lable at ScienceDirect

Journal of Cleaner Production 206 (2019) 688e700

Contents lists avai

Journal of Cleaner Production

journal homepage: www.elsevier .com/locate/ jc lepro

Emission pattern mining based on taxi trajectory data in Beijing

Tingting Li a, *, Jianping Wu a, Anrong Dang b, Lyuchao Liao a, Ming Xu a

a Department of Civil Engineering, Tsinghua University, Beijing, 100084, Chinab School of Architecture, Tsinghua University, 100084, China

a r t i c l e i n f o

Article history:Received 20 June 2017Received in revised form9 August 2018Accepted 7 September 2018Available online 18 September 2018

Keywords:Data miningTraffic analysis zoneEmission patternBeijing

* Corresponding author.E-mail addresses: ttlithu@163.com (T. Li), jianping

danrong@tsinghua.edu.cn (A. Dang), fjachao@gmail.cedu.cn (M. Xu).

https://doi.org/10.1016/j.jclepro.2018.09.0510959-6526/© 2018 Elsevier Ltd. All rights reserved.

a b s t r a c t

Traffic-related air pollution has been one of the major environmental problems in China. It is urgent toexplore the urban traffic emission patterns for the low-carbon urban planning and traffic management.With this purpose, a new urban traffic emission analysis model is proposed in this paper. The trafficanalysis zones (TAZs) are treated as the analysis unit. Then the spatial and temporal dynamic emissionpatterns are studied based on taxi GPS data in Beijing. The whole urban area of Beijing is divided into 33TAZs depending on the feature of road networks. And the trip patterns of TAZs are extracted. Theinstantaneous emissions of CO2, NOx, VOC and PM within and between TAZs are estimated. The re-lationships between emissions and road densities are studied. The results demonstrated that (1) thehighest taxi trips during the day occur at 10:00, 16:00 and 20:00. (2) The variations of the 4 pollutantswithin and between TAZs are similar. The emissions of TAZs with business centers, entertainment centersand transportation hubs are obviously higher than others. For TAZs within the 5th Ring Road, thenorthern emissions are stronger than the southern. Emission patterns can be divided into 3 types, cor-responding to the time periods of 0:00e3:00, 3:00e6:00 and 6:00e24:00. (3) There is a positive rela-tionship between emissions within and between TAZs and road densities. And when the road densities ofTAZs are bigger than 0.6, emissions within TAZs will rise up obviously with the increasing road densities.The policy implications of results include the region function planning, traffic planning, the publictransport improvement, optimal design of charging stations and new energy vehicles promotion.

© 2018 Elsevier Ltd. All rights reserved.

1. Introduction

Urban travel demand is growing quickly with the urbanizationprocess in China. And it causes many traffic problems. The rapidlyincreasing greenhouse gas (GHG) emissions and air pollutioncaused by fossil fuel from road traffic have attracted much atten-tion. It has become an imperative task to conduct energy-savingand emission reduction work. The central and local municipalgovernments in China have centered on emission mitigation (Wuet al., 2017). In the aspect of vehicles, there are efforts about thedevelopment of China 1e6 emission standards (MEE and SAMR,2001, 2005b, c, 2013, 2016), enhanced I/M programs for in-usevehicles (Wu et al., 2011), yellow-labeled vehicles classificationand regulations (State Council P.R. China, 2013), improvement onvehicle engines (Huo et al., 2012a) andmotorcycles and heavy-duty

wu@tsinghua.edu.cn (J. Wu),om (L. Liao), mxu@tsinghua.

truck restrictions (MEE and SAMR, 2005a). In the aspect of fuel, themain efforts contain the fuel quality improvement (Yue et al., 2015)and new energy vehicle promotions (State Council P. R. China,2012). In the aspect of traffic management, there are policiesabout public traffic infrastructure enhancement, public transportsubsidy (MOHURD. P. R. China, 2016), bus lanes (Yu et al., 2015), busrapid transit system (Liu and Teng, 2014), new energy buses (Linand Tan, 2017) and taxis, odd and even number restrictions (Wuet al., 2017) and old vehicle regulations (Zhang, 2014). As for eco-nomic measures, there are car purchase restrictions (Diao et al.,2016), increased parking fees, new energy vehicle subsidies(Helveston et al., 2015; Ma et al., 2017; Xu et al., 2017) and taxreduction. Besides, the policy about the congestion fee is beingconsidered.

The advanced information and communication technologiespromote the development of smart city. At the same time, varioustraffic detection data is applied into the broader research field. Thediversified traffic detection data can be divided into 3 classes. Thefirst one is the individual travel data, containing the smart cardrecords of buses and subway (Huang et al., 2017b), bike-sharing

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700 689

data (Zhang et al., 2015), car-sharing data (He et al., 2014; Yu et al.,2017), cellphone signal data (Mao et al., 2016), cellphone navigationdata and other application data of cellphone. The second one is theregulation data of enterprises, such as logistic data of commercialvehicles (Joubert and Meintjes, 2015). The third one is the roaddetection data, including GPS trajectory data (Hsueh and Chen,2018; Li et al., 2015), electronic toll data based on RFID (Tsenget al., 2014), vehicle detection data of inductive loop detectors,microwave detectors, infrared detectors, laser radar detectors, ul-trasonic detectors and closed-circuit television cameras (Seo et al.,2017). The appearance of big data mining techniques makes itpossible to explore the patterns of large-scale urban traffic and airpollution. Compared with the social media data, road detectiondata and mobile communication network data, taxi GPS trajectorydata has advantages of wide coverage, high spatial resolution andlow installation and maintenance cost. There have been relatedstudies based on GPS records in urban traffic management andcontrol such as traffic congestion and incident detection (D'Andreaand Marcelloni, 2017; Munoz-Organero et al., 2018), traffic stateestimation (Hsueh and Chen, 2018; Zhan et al., 2017) and routechoice (Ciscal-Terry et al., 2016).

