Use of Bayesian Inference Method To Model Vehicular Air Pollution in Local Urban Areas
Use of Bayesian Inference Method To Model Vehicular Air Pollution in Local Urban Areas
Use of Bayesian Inference Method To Model Vehicular Air Pollution in Local Urban Areas
A R T IC LE I N F O ABS TRA CT
Keywords: Traffic Related Air Pollution (TRAP) studies are usually investigated using different categories
Vehicular air pollution such as air pollution exposure for health impacts, urban transportation network design to miti-
Artificial intelligence gate pollution, environmental impacts of pollution, etc. All of these subfields often rely on a
Predictive modelling robust air pollution model, which also necessitates an accurate prediction of future pollutants. As
GIS
is widely accepted by the heath authorities, TRAP is considered to be the major health issue in
urban areas, and it is difficult to keep pollution at harmless levels if the time sequenced dynamic
pollution and traffic parameters are not identified and modelled efficiently. In our work here,
artificial intelligence techniques, such as Bayesian Networks with an optimized configuration, are
used to deliver a probabilistic traffic data analysis and predictive modelling for air pollution
(SO2, NO2 and CO) at very local scale of an urban region with up to 85% accuracy. The main
challenge for traditional data analysis is a lack of capability to reveal the hidden links between
distant data attributes (e.g. pollution sources, dynamic traffic parameters, etc.), whereas some
subtle effects of these parameters or events may play an important role in pollution on a long-
term basis. This study focuses on the optimisation of Bayesian Networks to unveil hidden links
and to increase the prediction accuracy of TRAP considering its further association with a pre-
dictive GIS system.
1. Introduction
The term air pollution is linked to harmful substances in the air such as gases, solid particles or liquid droplets. The more
specifically, traffic related air pollutants may be categorised as primary and secondary pollutants where the first category incudes
carbon monoxide, sulphur dioxide, nitrogen oxides (Alvarez-Vazquez et al., 2017) and also Hydrocarbons, particulate matter, mobile-
source air toxics such as lead, benzene, and formaldehyde (Health Effects Institute, 2010) whereas the second category covers ozone
and the other minor pollutants. The pollutant emission measures are usually described as the weight of pollutant divided by a unit
weight, volume, distance or duration of the pollution activity (US Environmental Pollution Agency, 1995). The air pollution is
automatically measured by Air quality stations located at the critical points of urban areas where population or traffic flow is at
certain density.
Nowadays with the increasing population of different vehicle types and by inadequate traditional transport or traffic system
designs, air pollution has become one of the key issues to be solved urgently for urban areas. Traffic related air pollution (TRAP),
which contains harmful chemicals, is a major threat for cities. In dense urban areas, vehicle emissions may be responsible for 90–95%
of carbon monoxide and 60–70% of nitrogen oxides within the atmosphere (Schwela, 2000). Recent epidemiological studies suggest
⁎
Corresponding author.
E-mail address: [email protected] (A. Orun).
https://doi.org/10.1016/j.trd.2018.05.009
that at present air pollution results in an average 7-month reduction in life expectancy and costs UK society up to £20 billion per year
(HCEA Committee, 2011).
The immediate solution is not easy if the time sequenced dynamic pollution and traffic system parameters are not properly
identified and modelled. Particularly multidisciplinary areas such as artificial intelligence (AI) methods (e.g. data mining, inference
methods, etc.), state-of-art instrumentation, supercomputing facilities, distributed sensors, etc. would be expected to bring the most
promising solutions to the problem. In our work here an artificial intelligence technique, Bayesian Networks, are used for a prob-
abilistic data analysis whose performance has already been proven by our previous works (Orun and Aydin, 2010; Orun, 2004). One
of the traditional common issues of manual data analysis is the lack of visibility of the hidden links between distant and the least
correlated data attributes (e.g. pollution sources, dynamic traffic parameters, geographic location characteristics, etc.); whereas some
subtle effects of these parameters or specific combinations of them may play an important role in traffic related air pollution on a
long-term basis. The proposed work here targets an optimised Bayesian model to unveil such hidden links between the TRAP factors
and parameters.
