage bounded by the buildings on both
sides including roadway with sidewalks. et al., 2001; Ghose et al., 2004; Ramachandra and Shwetmala, 2009). This

http://dx.doi.org/10.1016/j.scitotenv.2015.08.065
0048-9697/© 2015 Elsevier B.V. All rights reserved.
 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467
fact has forced researchers around the world to focus on the associa-
tion of traffic-related air pollution and its impact on human health,
along or in the close-proximity of traffic corridors. Several studies
have reported that negative health impacts such as heart attacks,
cancers, asthma, decreased lung function, etc. are closely associated
with traffic induced air pollution (Gasana et al., 2012; Lipfert et al.,
2006; Newman et al., 2014; Wilker et al., 2013). For this reason,
evaluating human exposure in traffic corridors is of growing concern
to assess human health risks.
Since air quality changes spatially and temporally, exposure also
varies in space and time (Dons et al., 2011; Steinle et al., 2013). It is,
therefore, essential that spatiotemporal concentrations are known for
accurate assessment of human exposure in traffic corridors. Ambient
air quality measurements at fixed multi locations to indirectly estimate
the exposure, or direct measurement of exposure at the breathing level
of a few voluntary individuals using corridors (i.e. personal monitoring)
or using mobile monitoring vans (Durant et al., 2010; Padró-Martínez
et al., 2012) is laborious and expensive (Carr et al., 2002). Several
studies are carried out on modeling of spatiotemporal concentrations,
but most have employed statistical methods, which do not capture
temporal trends (i.e. changes over time) in sources and meteorology.
For example, approaches such as spatial interpolation, land-use regres-
sion (LUR), hierarchical method and spatial proximity are widely used
to estimate spatial concentrations (Chen et al., 2010; Crouse et al.,
2009; Dons et al., 2013; Li et al., 2013; McAdam et al., 2011; O'Leary
and Lemke, 2014; Wheeler et al., 2008). These methods have been the
preferred methods for estimating intra-urban and fine scale concentra-
tion gradients. While these approaches are common, other methods
including dispersion models and hybrid modeling approaches have
also been used to estimate spatial concentrations. For example, several
dispersion models are used to evaluate roadside air quality (Sharma
and Khare, 2001; Vardoulakis et al., 2003), which incorporate emission
and meteorology. CALINE4 (California line source dispersion model,
version4) (Benson, 1984) is one such model, which is well evaluated
and validated air quality model for modeling pollutant concentrations
from vehicular sources and is commonly used (Heist et al., 2013;
Levitin et al., 2005; Zhang and Batterman, 2010); A hybrid modeling
approach using CAL3QHCR dispersion model was developed to produce
spatiotemporal concentrations based on long-term spatial, temporal
means and residuals (Keller et al., 2015). A similar study was carried
out by Wilton et al. (2010), in which CAL3QHCR dispersion model was
combined with LUR model to improve the spatial concentrations.
Spatial interpolation technique, such as kriging, uses observed con-
centrations of many locations to predict concentrations at unknown
locations. The number of available stations and their locations limits
the spatial resolution (Whitworth et al., 2011). The method produces
unbiased estimates if expanded routine monitoring of spatial pollut-
ant concentrations is available. The LUR models require observed
concentrations as response variables with surrounding land features,
population density and traffic characteristics as predictor variables.
Similarly, the hierarchical model needs a high spatiotemporal reso-
lution of observed concentrations, as demonstrated by Li et al.
(2013) on nitrogen dioxide (NO2), and Pirani et al. (2014) on particulate
concentrations.
In traffic corridors, a direct measurement or adopting any of these
methods is even more complicated due to limited spatial domain and
turbulent wind flow pattern. In this study, we combined a line-source-
dispersion model with a probability distribution model to develop a
method for calculating spatial concentrations in a highly trafficked
urban traffic corridor. Line-source-dispersion models are used for
estimating roadside air quality that relates traffic emissions with the
pollutant concentrations under a specified meteorological condition.
However, dispersion models predict average concentrations well, i.-
e. middle range of the concentration distribution and poorly predict
the extreme ranges (Gokhale and Khare, 2005, 2007). Further, these
models predict spatial concentrations well for a single source
459
(de Hoogh et al., 2014). On the other hand, probability distribution
models capture the dispersion of concentration values resulting
from the dilution of the pollutant, which remains nearly same at
any spatial point within the corridor and fit well to the larger con-
centration distribution range. However, they are data based and do
not take into account the emissions and meteorology (Gokhale and
Khare, 2005, 2007). Therefore, dispersion models are reasonably
good in predicting average concentrations and probability models
represent extreme concentrations well as they capture stochastic
variability (Gokhale and Khare, 2005, 2007; Jakeman et al., 1988;
Marani et al., 1986; Ott, 1995). This attribute of the probability distri-
bution models may eliminate large deviations between estimated
and measured concentrations. Therefore, a suitable probability
distribution model combined with the well suited line-source-
dispersion model can provide better estimates of concentration
distributions.
We developed a prediction method by combining CALINE4 with a
lognormal distribution model to estimate hourly average spatiotempo-
ral carbon monoxide (CO) concentrations at one location and validated
at two spatial locations. The method has been applied in an urban traffic
corridor to determine spatial concentrations and to identify locations of
higher probability of occurrences of pollutant concentration exceeding
national standards within the corridor.
2. Methods
A suitable probability distribution model for the concentrations
measured at three locations has been identified by standard goodness-
of-fit tests of which the location parameter has been calibrated
with the output of the widely used CALINE4 model. The calibration
factor was empirically estimated from the dispersion model output
and the location parameter of the identified probability distribution
model. This factor accounts for the average deviation of estimated
from the measured concentrations and, thus, considerably reduces
uncertainty. The step-by-step procedure of the method has been
described below:
i) identification of three suitable locations in the traffic corridor for
monitoring of pollutant concentrations, meteorological parameters
and traffic volume, ii) identification of the suitable probability
distribution model from a range of probability models at all the
selected locations, iii) concentration estimation with CALINE4
model, iv) development of the prediction method by combining the
CALINE4 model output with the identified probability distribution
model output using a calibration factor at one location v) validation
of the prediction method at other two locations, and vi) application
of the prediction method to estimate the probability of exceedances
at spatial locations within the traffic corridor.
2.1. Urban traffic corridor, monitoring locations and measurements
A highly trafficked 400 m long urban traffic corridor in Guwahati,
India was selected. It houses numerous commercial activities, offices,
public utilities and attracts about 100,000 of traffic volume daily. Three
monitoring locations (L1, L2 and L3 as shown in Fig. 1) were selected.
The distance between L1 and L3 was 296 m, between L1 and L2 was
127 m, and 176 m between L3 and L2. The road was 16 m wide with a
separator of 1 m in between two lanes each side and the width of each
lane was 3.75 m. Another location on a building rooftop, 18 m high
and 69 m from the road, was selected within the corridor for installing
weather station. Videotaping for recording traffic volume was done
near L1 along with the measurement of pollutant concentrations.
The CO concentrations were measured during daytime from 7 am to
7 pm at all the locations for a period of one week including weekdays
and weekend days in March 2014. The measurements at L1 were carried
out from March 1 to March 7, at L2 from March 8 to March 15, and at L3
from March 18 to March 24, 2014. The CO meter (make: Delta Ohm,
460 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467