In addition, energy-saving and emission-reduction work hasbecome a global issue, leading to some research on traffic emissions(Huang et al., 2017a, 2017b; Luo et al., 2017). There are two types ofapproaches in constructing high-resolution traffic emission in-ventories, that is, top-down and bottom-up methods (Liu et al.,2018). Top-down approaches are applied into the estimation ofannual vehicle emission inventories, in which the emission iscalculated relying on the average trips and traffic flow, averagevelocity, average mileage and fuel type and then allocated to eachcell (G�omez et al., 2018; Puliafito et al., 2015). But this allocationmethod may result in low accuracies and uncertainties in thespatial emission estimation (G�omez et al., 2018). What's more, therestricted access of high-emitting vehicles in some urban area maycause estimation errors. Thus, many works are relied on thebottom-up rather than top-down approaches. Some scholarsdemonstrated that the data collected in finer spatial scale inbottom-up method could be helpful for more accurate emissionestimation (L�opez-Aparicio et al., 2017; Pallavidino et al., 2014).Besides, the traffic simulationmodels are used to improve emissioninventories (Hofer et al., 2018; Osorio and Nanduri, 2015; Xu et al.,2018; Zhu and Zhang, 2017). However, the emission factor variesaccording to road conditions, vehicle mileage and locations. Thereal-time driving states are neglected in these aforementionedworks, which are neccessary for more accurate emission estimation(Nyhan et al., 2016). Even in the MOBILE model (USEPA, 2012),there are only 14 types of driving states. Early in 2003, Frey et al.(2003) demonstrated that the emissions of hydrocarbons and car-bon monoxide in the acceleration period were 5 times greater thanthe idling period. To solve the above problems, the vehicle drivingstates with speed and acceleration data should be extracted formore precise emission inventory estimation.

The taxi GPS trace has advantages of monitoring the whole tripand reflecting fine-gained individual space-time trajectories. Theindividual traces contain the instantaneous speed, accelerations,positions and directions (Nyhan et al., 2016), which are precise toestimate the traffic emissions. Relying on GPS data, some re-searchers allocated the accumulated emissions calculated bymicroscopic emission model into the road network and observedthe spatial distributions of emissions (Luo et al., 2017; Nyhan et al.,2016). Other scholars studied emission inventories in the resolutionof grid cells which split the geographic area into several parts, suchas 1� 1 km resolution (G�omez et al., 2018; Gois et al., 2004; Wanget al., 2010). However, the trips are not related to a single location or

a single road, they are regional behaviors. And a trip covers severalroads and regions. Thus it is necessary for scholars to analyze trafficemissions caused by trips from the viewpoint of regions. Particu-larly, traffic analysis zones (TAZs) are defined based on regionalattributes such as geographic features. The TAZ is the basic analysisunit of the traditional four-step urban travel demand forecastingmodel (Agrawal et al., 2018; Sun et al., 2016). The trip demandbetween the TAZs is estimated, which is useful for the heavilyloaded area identification. Then the planners make decisions bytaking the feature of TAZs into account. Poku-Boansi and Cobbinah(2017) studied the land use mode and urban travel based on TAZs.Jang et al. (2017) explored characteristics of urban built environ-ment by observing traffic crash occurrences of TAZs. Lee et al.(2018) analyzed the spatial patterns of vehicle crash proportion atTAZs level. TAZs are important analysis units in the relatedresearch. But to authors' knowledge, there is a lack of the researchabout traffic emissions of traffic analysis zones.

A new urban traffic emission analysis model is proposed in thispaper. Previous studies are related to emissions on roads. But thereis a lack of research related to emission patterns of TAZs. A TAZ is abasic unit in traffic planning and management. Thus the TAZ isregarded as the unit of real-time urban traffic emission patterns.The purpose of this study is to explore the pattern of traffic emis-sions of TAZs in the urban area. Hence, the whole urban area ofBeijing is divided into 33 TAZs based on the feature of road net-works. The GPS data of 12409 taxis in Beijing are analyzed.With thehelp of a microscopic emission model, emissions of CO2, NOx, VOCand PM are estimated. The emissions within and between the TAZsthereby are identified. What's more, the relationships betweenroad features and emissions of TAZs are explored.

2. Data and methodology

2.1. Data source and data processing

Emissions within and between the TAZs not only representspatial distributions of traffic pollution, but also reflect the trend ofurban trips. As a general case in this study, Beijing is one of thelargest metropolises all over the world, which has experiencedexplosive growth in recent years. At the end of the year 2015, thepopulation of Beijing is up to 21.71 million with the GDP per capitaRMB 106.28 thousand. The municipal administrative area is16410.54 km2. And there are 68284 taxis in Beijing according toWen et al. (2016). The data in this paper originates from taxi GPStrajectory records in Beijing on 1st November, 2012 (workday),which contains 12409 taxis. Each taxi is sampled average each 32 son a 24-h basis as Table 1 shows. The major area with taxi dataintensively available is within the 6th Ring Road. And the longitudeand latitude of the research area is from 116.164 to 116.575 and39.743 to 40.098. As Table 2 displays, each data record consists ofthe real-time data of identification, operation state, time stamp,longitude, latitude, speed and GPS state.

The data preprocessing can be divided into 2 steps.

(1) In the data cleaning step, the GPS records which are beyondthe research area are eliminated. The points with aninstantaneous velocity more than 150 km/h are removed. Asfor accelerations, it is determined based on the time andvelocities between each two sample points. Considering thereal driving cycle, the points with the acceleration lessthan �10m/s2 and more than 10m/s2 are deleted (Nyhanet al., 2016).

(2) In the filtering step, information of the taxi trips is analyzed.If the position of the taxi keeps the same and the operationstate is vacancy or parking, then the records will be

Table 1Description of taxi GPS data.

Taxi number Record number Time resolution (s) Duration (hour) Research area

12409 33,885,600 32 24 116.164e116.57539.743e40.098

Table 2Taxi GPS record sample.

ID Operation state Time stamp Longitude Latitude Speed (km/h) GPS state

426081 1 20121101170310 116.3532486 39.9433403 6 1490923 0 20121101170315 116.3369598 40.0275574 0 1

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700690

eliminated. Besides, the records with continuous time, thesame taxi ID and operation state are extracted as represen-tative trips.

2.2. Emission model

A complete taxi trajectory can be partitioned into several sub-divisions which correspond to GPS records with speed and accel-eration information. Pollutant emissions of each subdivision can beestimated referring to the microscopic emission model (Int Paniset al., 2006). Then they are summed up for working out emis-sions of each trip. The instantaneous traffic emission model isapplied to analyze emissions (Nyhan et al., 2016; Osorio andNanduri, 2015). And Harrington et al. (1998) believed that theemission rates can be calculated according to the base emission rateand the correction factors. In this paper, real-time emissions ofpollutants CO2, NOx, VOC and PM are estimated. For a taxi whose IDis signed with letter n, the emission rate of pollutant p in time in-terval D t can be calculated as follows.