Many works have previously investigated air pollution. The most similar methodology was used with Bayesian classifier by Corani
and Scanagatta (2016), but only for the prediction of PM2.5 (Atmospheric aerosol particles) pollutant and O3 (Ozone gas). In another
work Karatzas and Kaltsatos introduce a computational intelligence method (Karatzas and Kaltsatos, 2007) for air pollution mod-
elling by which the environmental system is simulated. Their work was also at a large geographic scale for a city area. Zhu et al.
(2015) investigate traffic-related air pollution in a street canyon by utilising a genetic algorithm-back propagation artificial neural
network but even though it introduces an AI-based approach, it is only based on a single pollutant parameter, nitrogen dioxide (NO2),
rather than focusing on a multi parameter solution. In (Passow et al. 2012) authors also used NO2 used as a tracer for pollution but
included additional factors derived from meteorological data, traffic data and earth observation data to create framework for near-
real time traffic management and air quality control. In contrast, the method introduced in this paper is dedicated to long-term
forecast to support design and re-assessment of air quality models in local urban areas. The method proposed here will also include
several diverse types of parameters (pollutant, environmental, etc.), processed in an interactive form, to identify subtle connection
between them.
On the other hand, geographic information systems (GIS) (Orun, 1993) are one of the major tools that could also potentially be
used for the transport planning (Niemeier and Beard, 1993). One of the earlier works (Thonga and Wongb, 1997) introduced a
specific database design for use by GIS for urban transport planning. In the work, even though some “predictive” techniques (called
what-if) were suggested, their implementation was not considered within the proposed system. Another statistical predictive model
based on a “distance decay regression” approach was introduced by Jason et al. (2009). Even though the method effectively used
statistical algorithms with high prediction accuracy, it was limited with NOx gases and no GIS utility was used or evaluated for a
pollution map formation. The majority of past studies Jason et al. (2009), Matthias et al. (2006), Zhu et al. (2015), and Shekarrizfard
et al. (2017) were limited to NOx pollutants, whereas in our work the additional SO2 and CO based pollutions were also studied.
Only a few earlier works combined the prediction methods with GIS for a pollution analysis. A direct estimation of traffic related
air pollution was made by a GIS based component-oriented integrated system which was developed by Rebolj and Sturm (1999). In
the paper, the generic outline of proposed system was introduced. However, the presentation did not exhibit the numerical test results
in details other than a single output. The current statistical traffic-related air pollution models (Gulliver and Briggs, 2005) used for
traffic design may cause several problems such as, Available data may not provide accurate pollution information for the future if
processed by inappropriate linear predictive models. Such inappropriate modelling may later cause bulky re-construction and de-
velopment for road network amendments.
Within this work we overcome the above issues by deploying a predictive modelling based GIS system in which AI technique,
Learning Bayesian Networks, is used for a new target layer generation.
This paper is organised as follow; The method and materials used here are introduced in Section 2, with a presentation of a limited
data sample set. In the subsections Bayesian Networks and predictive Bayesian model are also introduced by the presentation of a
flow chart and pseudo code of the proposed method. The results of model accuracy are finally represented in Section 3 as results and
discussion followed by the conclusion section.
The restricted data set consists of weekly recordings of traffic flows (at local data collection stations), air pollution values (e.g.
SO2, NO2, CO), local temperature readings, wind records, air pressure, rainfall and global radiation values within Leicester City local
urban areas for the year 2012. The local regions selected in this work are Newark and Aunsite. The whole data set was utilised for
Bayesian Network (BN) construction as seen in Fig. 3, where the abbreviations “st” refers to traffic data collection stations in the
Leicester city area. The traffic flow data were collected over the 56 station locations with additional 9 parameters including pollution
types, temperature, global radiation, air pressure, rainfall, wind speed and wind direction.