Fig. 1. The urban traffic corridor with monitoring locations.

model: HD37B17D) was used, which records concentrations at a
frequency of 3 s. Similarly, during the monitoring period, the meteoro-
logical parameters such as wind speed, wind direction, ambient
temperature, relative humidity and solar radiation were recorded by
a weather monitor (make: Davis weather, model: Vantage pro2).
The traffic volume survey was conducted by videotapes during the
monitoring for a period of one week, from March 1 to March 7, 2014.
The data were analyzed to determine 1-h average CO concentrations,
meteorological parameters, and traffic volume in different types
(two wheelers, three wheelers, cars, light duty commercial vehicles,
heavy duty diesel vehicles). The averaged CO concentrations resulting
zero were not considered in the analysis.
2.2. Identification of statistical distribution
A goodness-of-fit test was used to identify a suitable probability
distribution model, which fits best to the measured concentrations.
The measured 1-h average CO concentrations at L1, L2 and L3 were
utilized for this analysis. The probability distribution models such
as normal, lognormal, exponential, Gamma, Weibull, logistic, and
loglogistic were tested for the goodness-of-fit statistics using Dataplot
software (Heckert, 2001). The test includes the Kolmogorov–Smirnov
(KS) statistics, Anderson–Darling (AD) statistics, and Pearson correla-
tion coefficient (PCC). The lowest statistics of KS and AD, and the higher
of PCC indicated the better fit. These statistics have been used for such

Fig. 2. Measured concentrations at monitoring locations.

 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467 461

analysis in several studies (Gokhale and Khare, 2007; Lu, 2002; Marani
et al., 1986; P. Sharma et al., 2013; S. Sharma et al., 2013).
The KS statistic, Dn, is a non-parametric test, which compares
empirical cumulative distribution function (CDF) of measured data
with the theoretical CDF. The KS statistics is more sensitive at the middle
range of the distribution as compared to tails of the distribution. It is
defined by (Massey, 1951):