EpnðtÞ ¼ Epn; basicðtÞ$Cpf $C

pa$C

pe (1)

EpnðtÞ is the instantaneous emission rate of pollutant p (g/s) forthe vehicle n. Epn;basicðtÞ is the basic emission rate proposed by Int

Panis et al. (2006). And the variable p can represent CO2, NOx,VOC and PM. Cp

f , Cpa and Cp

e are correction coefficients of vehicle

emission control technology, taxi mileage and drivingenvironment.

(1) Basic emission rate

Epn; basicðtÞ ¼ maxhEp0; f

p1 þ f p2vnðtÞ þ f p3vnðtÞ

2 þ f p4anðtÞ þ f p5anðtÞ2

þ f p6vnðtÞanðtÞi

(2)

Variables vnðtÞ and anðtÞ represent the speed (m/s) and

Table 3Emission factors of LDGVs.

Pollutant CO2(g/km) (Carslaw et al.,2011)

NOx (g/km) (Huo et al.,2012b)

VOC (mg/km) (Cao2016)

Euro 0 228 1.9 469.3Euro 1 212 1.0 80.7Euro 2 204 0.47 56.8Euro 3 193 0.23 25.6Euro 4 178 0.05 14.9

acceleration (m/s2) of car n at time t respectively. Ep0 is the low

bound emission rate of pollutant p (g/s). Variables f1p, f2

p, f3p, f4

p,f5

p and f6p serve as the emission constants differing with taxi ID

and pollutants. And their values also vary depending on theinstantaneous moving state of the car. For pollutants NOx and VOC,the acceleration anðtÞ should be compared with �0.5m/s2 firstly.The specific values are determined according to emission functionsof petrol cars proposed by Int Panis et al. (2006).

(2) Correction coefficient of vehicle emission control technology

The coefficient Cpf corrects the emission difference caused by

vehicle emission control technologies of various emission stan-dards, which can be indirectly reflected by the average emissionfactors. The China 4 emission standard has been conducted since2008 for light-duty gasoline vehicles (Wu et al., 2011). And theshare of gasoline taxis in Beijing was 100% in 2010 (Yang, 2018;Zhang et al., 2014). The oldest taxis are 5 years old (Lin et al., 2009),which are in compliance with China 4 emission standard. Then thetaxis in this paper are believed to be gasoline cars with China 4emission standard. Besides, the vehicle emission standards China1e5 are considered to be equivalent to Euro 1e5 standards (Yueet al., 2015).

Cpf ¼ EFP4

,X3i¼0

Pi$EFpi (3)

As Eq. (3) shows, EFP4 means the emission factor of pollutant p inChina 4 standard. The variable i means the emission standard, forinstance, 0 for Euro 0, 1 for Euro 1. Pi is the ratio of vehicles in Euro istandard in Panis' measurement (Int Panis et al., 2006). And thevariable EFpi can represent the emission factor of various pollutantswhich are listed in Table 3.

(3) Correction coefficient of taxi mileage

The basic emission rates are not adequate for emission estima-tion of taxis in Beijing. Therefore, the taxi correction coefficient is

et al., PM (mg/km) (Hendriksen and Rijkeboer, 1993; Mamakos et al., 2012;Serves, 2005)

7.001.900.560.530.49

Table 4Fleet composition.

Taxi age (year) 1 2 3 4 5

MMEANi (� 104km) (Zhang et al., 2014) 12.6 25.2 37.8 50.4 63.0

Ri (%)(Lin et al., 2009)

14.5 35.4 25.2 16.1 8.8

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700 691

put forward in this paper. The age of the vehicle is very important toemission estimation. The corrected emission rate reflects thedegradation rate considering the vehicle mileage. As Eq. (4) shows,Ri can represent the age ratio of taxis in Beijing, which is listed in

Table 4 (Lin et al., 2009).MMEANi is the average mileage (Zhang et al.,

2014). Dpi is the mileage degradation rate which is function of the

pollutant and mileage. Lin et al. (2009) also researched the rela-tionship between the vehicle age and mileage in China, which washelpful in calculating the value of Dp

i . And the parameters are listedin Table 5. For VOC and PM, there is little degradation research,which will be supplemented in the further field study.

Cpa ¼

X5i¼1

Ri$Dpi (4)

Dpi ¼ ap �MMEAN

i þ bp (5)

(4) Correction coefficient of environment

The environment correction coefficient Cpe reflects the influence

of the geographical location on emissions. And it is the integrationof altitude, temperature and humidity. In this paper, the tempera-ture factor is considered. Joumard et al. (2009) researched re-lationships between temperature and emission as Eq. (6) shows. Tis the value of temperature with unit �C. The average temperatureof 1st November, 2012 is defined as 9 �C (php.weather.sina.com,2012). Variables gp and hp are the correction parameters ofpollutant p.

Cpe ¼ gp � T þ hp (6)

Fig. 1. Neighborhood structure of pixel p.

2.3. Map segmentation

Differing from previous studies related to traffic emissions ofurban roads, the spatial emission patterns within and betweenTAZs are explored in this paper. The research area is divided intoseveral TAZs with the following 3 rules.

(1) Each region is rational and meaningful in real Beijing map.(2) The features of various regions are comparable.(3) There is no obvious difference in road densities for the same

TAZ.

The urban area is splited by arterial roads into various parts(Yuan et al., 2012). And the roads have strong influence on socio-

Table 5Emission degradation rate due to mileage for Euro 4 petrol cars.

Parameter CO2 NOx (Ntziach

ap 0 3.986 e-06bp 1 0.932

MMEANi � 160;000 km 1 1.57

economic activities. For the first rule, the city map can besegmented based on original road networks. Citizens live in theseregions and take daily trips. Infrastructures with different functionsare located in these regions. Moreover, the area of each TAZ shouldbe approximate, leading to the meaningful comparison. After thesegmentation operation of arterial roads, these obtained small re-gions are clustered with the third rule.