Parameter units in Table 1: (NO2, SO2) = parts per billion, CO = parts per million, Temperature = °C, Global Radiation = W/m2,
Air pressure = mbar, Wind direction = Degree.M, Wind speed = m/s, Rainfall = mm/h, traffic flow (sti) = number of vehicles/h.
The collected raw data need additional processing in which different fractions of data segments with different data types were re-
arranged so they can be integrated within this work. This inevitably caused a limitation of data volume and relatively limited training
of the BN system at low efficiency.
237
A. Orun et al. Transportation Research Part D 63 (2018) 236–243
Table 1
Display of a sample for the traffic data set. The whole set contains total number of 56 stations (only station “st1” is shown below). Station values
correspond to traffic flow at the specific city locations.
Date NO2_aunsite CO_Newarke SO2_Newarke Temperature Global_radiation Air_pressure Wind_direction Wind_speed Rainfall St1
In general terms, Bayesian Networks (BN) or Casual Probabilistic Networks are very useful techniques which can achieve efficient
knowledge representation and reasoning. They are also capable of generating very accurate classification results under uncertainty
where the data set may include many uncertain conditions (Koski and Noble, 2009). Bayesian Networks graphically encode and
represent the conditional independence (CI) relationships among a set of data (Orun, 2004). In this work, a learning Bayesian
Network software tool (PowerConstructor™) (Cheng et al., 2002) is used for the analysis of air pollution, traffic and environmental
data and the Bayesian inference to construct the network (Fig. 3). The algorithm examines information flow between two highly
related variables (attributes) from a data set and decides if these variables (e.g. traffic parameters) are independent or linked and it
also investigates into how close the relationship between those variables is. This process continues for all variables in turn in the data
set. This information flowed between the two variables is called conditional mutual information of two variables Xi, Xj which may be
denoted as:
P (x i,x j |c )
I (Xi ,Xj |C ) = ∑ P (x i,x j,c )log
P (x i |c ) P (x j |c )
xi,xj,c (1)
In Eq. (1), C is a set of nodes and c is a vector (one instance of variables in C). If I(Xi,Xj|C) is smaller than a certain threshold t, then we
can say Xi and Xj are conditionally independent. In the equations P(xi, xj|c) may be extracted from the conditional probability tables,
where i, j are indices for the different cases of two variables whose link is investigated.
One early example in which a Bayesian approach for an analysis of air pollution data was introduced by Suggs and Curran (1983).
In their work air pollution data and air quality standards were compared and a combination of pollution history with instrumental
precision in a Bayesian probabilistic model was comprehensively discussed. Some of our previous works also focused on the different
application fields of Bayesian inference method and classification process separately, which provide a useful guidance for this work
(Orun, 2004; Orun and Aydin, 2010). In those works two different experiments were done by use of a Bayesian Network tools such as
PowerPredictor™ (Cheng et al., 2002) for the analysis of data produced by the real-time lab experiments. The other Bayesian package
used here is PowerConstructor (Cheng et al., 2002) which is a different tool than Bayesian classifier as it only exhibits the links
between the attributes in a semantic graphical domain. But both utilities use the Markov condition to obtain a collection of con-
ditional independence statements from the networks (Pearl, 1988). One of the advantages of Bayesian Networks over the other AI
systems (e.g. neural networks or fuzzy logic) is that, it identifies direct and indirect links between the attributes which can be easily
interpreted. In both utilities, the algorithms establish the links between the attribute nodes and show them graphically. The nodes
that are not connected are not necessarily excluded from the process, but it means that their contribution would not improve the
classification results further (Cheng et al., 2002).