Dn ¼ maxn j F n ðxÞ−F 0 ðxÞj ð1Þ

where, Fn(x) and F0(x) represents the empirical and theoretical CDF,
respectively, Dn represents the maximum vertical distance between
empirical and theoretical CDF, and x represents the measured data.
The AD statistic is a modified version of KS, which gives more
sensitivity to the tails of distribution. It is defined by (Kottegoda and
Rosso, 1998): Fig. 3. Lognormal probability distributions of 1 h-average measured CO concentrations at
L1, L2, and L3.
Xn ð2:i−1Þ
A2 ¼ −n− ½ ln F 0 ðxi Þ þ ln f1− F 0 ðxnþ1−i Þg ð2Þ
i¼1 n
2.3. Estimation of concentration using CALINE4
where, x i represents the measured data at ith term in increasing
order, n the number of observations and F(xi) the theoretical CDF at xi. CALINE4 is a Gaussian line source dispersion model, which considers
The PCC statistics is based on the linear relationship between the the mixing zone over the roadway to account for thermal and mechan-
empirical and theoretical distributions. The values range between ical turbulences generated due to vehicles. The model has the capability
0 and 1, in which, 1 indicates perfect relationship and 0 indicates no to estimate pollutant concentrations of CO, nitrogen dioxide and partic-
relationship. It is defined by (Gokhale and Khare, 2007): ulate matter within 500 m of roadway, for a given meteorological
condition, composite emission source strength and site geometry
X X
X i ¼ 1nxi i ¼ 1nyi (e.g. roadway type and links) (Benson, 1984). The model is simple
i ¼ 1nxi :yi − and flexible but considers only local sources and predicts poorly for
r¼ v n
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
u X 2 X 2 low wind conditions. For example, a study reported the poor performance
ð Þð Þ
u i ¼ 1nxi i ¼ 1nyi
t X X when applied for street canyons and low wind speeds (e.g. b0.5 m/s)
i ¼ 1nxi 2 − i ¼ 1nyi 2 −
n n (Broderick et al., 2005; Holmes and Morawska, 2006).
ð3Þ The important input parameters for the model are meteorology re-
lated such as wind speed, wind direction, temperature, stability classes,
where, (xi) represents the empirical PDF, (yi) the theoretical PDF. and mixing height; emissions related such as traffic volume per hour,
With these statistics, 2-parameter probability distribution model composite emission factor; and site related such as link geometry, recep-
was identified that had the lowest KS, AD statistics and highest PCC. tor locations with heights, and road grades. The hourly stability classes
The detailed description about the goodness-of-fit test can be found were estimated using Richardson number and Pasquill–Gifford stability
elsewhere (Gokhale and Khare, 2007; Kottegoda and Rosso, 1998). (Uehara et al., 2000). Richardson number simulates atmospheric condi-
tions well as mixing of pollutants in urban areas occurs due to convec-
Table 1 tive and wind induced turbulences (Nakamura and Oke, 1988). The
Results of goodness-of-fit test of CO dataset at (a) L1 (b) L2 (c) L3. Bold values represent hourly mixing heights were obtained from the Indian Meteorological
the best goodness-of-fit statistics. Department (Attri et al., 2008). The hourly traffic emission factors
Distribution types KS-test AD-test PCC were estimated using observed traffic characteristics and hot (exhaust)
KS-value p-Value AD-Value p-Value
emissions and COPERT III (Computer Program to calculate Emissions
from Road Traffic) methodology (Ntzaichristos and Samaras, 1999).
(a)
This methodology utilizes speed-dependent equations for estimating
Normal 0.225 0.00 6.716 0.00 0.92
Lognormal 0.059 0.96 0.255 0.97 0.99 emissions of a vehicle type depending upon the vehicle characteristics.
Exponential 0.129 0.18 1.382 0.04 0.92
Gamma 0.109 0.36 1.296 0.03 0.99
2.4. Development of prediction method and validation
Weibull 0.093 0.56 1.034 0.01 0.98
Logistic 0.223 0.00 4.609 0.01 0.96
Loglogistic 0.106 0.40 0.314 0.25 0.97 The prediction method was developed by combining a suitable
range of dispersion model output by trimming upper and lower percen-
(b)
Normal 0.236 0.00 6.57 0.00 0.94 tiles with the identified probability distribution model. The suitable
Lognormal 0.089 0.62 0.402 0.73 0.99 range of the dispersion model output was determined from the statisti-
Exponential 0.125 0.22 1.523 0.05 0.92 cal performance by estimating the index-of-agreement (d) (P. Sharma
Gamma – – 1.619 0.00 0.99
et al., 2013; S. Sharma et al., 2013) and the correlation coefficient2 (r)
Weibull 0.104 0.43 1.438 0.30 0.97
Logistic 0.208 0.00 4.206 0.00 0.98
between the measured and modeled concentrations. The percentile
Loglogistic 0.361 0.00 0.428 0.25 0.96 concentrations were compared for several ranges using d and r statis-
tics. The range that had the higher d and r values was selected and
(c)
Normal 0.191 0.01 3.893 0.01 0.97
considered as the most suitable one. The d determines the magnitude
Lognormal 0.066 0.91 0.427 0.81 0.99 of agreement between measured and modeled values taking into
Exponential 0.072 0.84 0.360 0.89 0.85 account the sensitivity of differences between the values. The value of
Gamma – – 0.542 0.20 0.99
Weibull 0.074 0.82 0.581 0.66 0.96 2
The correlation coefficient carries the usual meaning as explained by Eq. (3) but in this
Logistic 0.167 0.04 2.083 0.00 0.99
case, it represents the linear relationship between the modeled and measured
Loglogistic 0.380 0.00 0.579 0.09 0.95
concentrations.
462 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467

Fig. 4. The comparison of lognormal probability distributions of the measured and Fig. 6. Comparison of CO concentrations lognormal distribution of CALINE4 and prediction
CALINE4 modeled 1 h-average CO concentrations at L1. method output at L1.