2.3.1. Morphological operationMathematical morphology is a kind of theory in image and

signal processing, of which the basic idea is to identify and analyzethe image by utilizing structural elements with particular shape tomeasure and extract the corresponding shape in an image. In thispaper, urban roads are extracted as line elements. And the in-tersections, roundabouts, lanes and other elements of real roadnetworks are simplified by the morphological operation, leading tothe lower-complexity road networks. For instance, there are severalclosed inter-cells at roundabout in Fig. 2(a) where the road networkbinaryzation operation has been completed. The morphologicaloperation in this paper contains 2 steps.

(1) The first step is road dilation. Dilation is the basic operationof morphological process. The excess network informationsuch as the small connected parts caused by lanes, overpass,bridges and other elements can be eliminated with dilation.Letter A and B are defined as the sets in two-dimensionalinteger space Z2, which can be denoted as the expression A;B3 Z2. Dilation operations of set A with set B is signified inEq. (7). After parameter verification, proper structuralelement B is chosen. And the result proves that the gap be-tween roads and bridges can be eliminated.

A4B ¼ faþ bja2A; b2Bg ¼ Ub2B

Ab (7)

(2) The second step is network thinning. In the image thinningprocess, the foreground pixel p is deleted depending on theconnectivity of the 8 pixels, defined as a1, a2, a3,..., a8 (Lamet al., 1992). They besiege the pixel in the neighborhoodstructure as Fig. 1 shows. The connectivity is defined as XHðpÞ(Hilditch, 1969).

ristos and Samaras, 2010) VOC PM

e e

e e

e e

Fig. 2. Morphological operation.

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700692

XHðpÞ ¼X4i¼1

bi

bi ¼�1; when a2i�1 ¼ 0 and ða2i ¼ 1 or a2iþ1 ¼ 1Þ0; otherwise

(8)

According to (Lam et al., 1992), the essence of thinning process isthe iterative computation. Each iteration process is divided into 2steps. In the first step, the pixel pwill be deleted if Eqs. (9)e(11) aresatisfied. In the second step, the judgment criterions are similarwith the first one. And Eq. (11) should be replaced by Eq. (12). Aftermorphological operation, the network in Fig. 2(b) is generated withcomplete road connection information.

XHðpÞ ¼ 1 (9)

2 � min

(X4k¼1

a2k�1∨a2k;X4k¼1

a2k∨a2kþ1

)� 3 (10)

�a2∨a3∨a8

�∧a1 ¼ 0 (11)

�a6∨a7∨a4

�∧a5 ¼ 0 (12)

2.3.2. Spatial cluster model based on TAZ featuresIn order to obtain the appropriate TAZs, a new spatial cluster

model is proposed based on the similarity of regional features.Firstly, the research area is divided into 177 parts depending onmain roads. The 177 small parts are clustered into several biggerregions relying on cluster rules. The related definitions in spatial

Table 6The parameter notations.

Parameter Description

eki eki ¼ feki ; 1; ekiU U ¼ fUi; Uiþ1;

Areak the area of regiAreaUi

AreaUi¼ P

ki2Ui

A

l a predefined mlk;i the length of roDk Dk ¼ Pn

i¼1lk;i=ANki ;kj the adjacent rel

F F ¼ eki∩ekjDki

0 Dki0 ¼ ðDki � Dk

Ski ;kj Ski ; kj ¼���Dki

0 �

cluster model are listed in Table 6.The small regions with analogous road features are grouped into

big TAZs. And the emission patterns in diverse TAZs are explored.The TAZ set U is defined with two rules: (1) road density similarity.The road features of small regions in the same TAZ are most similar.This goal can be depicted as Ski ;kj � Ski ;km , where ki2Ui, kj2Ui,

km2Uj, Nki ;kj ¼ 1 and Nki;km ¼ 1 are satisfied. (2) Area restriction,

which can be expressed as AreaUi< l. The solution procedure of

spatial cluster model is shown in Table 7.The elbowmethod (Hardy, 1994) is applied to choose the proper

number of clusters and evaluate the result by standard deviation,which can be expressed in Eq. (13).

Std ¼ ð1=MÞ$XMj¼1

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffið1=NÞ$

XNi¼1

�Dki � DUj

�2vuut (13)

M denotes the number of cluster regions in the graph, and Ndenotes the number of small regions for each big region U. Based onthe variation trend of Std, number 33 is determined as the clusternumber as Fig. 4 displays. And these 33 TAZs do not destroy char-acteristics of the region formed by arterial roads such as the 2ndRing Road, the 3rd Ring Road, the 4th Ring Road and the 5th RingRoad.

3. Result

3.1. Trip patterns of TAZs

The origin and destination points of one trip are identified bythe operation state of GPS data. So do the time, longitude, latitudeand speed information. Trip patterns of the 33 TAZs are analyzedwith trip trajectories of taxis. The trip number of each TAZ in jth

; 2; eki ; 3; :::; eki ; n;g, the edge set of the region ki::: Ung, the set of regionson kreaki , the area of region Ui

aximum expected area of the traffic analysis zone Ui

ad i in region k

reak , road density of region kation of region ki and kj , if kiskj and Fs∅, Nki ;kj ¼ 1

;minÞ=ðDk;max � Dk;minÞDkj

0���, the road density similarity of region ki and kj

Table 7The solution procedure of spatial cluster model.

Step Operation

Step 1 Variable initialization. The region classification record matrix I ¼ 0. Iði; 2Þ means the ID of the big region which contains small region ki .Assign a specific value to l.

Step 2 Search for k which meets Areak � l in the set fk1; k2; k3; :::; kng and define region k as a big region.Step 3 Search for the smallest region ID which does not belong to any big region. If the search result is null, then stop iterating.

Otherwise, search for the region j which meets Nki ;kj ¼ 1 and Iðj;2Þ ¼ 0 as Fig. 3 depicts.Step 4 If J ¼ ∅, search for region kj where Nki ;kj ¼ 1, kj2Uj and Ski ;Uj

¼ min Sk;Uj, assign the value Iði;2Þ to Iðj;2Þ

and go to Step 3. Or else, go to Step 5.Step 5 Calculate Ski ;kj where kj2J, and search for the region kj with min Ski ;kj . Update AreaUi

and DUiwhich contains region ki .

Step 6 If Area Ui� l, then update I and go to Step 3. Otherwise, assign region ki and kj into a new region and mark it with variable ki .

Search for region kj which meets Nki ;kj ¼ 1 and Iðj;2Þ ¼ 0. And check whether J ¼ ∅ can be satisfied. If so, update I and go to Step 3. Or else, go to Step 5.