The inference characteristics of BN is concerned with the interpretation of data flow between the network attributes and es-
tablishing the links between those whose data flow exceed the certain threshold then they are automatically assigned as relevant. The
BN prediction process is achieved by two steps (a) training stage (learning) and (b) test stage (cross-validation) where in first stage
the network is trained by the data to establish rule-based model with the selected network parameters (e.g. data flow threshold,
discretisation method, etc.) and in the second stage the network processes the test data to achieve a classification. At BN training
stage, the cases (rows) in training data set are expected to be independent from each other to avoid any overfitting problem which
means, BN system works well with available data but outperformed with another data set. Overfitting problem is most common issue
of the classifiers but Bayesian networks have a distinctive advantage over the other classifiers such as Neural networks (Iruansi et al.,
2012)
Geographic information systems (GIS) are very popular multi-layer spatial data storage, analysing and presentation tools. By the
optimised integration of GIS (shown in Fig. 2) and Bayesian Networks for an intelligent query via the system layers (e.g. temperature,
pollutions, topography, etc.), the development of a predictive model is possible. Within this work a predictive model is created in
238
A. Orun et al. Transportation Research Part D 63 (2018) 236–243
Fig. 1. The pseudo code of a new GIS target layer generation loop by Bayesian classifier based predictive model (on the left) and its flow chart
representation (on the right). In the pseudo code a process loop is defined by Do-Return commands whereas exit command indicates the completion
of whole process.
which some critical target parameters (e.g. CO, SO2, NO2 pollutions) would be predicted in advance on a specific (city) location. Such
models could be invaluable at the planning stage for urban transportation networks designers as the outcomes from the model can
assist long-term planning to avoid later high-cost amendments, bulky re-construction of road or infrastructural network, or to avoid
hazardous situation (e.g. breaching the levels of air quality gas concentrations). The predictive model process uses an iterative
prediction loop to classify each point of the target layer (see Fig. 1) whose parametric data are provided to form a complete layer, but
its accuracy would depend on density of data collection stations distribution. In case of low distribution, the necessary interpolations
between the stations must be done. This iterative process can be described by the pseudo code and the flowchart shown in Fig. 1. The
Bayesian classifier used in the iterative process is fundamentally based on the Bayesian theorem that is described by the formula:
P (B|A) P (A)
P (A|B ) =
P (B ) (2)
where P(A) and P(B) are probability of observation of A and B, respectively. P(A|B) probability of observing A given that B is true. P
(B|A) is the probability of observing B given that A is true. This principal is related to the Bayesian classifier cited in Fig. 1.
As is seen in the Bayesian Network (Fig. 3) which was built after Bayesian inference based data processing by use of the package
Fig. 2. Traditional query based GIS system (on the left) and predictive model based GIS (on the right) in which a prediction of each sample layer
(CO) point is calculated. In our experiments, the sample prediction has been made with 81% accuracy with the limited data set by use of Bayesian
classifier.
239
A. Orun et al. Transportation Research Part D 63 (2018) 236–243
Link 1
Link 2
Link 3
Link 4
Fig. 3. Established Bayesian Network (BN) configuration created with the use of inference tool, PowerConstructor™ to connect links between the
data attributes. The network connection threshold was kept at min level (t = 0.1) to maximize the number of links between the nodes. As is seen in
the figure, different cause-effect relations of the attributes are represented by different link characteristics (links 1–4).
PowerConstructor™, the following conclusions would be drawn to interpret the links between the attributes in the network.
• Traffic data collection Station56 (shown as St56) has a vital role as it has 7 connections with other stations (for vehicle flow data).