Fig. 5. The values of (a) index of agreement, d (b) Pearson correlation coefficient, r for different percentile rangesof the measured and CALINE4 modeled CO concentrations at L1.
 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467
Fig. 7. Comparison of the CO concentrations modeled by CALINE4, earlier hybrid method,
and the prediction method, with the measured concentrations at L1.
d has a range of 0 ≤ d ≤ 1, where 1 indicates prefect agreement between
measured and modeled concentrations and 0 indicates complete
disagreement. It is defined by Eq. (4) (Marmur and Mamane, 2003).
X 3.1. Lognormal distribution model
ðM e −M o Þ2
d ¼ 1− X   
Me −Mo  þ Me −Mo  2
where, Me and Mo represent measured and modeled concentration. The
bar over the notation represents the average value.
With the middle range concentrations estimated by CALINE4
model, the location and scale parameters of the identified probabil-
ity distribution model were estimated by using the method of
maximum likelihood estimation (MLE) (Mage and Ott, 1984).
The MLE maximizes the log-likelihood function, ln L(θ|y). The
likelihood function is given by Eq. (5) (Myung, 2003; Kottegoda
and Rosso, 1998).
LðθjyÞ ¼ f ðyjθÞ
where, L(θ|y) represents the likelihood of the parameter θ, given the
measured data y, which can be represented by probability distribution
model f(y|θ).
In this study, in addition to the combination of the models, a calibra-
tion factor has been applied to improve the estimates and thus, is a
modified method. The calibration factor was estimated by comparing
the respective location parameters (LP) of the probability distributions
of the modeled and measured concentrations. The modeled concentra-
tions were divided by this factor to re-estimate the concentrations.
This factor is the ratio between the LP of the modeled and measured
concentrations, as given by Eq. (6).
P mo
¼ Cf
P me
where, Pmo and Pme are the parameters of probability distribution of
the modeled and measured concentrations, respectively; and Cf is
the calibration factor. For lognormal distribution, Pmo = exp(Lmo)
and Pme = exp(Lme), where Lmo and Lme are the geometric mean and
Table 2
Statistics of prediction model performance at L1, L2, and L3.
Locations Performance statistics
Index-of-agreement (d) Pearson correlation coefficient (r)
L1 0.99 0.97
L2 0.98 0.96
L3 0.91 0.96
463
geometric standard deviation of the modeled and measured concentra-
tions, respectively.
The prediction method was developed at L1 and validated at L2 and
L3. Its performance was evaluated using the widely used statistical
measures such as Pearson correlation coefficient (r), normalized mean
square error (NMSE), fractional bias (FB), fractional variance (FS) and
d. The numerical value of r ranges from 1 to − 1, where 1 and − 1
represents perfect correlation and 0 indicates no correlation. The
NMSE ranges from 0 to ∞, where 0 represents perfect agreement. The
FB and FS ranges from − 2 to 2, where, − 2 represents extreme
under-prediction and + 2 extreme over-prediction, and d value (Cox
and Tikvart, 1990; Marmur and Mamane, 2003).
2.5. Application of the prediction method
The traffic corridor was divided into a number of grids and the pre-
diction method was applied to estimate the spatiotemporal probability
of occurrences of hourly average concentrations.
3. Results
ð4Þ The measured 1-h average CO concentrations from the three
monitoring locations, i.e. L1, L2 and L3 were used to identify the suitable
probability distribution model. The data were first put to statistical
quality check. Fig. 2 shows the important statistics of the measured con-
centrations at three locations. The distribution of the CO concentrations
was skewed from the normal and the coefficients of variations
were 1.22, 1.32, and 1.04 at L1, L2, and L3 respectively. It implied that
concentrations were temporally dispersed but location to location, the
variation was insignificant. Table 1 shows the results of the goodness-
of-fit tests. The CO concentrations at all the three locations fitted well
to the lognormal distribution (LND) model, with KS (p = 0.96), AD
(p = 0.97) at L1, similarly, KS (p = 0.91) and AD (p = 0.81) at L2,
and KS (p = 0.62), AD (p = 0.73) at L3 and PCC value was 0.99 at all
ð5Þ the locations. Therefore, LND was selected as the most suitable probabil-
ity distribution model. The differences in location parameters
(−0.9592, −1.090, −1.171 at L1, L2, L3, respectively) and scale param-
eters (1.107, 0.9957, 0.9194 at L1, L2, L3, respectively) were negligible
indicating that the CO concentration distributions were same at all the
selected locations. Fig. 3 shows the LND fit to the CO concentrations at
three locations.
3.2. CALINE4
Fig. 4 shows the measured and modeled concentrations at L1. The
model shows considerable deviation from the measured concentrations
in the lower and upper ranges. There was a deviation in the middle
range as well but was relatively less. The statistical indicators d, r,
NMSE, FB, and FS were found to be 0.68, 0.44, 0.87, 0.03, and − 0.30,
ð6Þ respectively, which show poor performance of the model.
To develop prediction method, the CALINE4 model performance was
tested by censoring the extreme ranges of its output of different percen-
tiles and comparing with the measured concentrations to obtain mini-
mum deviations. Fig. 5(a,b) show d and r for different percentile
ranges of measured and modeled concentrations, which were improved
Normalized mean square error (NMSE) Fractional bias (FB) Fractional variance (FS)
0.02 0.02 −0.07
0.05 0.23 0.05
0.22 0.33 0.49
464 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467

Fig. 8. Statistic of measured, CALINE4 and prediction method estimated concentrations at L1.