Fig. 3. Region search and cluster.

Fig. 4. Cluster result.

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700 693

hour is defined as follows.

GaðjÞ ¼ Gina ðjÞ þ Gout

a ðjÞ (14)

Gina ðjÞ ¼

XT

jfn2ð1;NÞjDn2Cagj (15)

Gouta ðjÞ ¼

XT

jfn2ð1;NÞjOn2Cagj (16)

Gina ðjÞ means the number of inflow trips in TAZ a, and Gout

a ðjÞ isthe number of outflow trips. Variable T represents the set of trips,that is, all trips of N taxis in jth hour. In this paper, the tripsrepresent the driving process with and without customers. Ca is thecoordinate set of TAZ a. Variable On indicates the origin point of thetrip, and Dn indicates the destination point. For the hourly trippatterns, the time interval is defined as one hour. The hourly tripnumber GaðjÞ represents the trip number of TAZ a in jth hour on 1stNovember, 2012. The relative trip number is defined as GaðjÞ0 in Eq.(17). And the hourly trip number of all TAZs is defined as GðjÞ in Eq.(18).

GaðjÞ0 ¼ ½GaðjÞ � Gmin�=ðGmax � GminÞ (17)

GðjÞ ¼X33a¼1

GaðjÞ (18)

Gmin and Gmax are the minimum andmaximum of GaðjÞ. And thevalue of GaðjÞ0 is shown in Fig. 5. It can be observed from Fig. 5 thatthe relative trip number of most TAZs changes greatly during theday, especially for the TAZ identified with ID number 10, 11, 12, 14,16e23. In Beijing map, these TAZs are located within the 4th RingRoad or in the northern area between the 4th Ring Road and the 5thRing Road, confirming that the lively area is urban center. Evidently,the high relative trip numbers are in the daytime and early night(6:00e24:00).

The numbers of real-time trips GðjÞ and running taxis are shownin Fig. 6. The trip number changes greatly during the day. It beginsto rise up at 5:00 in the morning. And the first peak appears around9:00e10:00. Then it drops at noon. And the trip number reaches alow level around 13:00. In the time period of 13:00e16:00, tripnumber goes up steadily. And the highest value appears at around15:00e16:00 rather than the peak hours (7:00e9:00 and

Fig. 5. Relative taxi trip number of each TAZ (the value of GaðjÞ0).

Fig. 6. Number of taxis and trips.

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700694

17:00e19:00) of urban traffic demand. From 17:00 to 18:00, the tripnumber keeps getting smaller owing to the heavier and heaviertraffic. After the turning point at around 19:00, arrangements ofdining out and other leisure activities cause another trip peak. Andafter 23:00, taxi trips decrease quickly, implying that people tend togo to sleep. As for the running taxi number, it has the same varia-tion trend with the trip number before 16:00. The taxi number ataround 16:00 is more than the morning and evening peak hours,which is similar with the trip number. It can be explained that sometaxi drivers in Beijing tend to not work in peak hours because of thetraffic congestion (http://news.sina.com.cn, 2013). The average triptime in rush hours is longer than other time, leading to higheroperating cost. It is investigated that the operating revenue is lessthan cost in peak hours. Thus some taxi drivers usually choose tochange shifts, take a rest and park taxis in morning and eveningrush hours. Compared with rush hours, there are more runningtaxis in the afternoon, causing the highest trip number in the timeperiod of 15:00e16:00.

3.2. Emission patterns

3.2.1. Emission patterns between TAZsThe spatial emission patterns of TAZs are explored. The emission

quantities of pollutant p from TAZ a to TAZ b in the time period oft1 � t2 are defined in Eq. (19).

Eex; pa; b ðt1; t2Þ ¼Xl

t¼0

Xkn¼1

EpnðtÞDt (19)

EpnðtÞ is the emission rate of taxi trips from TAZ a to TAZ b.

Variable k is the trip number from TAZ a to b. And lmeans the traveltime. The accumulated emissions of pollutant CO2, NOx, VOC andPM are estimated according to Eq. (19). As Fig. 7 depicts, the CO2emissions are accumulated for each 3 hours, leading to 8 time in-tervals for the 24-h data. The straight lines connect geometricalcenters of TAZs. The color and width of lines stand for emission

quantities. For instance, Eex; CO2

a; b ð0;3Þ means the CO2 emission

quantity caused by trips between TAZ a and TAZ b in the time

period of 0:00e3:00. The relative emission quantity Eex;pa; b ðt1; t2Þ0is

defined in Eq. (20). Eex;pmax and Eex;pmin are the minimum and maximum

of Eex; pa; b ðt1; t2Þ.

Eex; pa; b ðt1; t2Þ0 ¼hEex; pa; b ðt1; t2Þ � Eex; pmin

i.�Eex; pmax � Eex; pmin

�(20)

Fig. 7 shows the value of Eex; CO2

a; b ðt1; t2Þ0between each two TAZs.

It can be observed that emissions center on the daily time and theearly night. Besides, emissions are chiefly located within the 5thRing Road, the downtown area. From 0:00 to 3:00, the majoremissions are distributed in the eastern area within the 5th RingRoad. And the high-emission area covers Houhai, Nanluoguxiang,Sanlitun and Workers' Stadium, the lively recreational areas inBeijing. The lowest emissions occur in the time period of3:00e6:00. With the beginning of human activities, emissionswithin the 5th Ring Road rise up during 6:00e9:00. Apparently, inthe time period of 9:00e12:00, the emissions between TAZs reach ahigh level. For the business areas such asWudaokou and Xidan, andtransportation junctions like Beijing South Railway Station andBeijing West Railway Station, there tend to be strong emissions

Fig. 7. Relative CO2 emission quantities between each two TAZs (the value of Eex; CO2

a; b ðt1; t2Þ0).

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700 695

between them and other TAZs. It reveals that the major daily hu-man activities are related to business areas and traffic hubs. Similarwith the former time intervals, spatial relationships in the timeperiod of 12:00e15:00 keep stable. After a drop between 15:00 and18:00, emission quantities in the period of 18:00e24:00 remainunchanged. By comparing emissions in different time intervals, itcan be identified that there exists temporal and spatial tripdisequilibrium between TAZs. And the hot spots practically arelocated within the 5th Ring Road. The instantaneous distribution ofemissions indicates that activities in northern area are livelier thansouthern area.