This means that any structural change on Station56 would have substantial effect on the other stations. (Link 1)
• Air pollution attributes have natural links with some parameters like: temperature, rainfall, CO_Newarke, air pressure, wind
speed, NO . This will lead to a natural modelling to be associated with the main traffic air pollution model (Link2, Link 3)
2
• CO in Newarke area has a link with traffic Station33, NO Aunsite area has link with Station10, SO Newarke area has link with
2 2
Station59, it has also link with wind direction whose cause-effect relationships would need an efficient interpretation before a
modelling (Link 4)
• The stations (Sti) for vehicle flow data collection have links with each other’s, which may give an idea about the interactions
between the high-density vehicle flow regions in the city. (Link 1)
• One of examples for a hidden interaction would be between temperature and CO pollution. Which may help in traffic planning to
bring some restriction on CO emission (e.g. by speed reduction in warm days of the year) (Link 3)
The attributes’ cause-effect connections can be further exhibited as demonstrated in Figs. 2–5 by the separate local links. Note that
the direction of arcs in Figs. 1–5 should to be ignored as the formation of arrows are due to software utility characteristics and have
no specific function other than link.
The effectiveness of artificial intelligence techniques which support data interpretation is the most essential and challenging
St33
CO_Newarke Temperature
Fig. 4. Traffic flow effect of Station33 on CO pollution in Newark area with an influence of temperature variations.
240
A. Orun et al. Transportation Research Part D 63 (2018) 236–243
St19
St10
NO2_Aunsite Wind_speed
Fig. 5. Traffic flow effects of stations St19 and St10 on NO2 pollution in Aunsite area with the influence of wind speed variations.
factor for the data analysis which would otherwise not be possible to achieve manually. In our work, local link analyses have been
made as shown in Figs. 4–7 where each was derived from the global Bayesian Network configuration (grey boxes indicate the traffic
data collection stations).
In our work, the BN attribute connection threshold “t” is set to 0.1 for establishing the maximum number of links. The dis-
cretization method for data set values was selected as equal frequency. The examples of local cause-effect connections (in Figs. 4–7)
may provide useful initial information for design strategy of the air pollution predictive model. In the example, an interpretation of
configuration can be made as shown in Fig. 4 where the CO pollution in Newarke area is caused by Station 33 as the temperature
variation also seems to be an impact factor. If there would be a direct link between the temperature and Station33, then the possible
conclusion would be made that the temperature variation might be caused by high density of traffic flow or traffic jam. But in this
circumstance temperature is possibly the function of CO gas emission. In Fig. 5, traffic flow effects of stations St19 and St10 on NO2
pollution in the Aunsite area with the influence of wind speed variations, which is concerned with topographic characteristics of the
location. Similarly, in Fig. 4, the pollution caused by Station59 (labelled as st59) is under the influence of wind direction where it has
to be taken into account during the traffic network design phase in regards to geographic location. In Fig. 7 an interaction between
the stations provide a beneficial information for easing the traffic load on any of those station junctions by transferring its traffic flow
to the other. In Fig. 6, SO2 Newarke area has link with Station59, it has also link with wind direction whose cause-effect relationships
requires an interpretation of the wind effect.
The predictive GIS layer generation for a single point is shown in Fig. 1. The same procedure (for each desired location) can be
repeated by assigning each layer point (e.g. traffic flow, other gas pollutions) as a class node in turn in the predictive Bayesian model
(Fig. 1). It needs to be highlighted that the calculation of each predicted pollution value on the generated GIS layer will depend on the
availability data collection stations at sufficient frequency so that a feasible interpolation between the stations can be done to
calculate a map grid. Table 2 shows that NO2 and CO pollutions are only linked to traffic flow data where for the SO2 the en-
vironmental parameters are additionally involved. The likely reason would be that in the formed BN the environmental attributes
have very indirect effects on such pollutions, and their additional links in BN would not improve the prediction accuracy further. The
prediction of pollution/air quality linked to the traffic flows (at the data collection station areas) would give us the numerical results
(e.g. pollution below or above the hazardous threshold) which means that we can predict the pollution phenomenon such as, if traffic
flow at Station14 exceeds a specified threshold value (e.g. specified at network design stage) in association with certain wind speed/
direction, then SO2 pollution would become > 0.18 ppb (parts per billion) which is specified as high pollution. Similarly, if traffic
flows exceed the specific values at stations 1, 33, 59 simultaneously, then NO2 air pollution would be > 8.41 ppb that is beyond the
specified pollution level (please note that the pollution thresholds are selected according to our limited data set symbolically to
enable class attribute values to include both (high/low) classes for proper BN training whereas at the system design stage those
thresholds will be selected more properly specific to the residential issues whether for school, hospital or less sensitive urban areas. In
St59
SO2_Newarke
Wind direction
Fig. 6. Traffic flow effects of stations St59 on SO2 pollution in Newarke area with the influence of wind direction parameters.