Fig. 9. The validation results of the prediction method and comparison with the CALINE4 modeled and measured concentrations at (a) L2, and (b) L3.
 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467 465

Fig. 10. Contours showing probability of exceedances greater than 3.5 ppm (NAAQS).

for the range of 5 to 90 percentiles, e.g. d N 0.9 from 0.68 and r N 0.8 from
0.44. Therefore, the percentile range of 5–90 was selected as the reliable
range to develop prediction method and the concentrations below 5
and above 90 percentiles were censored.
3.3. The prediction method
i) As found out, the range (i.e. 5–90 percentiles) of concentrations
of dispersion model at L1 was utilized to estimate the parameters
of the LND (i.e. location and scale parameters) using MLE. The
location and scale parameters were found to be − 0.7497 and
0.9380, respectively.
ii) The LND with these parameters (i.e. location and scale) was
applied back to estimate extreme CO concentrations for the
censored percentiles ranges of less than 5 and greater than 90.
Thus, the deviations at extreme ranges were reduced to minimal.
However, further in order to reduce deviation in the middle
range, a calibration factor was used to bring the entire range
closer to the measured concentrations.
iii) The calibration factor was calculated by comparing the ratio of
the respective distribution parameter (e.g. location parameters)
of the modeled entire range (including the 5–90 percentiles
and the revised extreme concentrations for ranges less than 5
and more than 90) and the entire range of measured concentra-
tions. The factor was found to be about 1.3.
iv) The 5–90 percentile range of the modeled concentrations was
corrected by the calibration factor to get a new entire range of
the modeled concentrations. Thus in this method, the middle as
well as the extreme ranges were improved.
v) This new range of the modeled concentrations was put into
LND model. Fig. 6 shows the comparison of the measured con-
centrations, CALINE4 concentrations and the concentrations
estimated with the prediction method. Fig. 7 shows the
comparison of the concentrations modeled by CALINE4 model,
earlier hybrid approach (Gokhale and Khare, 2005) and the
prediction method with the measured concentrations. The
results of the prediction method have been significantly
improved compared with the results of the earlier hybrid in
the middle range.
The prediction method produced considerably better estimates
of the concentrations than CALINE4 alone. The statistical indicators
between the measured and the prediction method concentrations
ascertained good match for the entire range as shown in Table 2,
i.e. d(0.99), r(0.97), NMSE(0.02), FB(0.02), and FS(− 0.07). These
statistics indicate that the prediction method minimized the error to
about 1%. Fig. 8 shows the graphical statistics of measured, CALINE4
and prediction method estimated concentrations at L1. It can be seen
that the statistics of measured and prediction method are quite similar
as compared to CALINE4.

Fig. 11. Contours showing probability of exceedances greater than 1 ppm.