In order to explore the emission of various pollutants, emission

quantities during the whole day are accumulated, i.e. Eex; CO2

a; b ð0; 24Þ,Eex; NOx

a; b ð0; 24Þ , Eex; VOCa; b ð0; 24Þ, Eex; PMa; b ð0; 24Þ. Then the value of

Eex; CO2

a; b ð0; 24Þ0, Eex; NOx

a; b ð0; 24Þ0, Eex; VOCa; b ð0; 24Þ0 and Eex; PMa; b ð0; 24Þ0

can be obtained according to Eq. (20). As Fig. 8 shows, the emissiondistributions of the 4 pollutants are similar. The emission quantitiesare pretty strong between TAZs in urban center, with a downtrendfrom urban center to suburbs. In order to acquire the main spatialemission patterns, the first 12 strongest emission quantities be-tween 2 TAZs of the 4 pollutants are extracted as a new network asFig. 9 presents. It is highlighted that the main emission networkcovers all TAZs within the 4th Ring Road, with 10 points and 12edges. On the one hand, in the perspective of TAZ type, core areasembody large business centers and important transportation hubs.On the other hand, the emission quantities in eastern TAZs arestronger than the western TAZs. And the emission quantities be-tween the southern TAZs are stronger than the northern TAZs.During the whole day, the pollutant emissions show a polycentricdistribution.

3.2.2. Emission patterns within TAZsThe accumulated emission quantity within TAZ a in the time

period of t1 � t2 is defined as Ein; pa ðt1; t2Þ. As Eq. (21) shows, EpnðtÞ isthe emission rate of taxi trips within TAZ a. Variable k is the trip

number within TAZ a. And l means the travel time. The relative

emission quantities are defined as Ein; pa ðt1; t2Þ0.

Ein; pa ðt1; t2Þ0 ¼Xl

t¼0

Xkn¼1

EpnðtÞDt (21)

Ein; pa ðt1; t2Þ0 ¼hEin; pa ðt1; t2Þ � Ein; pmin

i.�Ein; pmax � Ein; pmin

�(22)

Ein; pmax and Ein; pmin are maximum and minimum of Ein; pa ðt1; t2Þ. Thevalue of Ein; CO2

a ðt1; t2Þ0 is shown in Fig. 10. The CO2 emission vari-ations within TAZs and between TAZs are similar. It hints that for acertain TAZ, the CO2 emissions within this TAZ have a positiverelationship with the emissions between this TAZ and other TAZs. Itis clear that a small amount of CO2 emission is caused in thenorthern and southeastern TAZs within the 4th Ring Road in thetime period of 0:00e3:00. In the later 3 hours, emission quantitieskeep a low level. Corresponding to commuting peak hours from6:00 to 9:00, regional emissions rise up within the 4th Ring Roadand in the northern area between the 4th Ring Road and the 5thRing Road. It can be proved that there is a short distance betweenresidence and work place for some citizens, which is helpful forurban commuting research. At 9:00, the CO2 emissions reach a highlevel. And emission quantities within the 5th Ring Road are higherthan other areas, especially the northern area within the 4th Ring

Road. For instance, the value of Ein; CO2a ðt1; t2Þ

0in the northern area

within the 2nd Ring Road can be up to 0.75e1, which is 10 times ofthe southern areawithin the 5th Ring Road from 9:00 to 12:00. Anda slightly downtrend occurs at around 21:00. It can be highlightedthat CO2 emission quantity generated by the short-distance taxitrips in central area is distinctly higher than the fringe area.

3.3. Correlations between emission quantities and TAZ features

Region segmentation based on spatial cluster model guarantees

Fig. 8. Relative pollutant emission quantities between each two TAZs during the whole day (the value of Eex; CO2

a; b ð0; 24Þ0 , Eex; NOx

a; b ð0; 24Þ0 , Eex; VOCa; b ð0; 24Þ0 and Eex; PMa; b ð0; 24Þ0).

Fig. 9. The first 12 strongest emissions of the 4 pollutants.

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700696

the maximum similarity of clustered areas. Road characteristicshave influence on TAZ trips (Crane and Crepeau, 1998), furtheraffecting emissions between TAZs. TAZs with high road densitiesmaybe attract more trips, resulting in influence on the speed andemissions. The road density of TAZ a is defined as variable Da. Therelative road density Da

0 is defined as follows.

Da0 ¼ ðDa � DminÞ=ðDmax � DminÞ (23)

Dmax and Dmin are the maximum and minimum of Da. Fig. 11shows the value of Da. It can be observed that the road densitieswithin the 5th Ring Road are higher than the fringe area, similarwith traffic emissions.

3.3.1. Correlations between emission quantities between TAZs andTAZ features

Accumulated emissions between TAZ a and other TAZs duringthe whole day are defined as Eex;pa ð0;24Þ. Variable U represents theset of TAZs. And the value of Eex;pa ð0;24Þ is normalized to

Eex;pa ð0;24Þ0 according to Eq. (20). The relationships between rela-

tive emissions Eex;pa ð0;24Þ0 of pollutant CO2, NOx, VOC and PM andrelative road densities Da

0 are indicated in Fig. 12. It implies that the

relationships between road densities and the emissions of 4 pol-lutants are similar. For TAZs with road densities higher than 0.6, theemissions rise up with the increasing road densities. The co-ordinates are divided into 4 parts by the median values of roaddensity and emission. And the 4 parts correspond to high roaddensity-high emission, high road density-low emission, low roaddensity-low emission and low road density-high emission. MostTAZs are located in high road density-high emission and low roaddensity-low emission area. Thus it can be denoted that there is apositive relationship between emission quantities and roaddensities.

3.3.2. Correlations between emission quantities within TAZs andTAZ features

The accumulated emissions within TAZ a during the whole day

can be presented as Ein; pa ð0;24Þ. And the relative emissions

Ein; pa ð0;24Þ0 can be obtained according to Eq. (20). The relationships

of Ein; pa ð0;24Þ0 and Da0 are exhibited in Fig. 13. When road densities

of TAZs are in the range of 0e0.6, the emission quantities withinTAZs will stay a low level. And for TAZs whose densities are higherthan 0.6, the positive relationships between emission quantitiesand road densities are obvious. It can be observed from Figs. 12 and13 that TAZs with road densities higher than 0.6 are located withinthe 4th Ring Road andmiddle-west and northeast area between the4th Ring Road and the 5th Ring Road. And high road densitiescorrespond to more business activities, recreation activities andtrip demand.