241
A. Orun et al. Transportation Research Part D 63 (2018) 236–243
St24
St48
Fig. 7. Traffic flow interaction between the stations St24 and St48 which may be taken into account for easing the traffic load of each one at traffic
network design stage.
Table 2
Prediction accuracies of the predicted pollution values of a new GIS layer calculated by BN classifier with the PowerPredictor system options, and
automatically selected attributes that primarily influence the prediction process.
Gas pollutant Prediction accuracy (in GIS layer) Influencing attributes Bayesian network options
SO2 85% Wind direction, wind speed, traffic flow at Station14 t = 0.1, equal frequency discretization
NO2 78% Traffic flow at Stations 1, 33, 59 t = 0.1, equal frequency discretization
CO 81% Traffic flow at Station 42 t = 0.1, equal freq. discretization
the prediction results (Table 2) calculated by use of Bayesian classifier (PowerPredictor™) (Cheng et al., 2002), SO2 values are
calculated with 85% accuracy of low/high gas emission level prediction, based on: wind direction, wind speed, station 14 traffic flow
attributes. Meanwhile NO2 values are calculated with 78% accuracy of low/high gas emission level prediction, based on: station1, 33,
59 traffic flow attributes and CO values with 81% accuracy of low/high gas emission level prediction, based on: station 42 traffic flow
attributes. This means that new pollution layers (maps) in GIS can be predicted as (low = harmless, high = harmful) with 78–85%
accuracy at the design stage of a road network, residential areas, hospitals, etc. before their construction stage to avoid harmful
effects of air pollution (see Table 3).
4. Conclusion
The method proposed in this article aims on establishing relation between traffic and traffic-induced air pollutants, (SO2, NOx,
CO) whose considerable impact on public health and environment had already been highlighted, e.g. by Health Effects Institute
(2010) and Alvarez-Vazquez et al. (2017). Meanwhile the distinguished role of GIS systems that associate with the pollution models
have also been emphasized by several authors (Gulliver and Briggs, 2005; Hochadel et al., 2006). In our work we consider the above
factors to develop a Bayesian predictive model by inclusion of the major pollutants. To achieve this by the experiments, the sample
layers (SO2, NO2 and CO pollutions assigned as class nodes in turn) are constructed using a Bayesian Classifier (PowerPredictor™)
utility with 85%, 78% and 81% classification accuracies respectively by an optimised network parameters selection (Table 2), but
based on limited and partially missing data set due to local stations sensors’ data capturing discontinuity. By the experiments it has
been proven that the proposed method would lead to development of predictive layers to be integrated with a GIS system with
prediction capability. As stated in the past literature, the learning classification systems like Bayesian Networks used in this appli-
cation have several advantages over the other traditional data analysis systems in terms of flexibility of adaptation against the
potential changes (e.g. additional data inclusion without any structural change of system). These advantages are still in question even
though the earlier non-learning systems presented in the introduction section above provide higher prediction accuracy but with a
potential of overfitting problem. In the literature, it is indicated that Bayesian method guards against overfitting (Kass and Raftery,
1995) and it also exhibits better performance than other classifiers like Neural Nets or Maximum likelihood (Denzer et al., 1995).