466 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467
3.4. Validation
The prediction method was validated at L2 and L3 spatially located
within the traffic corridor (Fig. 1). Fig. 9(a,b) shows the comparison of
the measured and the prediction method concentrations at L2 and L3,
respectively. The CALINE4 concentrations have also been plotted
for comparison. It was evident that dispersion modeling to produce spa-
tiotemporal concentrations was not enough since it failed to estimate
the entire concentration distributions and deviated from the measured
concentrations. The prediction method provided the best estimates of
the concentrations for the entire range at L2 and L3, which validates
that the prediction method may be applied to estimate concentration
distributions to spatial locations. The statistics as shown in Table 2,
e.g. d, r, NMSE, FB, and FS were 0.98, 0.96, 0.05, 0.23, 0.05 at L2 and
0.91, 0.96, 0.22, 0.33, 0.49 at L3, respectively indicate that the prediction
method estimates the concentrations with more accuracy even at far
spatial locations within the traffic corridor.
3.5. Spatiotemporal air quality
The receptors of 3 m height were selected at 10 m and 20 m from the
center of the road on both sides for every 10 m along the corridor
(Fig. 1). The prediction method was applied to estimate hourly concen-
trations at those receptors. The probability of occurrences greater than
the national ambient air quality standard (NAAQS) of 3.5 ppm (CPCB,
2010) was estimated at each receptor. For convenience in identification
of locations, two sides of the road were named as north-east side (NES),
where L1 and L3 located, and south-west side (SWS), where L2 located.
Fig. 10 shows the probability of occurrence of hourly concentrations
exceeding the NAAQS value (e.g. 3.5 ppm). The 2% to 3% probability of
occurrence indicates that at least once a week the concentrations
exceeded the NAAQS value (i.e. 12 samples each day for a week,
12 × 7 = 84 samples, therefore 0.02% of 84 samples = 1.68), while 3%
to 4% means it exceeded at least twice a week (i.e. 0.03 × 84 = 2.52).
Fig. 11 shows the probability of occurrence N 1 ppm. At about 10 m
from the road center, it was 12% to 15% at NES and 18% to 21% at SWS,
whereas, 15% to 18% at other locations. This means the likelihood of con-
centrations N 1 ppm for 10 h in a week was at about 10 m from the road
centerline on SWS, and for at least 15 h at NES, and at other locations, it
occurred for at least 12 h.
4. Discussion
In this paper, a method for predicting spatiotemporal concentrations
in traffic corridors has been presented. The prediction method has been
developed using observed concentrations, dispersion model output at
one monitoring location and the suitable probability distribution
model identified in the corridor, and validated at two spatial locations.
The method has been applied to produce probability of exceedances
of concentrations over the national ambient air quality standards.
The earlier hybrid modeling technique (Gokhale and Khare, 2005;
P. Sharma et al., 2013; S. Sharma et al., 2013; Taylor et al., 1985) was
also applied and compared with the prediction method, which shows
large deviations in the concentrations predicted by the hybrid
technique even within the middle range (Fig. 5b). With the use of a cal-
ibration factor in the prediction method, a considerable improvement
was observed. The results have, therefore, demonstrated that the
prediction method is a modified hybrid modeling technique, which
provides better estimates of the entire concentration distributions.
The method is useful since its important component i.e. probability
distribution model, filters out the measurement errors as well as
reduces the deviation in the modeled concentrations (dispersion
model). Probability distribution models capture the stochastic variabil-
ity of pollutant concentrations arising due to influencing factors such as
wind-flow pattern, emission pattern, urban features, and traffic-flow
pattern (Gokhale and Khare, 2007; Jakeman et al., 1988).
From the study, it has been observed that the probability distribution
of a pollutant remains same within the traffic corridor and, therefore,
the prediction method may be applied at spatial locations using the
meteorology of the urban area, traffic volume of the corridor of interest
and pollutant concentrations of at least one location in the corridor. The
prediction method, therefore, is a modified method and can provide
better estimates of the concentration distributions and proves to be
economical as compare to the statistical methods discussed in the
introduction, which need concentrations data on multiple locations.
However, the promising results of the study have limitations
because it was assumed that the CO concentrations were originated
from traffic alone and other sources were not considered; and the
locations were all within the same corridor. Therefore, further analysis
is required to investigate whether the prediction method will produce
similar results when applied to the other road segments within the
urban area.
5. Conclusion
The prediction method is simple, needs less number of ambient
monitoring stations to produce a reliable estimate of concentration
distributions at spatial locations in a traffic corridor. The important com-
ponent of the method is probability distribution, which incorporates
spatial distribution of concentrations arising due to local wind-flow,
urban feature, changes in traffic-flow pattern. It also proved to be