4. Discussion

The emission patterns of TAZs are studied in this paper, whichare helpful for urban planning and management.

(1) The temporal and spatial emission variations of TAZs areidentified with GPS data. And the significant disparity in

Fig. 10. Relative CO2 emission quantities within TAZs (the value of Ein; CO2a ðt1; t2Þ0).

Fig. 11. Relative road densities of TAZs. (the value of Da0).

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700 697

emissions across TAZs is verified. The government is sug-gested to reduce the emission disparity and improve thesupply of infrastructures with different functions in TAZs.That is, for different regions, the functions of business and

Fig. 12. Correlations between the relative emission quantities between TAZ

entertainment should be satisfied asmuch as possible so thatcitizens don't need to take long-distance trips with variouspurposes. What's more, the urban planners should alsoconsider the emisson bearing capacity of regions.

(2) As the basic unit of traffic planning, the features of TAZs suchas land use, population composition and infrastructuresshould be investigated during the first stage of planning.However, the current traffic emissions of TAZs are notconsidered in the previous study. As the major environ-mental problem, the air pollution hinders the economicdevelopment in China. Traffic planners can get policy sup-port from this study.

(3) The public transport priority policy has been applied inmanycities in China, which is expected to relieve air pollution. Ourresults lay a foundation for better public transit assignment.With the dynamic emission inventories in this paper, thegovernment officials are advised to increase bus lines be-tween high-emission TAZs in the high-emission time period.Then the bus accessibility can be improved, resulting in moreconvenient trips and less traffic emissions. Furthermore, thetaxi services also can be improved by designing more taxireception stations in high-emission TAZs.

s and road densities (the relationship between Eex;pa ð0; 24Þ0 and Da0).

Fig. 13. Correlations between the relative emission quantities within TAZs and road densities (the relationship between Ein; pa ð0; 24Þ0 and Da0).

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700698

(4) Results in this paper are helpful for government decisions onlocation assignment of electric vehicle charging stations andnew energy vehicles. Chinese government has recognizedthe popularization of EVs as a key method to alleviate urbanpollution. Based on this paper, the infrastructures of EVs andnew energy vehicles in high-emission regions should beimproved, resulting in more efficient spatial planning.

However, some details are neglected in this paper. Although themicroscopic emission model has been corrected with coefficients,there are still some inaccuracies for emission estimations. For themore accurate emissions, the field experiment of vehicle emissionsshould be performed in further study. Moreover, Int Panis et al.(2006) only focused on the emission measurement in urbantraffic. For highways with higher speed, this emission models aredifferent. And the emission estimation on highways can be insuf-ficiently represented based on the model used. Therefore, the au-thors plan to conduct the field test and compare the emissionsestimated with the test results in further study.

5. Conclusions

The 33 TAZs of the research area are defined as the analysisunits. The urban trip and emission patterns are explored based onreal-time taxi trajectories of 12409 taxis in Beijing. A new trafficemission patterns analysis model is proposed. Analysis is per-formed on emissions of CO2, NOx, VOC and PM, temporal andspatial variations and the relationship with the TAZ features. Theconclusions are as follows.

(1) There exist 3 peak time periods for TAZ trips, that is 10:00,16:00 and 20:00. And trips in TAZs within the 4th Ring Roador in northern area between the 4th Ring Road and the 5thRing Road are higher than others and fluctuate greatly.

(2) Generally speaking, the variations of the four pollutants aresimilar between TAZs. Spatially, emission quantities arerelated to the number and type of infrastructures and loca-tions of TAZs. The emissions of TAZs with business centers,entertainment centers and transportation hubs are obviouslyhigher than others. Emissions of TAZs within the 5th RingRoad are higher than the fringe TAZs in Beijing. And for TAZswithin the 5th Ring Road, the northern emissions arestronger than the southern. Temporally, emissions can bedivided into 3 types, corresponding to time periods of0:00e3:00, 3:00e6:00 and 6:00e24:00. In the time periodof 0:00e3:00, the emissions between TAZs with entertain-ment functions are highest. In the time period of 3:00e6:00,emissions keep a low level. In the time period of 6:00e24:00,

emissions of different TAZs varies obviously in time andspace. For emissions within TAZs, the variations of emissionsare analogous with emissions between TAZs.

(3) There is a positive relationship between road densities andemissions within and between TAZs. Furthermore, the pat-terns between road densities and emission quantitiesresemble across various pollutants. When the road densitiesof TAZs are bigger than 0.6, emissions within TAZs will riseup obviously with the increasing road densities. It can beconcluded that TAZ road densities have an influence ontraffic emissions within and between TAZs.

(4) The policy implications of results are discussed, including theregion function planning, traffic planning, the public trans-port improvement and optimal design of charging stationsand new energy vehicles promotion.

(5) The result of this paper is meaningful for urban taxi emissionanalysis. It should be noted that this study can be extended.For emission analysis in other traffic scenes, the emissionestimation model should be localized. In the further study,the experiment on emission estimations and comparisonwillbe conducted.

Acknowledgements

This research is supported in part by National Natural ScienceFoundation of China (U1509205) and Tsinghua University InitiativeScientific Research Program (2015THZ01). The authors are gratefulto anonymous reviewers for their constructive comments.

References

Agrawal, A., Udmale, S.S., Sambhe, V.K., 2018. Extended four-step travel demandforecasting model for urban planning. In: Mishra, D.K., Nayak, M.K., Joshi, A.(Eds.), Information and Communication Technology for Sustainable Develop-ment. Springer Singapore, Singapore, pp. 191e198.

Cao, X., Yao, Z., Shen, X., Ye, Y., Jiang, X., 2016. On-road emission characteristics ofVOCs from light-duty gasoline vehicles in Beijing, China. Atmos. Environ. 124,146e155.

Carslaw, D.C., Beevers, S.D., Tate, J.E., Westmoreland, E.J., Williams, M.L., 2011.Recent evidence concerning higher NOx emissions from passenger cars andlight duty vehicles. Atmos. Environ. 45 (39), 7053e7063.