Table 3
Bayesian statistical results about SO2 pollution with 85% accuracy, 90% sensitivity and 80% specificity. Sensitivity value is linked to high pollution
detection with 90% accuracy which is more important than the detection of non-pollution that refers to specificity (with 80% accuracy). The lift
index indicates the ratio between the results predicted using the classification model and the results using no model (predicted randomly). The ROC
area under curve (AUC) measure of the accuracy values.
Predicted High pollution (SO2) Low pollution (SO2) Lift index
ROC indices
High pollution (SO2) 0.00 0.90
Low pollution (SO2) 0.90 0.00
Accuracy = 85%, Sensitivity = 90%, Specificity = 80%
242
A. Orun et al. Transportation Research Part D 63 (2018) 236–243
An accurate and reliable modelling is always a big challenge for high parameter-interactive time-sequenced domains, like air
pollution measures of a traffic area. Such a modelling issue would only be solved by state-of-art techniques such as artificial in-
telligence assisted prediction in association with efficiently distributed low-cost sensor networks, etc. Our ultimate target within this
framework was an optimised method for a predictive traffic air pollution modelling at very local scale of urban areas by which
maximum desired reliability and accuracy would be obtained by use of feasible instrumentation and labour work at moderate cost
(that is affordable by local governments). The method would be particularly useful at traffic network design stages where subtle
parametric impacts would be more effective than expected on the environment and economy on the long-term basis.
The proposed predictive model introduced here would be further extended to be used in areas with critical importance (e.g.
nuclear plants, airports, hospitals, schools or other strategic and public constructions) with aim to support planners and potentially
avoid irreversible fatal mistakes or significant economic implications. For each application, a specific data layers and prediction
layers that are expected to be in interaction in a reasonable time domain should be designed and developed with efficient data
collection. The accuracy result obtained would be further improved by availability of data (e.g. for longer period and more efficient
data types integration). As an overall conclusion, the proposed work will enable conventional GIS systems to be upgraded to an
intelligent level in assistance with AI methods.
References
Alvarez-Vazquez, L., Garcia-Chan, N., Martinez, A., Vazquez-Mendez, M.E., 2017. Numerical simulation of air pollution due to traffic flow in urban networks. J.
Comput. Appl. Math. 326, 44–61.
Cheng, J., Bell, D., Liu, W., 2002. Learning Bayesian Networks from Data: An Efficient Approach based on Information Theory. < http://www.cs.ualberta.ca/~jcheng/
lab.htm > .
Corani, G., Scanagatta, M., 2016. Air pollution prediction via muti-label classification. Environ. Modell. Softw. 80 (2016), 259–264.
Denzer, R., Russell, D., Schimak, G.R., 1995. Environmental Software Systems. Springer-Science & Business Media, USA, pp. 230.
Gulliver, J., Briggs, D.J., 2005. Time-space modelling of journey-time exposure to traffic-related air pollution using GIS. Environ. Res. 97(1), 10–25.
Hochadel, M., Heinrich, J., Gehring, U., Morgenstern, V., Kuhlbusch, T., Link, E., Wichmann, H.E., Kramer, U., 2006. Predicting long-term average concentrations of
traffic-related air pollutants using GIS-based information. Atmos. Environ. 40, 542–553.
House of Commons Environmental Audit Committee, 2011. Air Quality: A Follow up Report, Ninth Report of Session 2010-12, vol. 1. < www.parliament.uk/
eacom > .
Health Effects Institute, 2010. Traffic Related Air Pollution: A Critical Review of the Literature on Emissions, Exposure and Health Effects. Special Report 17, Jan 2010,
Boston, MA, USA.
Iruansi, O., Guadagnini, M., Pilakoutas, K., Neocleous, K., 2012. Predicting the shear resistance of RC beams without shear reinforcement using a Bayesian neural
network. Int. J. Reliab. Saf. 6(1/2/3), 82–109.