better than the existing hybrid modeling approach and, therefore, is a
modified hybrid modeling methodology. Further investigation of the
method to determine its applicability on other road segments of the
urban area is needed.
References
Attri, S.D., Siddhartha, S., Mukhopadhyay, B., Bhatnagar, A.K., 2008. Atlas of Hourly Mixing
Height and Assimilative Capacity of Atmosphere in India. Indian Meteorological
Department Publication on Environmental Meteorology, Government of India.
Benson, P.E., 1984. CALINE4- A Dispersion Model For Predicting Air Pollutant
Concentrations Near Roadways. Office of Transportation Laboratory, California
Department of Transportation, Sacramento, USA (Report No. FHWA/CA/TL-84/15).
Broderick, B.M., Budd, U., Misstear, B.D., Ceburnis, D., Jennings, S.G., 2005. Validation of
CALINE4 modeling for carbon monoxide concentrations under free-flowing and
congested traffic conditions in Ireland. Int. J. Environ. Pollut. 24, 104–113.
Carr, D., von Ehrenstein, O., Weiland, S., Wagner, C., Wellie, O., Nicolai, T., von Mutius, E.,
2002. Modeling annual benzene, toluene, NO2, and soot concentrations on the basis
of road traffic characteristics. Environ. Res. 90, 111–118.
Chen, L., Bai, Z., Kong, S., Han, B., You, Y., Ding, X., Du, S., Liu, A., 2010. A land use regression
for predicting NO2 and PM10 concentrations in different seasons in Tianjin region,
China. J. Environ. Sci. 22, 1364–1373.
Colvile, R.N., Hutchinson, E.J., Mindell, J.S., Warren, R.F., 2001. The transport sector as a
source of air pollution. Atmos. Environ. 35, 1537–1565.
Cox, W.M., Tikvart, J.A., 1990. A statistical procedure for determining the best performing
air quality simulation model. Atmos. Environ. Part A 24, 2387–2395.
CPCB (Central Pollution Control Board), 2010. National Ambient Air Quality Status and
Trends in India. P.R. Division, Ministry of Environment and Forests, Delhi, India
http://www.cpcb.nic.in (January 2012).
Crouse, D.L., Goldberg, M.S., Ross, N.A., 2009. A prediction-based approach to modeling
temporal and spatial variability of traffic-related air pollution in Montreal, Canada.
Atmos. Environ. 43, 5075–5084.
de Hoogh, K., Korek, M., Vienneau, D., Keuken, M., Kukkonen, J., Nieuwenhuijsen, M.J.,
Badaloni, C., Beelen, R., Bolignano, A., Cesaroni, G., Pradas, M.C., Cyrys, J., Douros, J.,
Eeftens, M., Forastiere, F., Forsberg, B., Fuks, K., Gehring, U., Gryparis, A., Gulliver, J.,
Hansell, A.L., Hoffmann, B., Johansson, C., Jonkers, S., Kangas, L., Katsouyanni, K.,
Künzli, N., Lanki, T., Memmesheimer, M., Moussiopoulos, N., Modig, L., Pershagen,
G., Probst-Hensch, N., Schindler, C., Schikowski, T., Sugiri, D., Teixidó, O., Tsai, M.-Y.,
Yli-Tuomi, T., Brunekreef, B., Hoek, G., Bellander, T., 2014. Comparing land use regres-
sion and dispersion modeling to assess residential exposure to ambient air pollution
for epidemiological studies. Environ. Int. 73, 382–392.
Dons, E., Int Panis, L., Van Poppel, M., Theunis, J., Willems, H., Torfs, R., Wets, G., 2011.
Impact of time–activity patterns on personal exposure to black carbon. Atmos.
Environ. 45, 3594–3602.
Dons, E., Van Poppel, M., Kochan, B., Wets, G., Int Panis, L., 2013. Modeling temporal and
spatial variability of traffic-related air pollution: hourly land use regression models
for black carbon. Atmos. Environ. 74, 237–246.
Durant, J.L., Ash, C.A., Wood, E.C., Herndon, S.C., Jayne, J.T., Knighton, W.B., Canagaratna,
M.R., Trull, J.B., Brugge, D., Zamore, W., Kolb, C.E., 2010. Short-term variation in
near-highway air pollutant gradients on a winter morning. Atmos. Chem. Phys. 10,
5599–5626 (Print).
 N.P. Singh, S. Gokhale / Science of the Total Environment 538 (2015) 458–467
Gasana, J., Dillikar, D., Mendy, A., Forno, E., Ramos Vieira, E., 2012. Motor vehicle air
pollution and asthma in children: a meta-analysis. Environ. Res. 117, 36–45.
Ghose, M.K., Paul, R., Banerjee, S.K., 2004. Assessment of the impacts of vehicular
emissions on urban air quality and its management in Indian context: the case of
Kolkata (Calcutta). Environ. Sci. Pol. 7, 345–351.
Gokhale, S., Khare, M., 2005. A hybrid model for predicting carbon monoxide from
vehicular exhausts in urban environments. Atmos. Environ. 39, 4025–4040.
Gokhale, S., Khare, M., 2007. Statistical behavior of carbon monoxide from vehicular
exhausts in urban environments. Environ. Model. Softw. 22, 526–535.
Heckert, N.A., 2001. Dataplot website at the National Institute of Standards and
Technology. http://www.itl.nist.gov/div898/software/dataplot.html (June).
Heist, D., Isakov, V., Perry, S., Snyder, M., Venkatram, A., Hood, C., Stocker, J., Carruthers, D.,
Arunachalam, S., Owen, R.C., 2013. Estimating near-road pollutant dispersion:
a model inter-comparison. Transp. Res. Part D: Transp. Environ. 25, 93–105.
Holmes, N.S., Morawska, L., 2006. A review of dispersion modeling and its application to
the dispersion of particles: an overview of different dispersion models available.
Atmos. Environ. 40, 5902–5928.
Jakeman, A.J., Simpson, R.W., Taylor, J.A., 1988. Modeling distributions of air pollutant
concentrations—III. The hybrid deterministic-statistical distribution approach.
Atmos. Environ. 22, 163–174.
Keller, J.P., Olives, C., Kim, S.-Y., Sheppard, L., Sampson, P.D., Szpiro, A.A., Oron, A.P.,
Lindström, J., Vedal, S., Kaufman, J.D., 2015. A unified spatiotemporal modeling
approach for predicting concentrations of multiple air pollutants in the multi-ethnic
study of atherosclerosis and air pollution. Environ. Health Perspect. 123, 301–309.