Jason, G.S., Jerrett, M., Bckerman, B., Wilhelm, M., Ghosh, J.K., Ritz, B., 2009. Predicting traffic-related air pollution in Los Angles using a distance decay regression
selection strategy. Environ. Res. 109(6), 657–670.
Koski, T., Noble, J., 2009. Bayesian Networks: An Introduction. John Wiley and Sons, UK.
Karatzas, K.D., Kaltsatos, S., 2007. Air pollution modelling with the aid of computational intelligence methods in Thessaloniki, Greece. Simulat. Modell. Pract. Theory,
15(10), 1310–1319.
Kass, R.E., Raftery, A.E., 1995. Bayes factors. J. Am. Stat. Assoc. 90 (430), 791. http://dx.doi.org/10.2307/2291091.
Matthias, H., Heinrich, J., Gehring, U., Morgenstern, V., Kuhlbush, T., Link, E., Wichmann, H., Kramer, U., 2006. Predicting long-term average concentrations of
traffic-related air pollutants using GIS-based information. Atmos. Environ. 40 (2006), 542–553.
Niemeier, D.A., Beard, M.K., 1993. GIS and transportation planning: a case study. Comput. Environ. Urban Syst. 17(1), 31–43.
Orun, A.B., Aydin, N., 2010. Variable optimisation of medical image data by the learning Bayesian Network reasoning. In: 32nd Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC'10), Buenos Aires, Argentina, 1st–4th September.
Orun, A.B., 2004. Automated identification of man-made textural features on satellite imagery by Bayesian networks. Photogram. Eng. Rem. Sens. 70(2).
Orun A.B., 1993. Geographic information systems integration with the remote sensing data and the expert systems. In: 25. Environment and Remote Sensing
Conference, Austria – Graz, April 1993.
Passow, B.N., Elizondo, D., Goodyer, E.N., Leigh, R.J., Lawrence, J.P., Shah, S., Obszynska, J., Brown, S., Gustaffsson, S., Hübner, N., 2012. iTRAQ – an integrated
traffic management and air quality control system using space services. In: 4th International Conference on Space Applications, Toulouse Space Show, Toulouse,
France, 25–28 June 2012.
Pearl, J., 1988. Probabilistic Reasoning in Intelligence Systems: Network of Plausible Inference. Morgan Kaufmann, San Francisco, California.
Rebolj, D., Sturm, P.J., 1999. A GIS based component-oriented integrated system for estimation, visualization and analysis of road traffic air pollution. Environ.
Modell. Softw. 14(6), 531–539.
Thonga, C.M., Wongb, W.G., 1997. Using GIS to design a traffic information database for urban transport planning. Comput. Environ. Urban Syst. 21(6), 425–443.
Schwela, D., 2000. Air pollution and health in urban areas. Rev. Environ. Health 15(1–2), 13–42.
Shekarrizfard, M., Imani, A.F., Tetreault, L.F., Yasmin, S., Reynaud, F., Morency, P., Plante, C., Drouin, L., Smargiassi, A., Eluri, N., Hatzopoulou, M., 2017. Regional
assessment of exposure to traffic-related air pollution: impacts of individual mobility and transit investment scenarios. Sust. Cities Soc. 29 (2017), 68–76.
Suggs, J.C., Curran, T.C., 1983. An empirical Bayes method for comparing air pollution data to air quality standards. Atmos. Environ. 17(4), 837–841.
US Environmental Pollution Agency, 1995. Compilation of Air Pollutant Emissions Factors (AP-42), 5. Edition of AP-42 Report, North Carolina, USA.
Zhu, G., Zhang, P., Tshukudu, T., Yin, J., Fan, G., Zheng, X., 2015. Forecasting traffic- related nitrogen oxides within a street canyon by combining a genetic algorithm-
back propagation artificial neural network and parametric models. Atmos. Pollut. Res. 6(6), 1087–1097.
243