Kottegoda, N.T., Rosso, R., 1998. Statistics Probability and Reliability for Civil and
Environmental Engineers. The McGraw Hills Companies Inc.
Levitin, J., Härkönen, J., Kukkonen, J., Nikmo, J., 2005. Evaluation of the CALINE4 and CAR-
FMI models against measurements near a major road. Atmos. Environ. 39, 4439–4452.
Li, L., Wu, J., Ghosh, J.K., Ritz, B., 2013. Estimating spatiotemporal variability of ambient air
pollutant concentrations with a hierarchical model. Atmos. Environ. 71, 54–63.
Lipfert, F., Wyzga, R., Baty, J., Miller, J., 2006. Traffic density as a surrogate measure of
environmental exposures in studies of air pollution health effects: Long-term
mortality in a cohort of US veterans. Atmos. Environ. 40, 154–169.
Lu, H.-C., 2002. The statistical characters of PM10 concentration in Taiwan area. Atmos.
Environ. 36, 491–502.
Mage, D.T., Ott, W.R., 1984. An evaluation of the methods of fractiles, moments and
maximum likelihood for estimating parameters when sampling air quality data
from a stationary lognormal distribution. Atmos. Environ. 18, 163–171.
Marani, A., Lavagnini, I., Buttazzoni, C., 1986. Statistical study of air pollutant concentrations
via generalized gamma distributions. J. Air Pollut. Control Assoc. 36, 1250–1254.
Marmur, A., Mamane, Y., 2003. Comparison and evaluation of several mobile-source and
line-source models in Israel. Transp. Res. Part D: Transp. Environ. 8, 249–265.
Massey, F.J., 1951. The Kolmogorov–Smirnov test for goodness of fit. J. Am. Stat. Assoc.
46, 68–78.
McAdam, K., Steer, P., Perrotta, K., 2011. Using continuous sampling to examine the
distribution of traffic related air pollution in proximity to a major road. Atmos.
Environ. 45, 2080–2086.
Myung, I.J., 2003. Tutorial on maximum likelihood estimation. J. Math. Psychol.
47, 90–100.
Nakamura, Y., Oke, T.R., 1988. Wind, temperature and stability conditions in an east–west
oriented urban canyon. Atmos. Environ. 22, 2691–2700.
467
Newman, N.C., Ryan, P.H., Huang, B., Beck, A.F., Sauers, H.S., Kahn, R.S., 2014.
Traffic-related air pollution and asthma hospital readmission in children: a longitudinal
cohort study. J. Pediatr. 164, 1396–1402 (e1391).
Ntzaichristos, L., Samaras, Z., 1999. COPERT III. Computer Programme to Calculate
Emissions From Road Transport. Methodology and emissions factors (Version 2.1),
EEA, Technical Report No. 49, Copenhagen.
O'Leary, B.F., Lemke, L.D., 2014. Modeling spatiotemporal variability of intra-urban air
pollutants in Detroit: a pragmatic approach. Atmos. Environ. 94, 417–427.
Ott, W.R., 1995. Environmental Statistics and Data Analysis. Lewis Publishers.
Padró-Martínez, L.T., Patton, A.P., Trull, J.B., Zamore, W., Brugge, D., Durant, J.L., 2012.
Mobile monitoring of particle number concentration and other traffic-related air
pollutants in a near-highway neighborhood over the course of a year. Atmos. Environ.
61, 253–264.
Pirani, M., Gulliver, J., Fuller, G.W., Blangiardo, M., 2014. Bayesian spatiotemporal
modeling for the assessment of short-term exposure to particle pollution in urban
areas. J. Expo. Sci. Environ. Epidemiol. 24, 319–327.
Ramachandra, T.V., Shwetmala, 2009. Emissions from India's transport sector: statewise
synthesis. Atmos. Environ. 43, 5510–5517.
Sharma, P., Khare, M., 2001. Modeling of vehicular exhausts — a review. Transp. Res. Part
D: Transp. Environ. 6, 179–198.
Sharma, P., Sharma, P., Jain, S., Kumar, P., 2013. An integrated statistical approach for
evaluating the exceedence of criteria pollutants in the ambient air of megacity
Delhi. Atmos. Environ. 70, 7–17.
Sharma, S., Sharma, P., Khare, M., 2013. Hybrid modeling approach for effective
simulation of reactive pollutants like Ozone. Atmos. Environ. 80, 408–414.
Steinle, S., Reis, S., Sabel, C.E., 2013. Quantifying human exposure to air pollution—moving
from static monitoring to spatio-temporally resolved personal exposure assessment.
Sci. Total Environ. 443, 184–193.
Taylor, J.A., Simpson, R.W., Jakeman, A.J., 1985. A hybrid model for predicting the
distribution of pollutants dispersed from line sources. Sci. Total Environ. 46, 191–213.
Uehara, K., Murakami, S., Oikawa, S., Wakamatsu, S., 2000. Wind tunnel experiments on
how thermal stratification affects flow in and above urban street canyons.
Atmos. Environ. 34, 1553–1562.
Vardoulakis, S., Fisher, B.E.A., Pericleous, K., Gonzalez-Flesca, N., 2003. Modeling air
quality in street canyons: a review. Atmos. Environ. 37, 155–182.
Wheeler, A.J., Smith-Doiron, M., Xu, X., Gilbert, N.L., Brook, J.R., 2008. Intra-urban
variability of air pollution in Windsor, Ontario—measurement and modeling for
human exposure assessment. Environ. Res. 106, 7–16.
Whitworth, K., Symanski, E., Lai, D., Coker, A., 2011. Kriged and modeled ambient air
levels of benzene in an urban environment: an exposure assessment study. Environ.
Heal. 10, 1–10.
Wilker, E.H., Mostofsky, E., Lue, S.-H., Gold, D., Schwartz, J., Wellenius, G.A., Mittleman,
M.A., 2013. Residential proximity to high-traffic roadways and poststroke mortality.
J. Stroke Cerebrovasc. Dis. 22, 366–372.
Wilton, D., Szpiro, A., Gould, T., Larson, T., 2010. Improving spatial concentration
estimates for nitrogen oxides using a hybrid meteorological dispersion/land use
regression model in Los Angeles, CA and Seattle, WA. Sci. Total Environ. 408,
1120–1130.
Zhang, K., Batterman, S., 2010. Near-road air pollutant concentrations of CO and PM2.5:
a comparison of MOBILE6.2/CALINE4 and generalized additive models. Atmos.
Environ. 44, 1740–1748.