Ciscal-Terry, W., Dell'Amico, M., Hadjidimitriou, N.S., Iori, M., 2016. An analysis ofdrivers route choice behaviour using GPS data and optimal alternatives.J. Transport Geogr. 51, 119e129.

Crane, R., Crepeau, R., 1998. Does neighborhood design influence travel?: Abehavioral analysis of travel diary and GIS data. Transport. Res. Transport En-viron. 3 (4), 225e238.

D'Andrea, E., Marcelloni, F., 2017. Detection of traffic congestion and incidents fromGPS trace analysis. Expert Syst. Appl. 73, 43e56.

Diao, Q., Sun, W., Yuan, X., Li, L., Zheng, Z., 2016. Life-cycle private-cost-basedcompetitiveness analysis of electric vehicles in China considering the intan-gible cost of traffic policies. Appl. Energy 178, 567e578.

Frey, H.C., Unal, A., Rouphail, N.M., Colyar, J.D., 2003. On-road measurement ofvehicle tailpipe emissions using a portable instrument. J. Air Waste Manag.

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700 699

Assoc. 53 (8), 992e1002.G�omez, C.D., Gonz�alez, C.M., Osses, M., Aristiz�abal, B.H., 2018. Spatial and temporal

disaggregation of the on-road vehicle emission inventory in a medium-sizedAndean city. Comparison of GIS-based top-down methodologies. Atmos. Envi-ron. 179, 142e155.

Gois, V., Inventar, H.M., Proj, E., Unip, Lda, R., Sarmento, Nogueira, L., Almeida, C.,Torres, P., Mesquita, S., 2004. A Detailed Urban Road Traffic Emissions InventoryModel Using Aerial Photography and GPS Survey.

Hardy, A., 1994. An examination of procedures for determining the number ofclusters in a data set. In: Diday, E., Lechevallier, Y., Schader, M., Bertrand, P.,Burtschy, B. (Eds.), New Approaches in Classification and Data Analysis. SpringerBerlin Heidelberg, Berlin, Heidelberg, pp. 178e185.

Harrington, W., Mcconnell, V.D., Cannon, M., 1998. A Behavioral Analysis of EPA'sMOBILE Emission Factor Model. Federal Highway Administration.

He, W., Kai, H., Li, D., 2014. Intelligent carpool routing for urban ridesharing bymining GPS trajectories. IEEE Trans. Intell. Transport. Syst. 15 (5), 2286e2296.

Helveston, J.P., Liu, Y., Feit, E.M., Fuchs, E., Klampfl, E., Michalek, J.J., 2015. Willsubsidies drive electric vehicle adoption? Measuring consumer preferences inthe U.S. and China. Transport. Res. Pol. Pract. 73, 96e112.

Hendriksen, P., Rijkeboer, R.C., 1993. Regulated and Unregulated Exhaust GasComponents from LD Vehicles on Petrol, Diesel. LPG and CNG. TNO, Delft.

Hilditch, C., 1969. Linear skeleton from square cupboards. Mach. Intell. 6, 403e420.Hofer, C., J€ager, G., Füllsack, M., 2018. Large scale simulation of CO2 emissions

caused by urban car traffic: an agent-based network approach. J. Clean. Prod.183, 1e10.

Hsueh, Y.-L., Chen, H.-C., 2018. Map matching for low-sampling-rate GPS trajec-tories by exploring real-time moving directions. Inf. Sci. 433e434, 55e69.

2013. (Ten Thousand Taxi Drivers in Beijing do not Work in Peak Hours). http://news.sina.com.cn.

Huang, Z., Cao, F., Jin, C., Yu, Z., Huang, R., 2017a. Carbon emission flow from self-driving tours and its spatial relationship with scenic spots e a traffic-relatedbig data method. J. Clean. Prod. 142, 946e955.

Huang, Z., Zheng, P., Xu, W., Ren, G., 2017b. SAE for the prediction of road trafficstatus from taxicab operating data and bus smart card data. Int. J. Mod. Phys. C48, 1750121-1e1750121-10.

Huo, H., Wang, M., Zhang, X., He, K., Gong, H., Jiang, K., Jin, Y., Shi, Y., Yu, X., 2012a.Projection of energy use and greenhouse gas emissions by motor vehicles inChina: policy options and impacts. Energy Pol. 43, 37e48.

Huo, H., Yao, Z., Zhang, Y., Shen, X., Zhang, Q., Ding, Y., He, K., 2012b. On-boardmeasurements of emissions from light-duty gasoline vehicles in three mega-cities of China. Atmos. Environ. 49, 371e377.

Int Panis, L., Broekx, S., Liu, R., 2006. Modelling instantaneous traffic emission andthe influence of traffic speed limits. Sci. Total Environ. 371 (1), 270e285.

Jang, K., Kim, K., Kim, H.Y., Kang, S., 2017. A multi-level analysis of the relationshipbetween urban built environment and severe injury traffic crashes. KSCE J. CivilEng. 22 (4), 1427e1435.

Joubert, J.W., Meintjes, S., 2015. Repeatability & reproducibility: implications ofusing GPS data for freight activity chains. Transp. Res. Part B Methodol. 76,81e92.

Joumard, R., Laurikko, J., Han, T.L., Geivanidis, S., Samaras, Z., Mer�etei, T., Devaux, P.,Andr�e, J.-M., Cornelis, E., Lacour, S., Prati, M.V., Vermeulen, R., Zallinger, M.,2009. Accuracy of exhaust emission factor measurements on chassis dyna-mometer. J. Air Waste Manag. Assoc. 59 (6), 695e703.

L�opez-Aparicio, S., Guevara, M., Thunis, P., Cuvelier, K., Tarras�on, L., 2017. Assess-ment of discrepancies between bottom-up and regional emission inventories inNorwegian urban areas. Atmos. Environ. 154, 285e296.

Lam, L., Lee, S.W., Suen, C.Y., 1992. Thinning methodologies-a comprehensive sur-vey. IEEE Trans. Pattern Anal. Mach. Intell. 14 (9), 869e885.

Lee, J., Yasmin, S., Eluru, N., Abdel-Aty, M., Cai, Q., 2018. Analysis of crash proportionby vehicle type at traffic analysis zone level: a mixed fractional split multino-mial logit modeling approach with spatial effects. Accid. Anal. Prev. 111, 12e22.

Li, H., Kulik, L., Ramamohanarao, K., 2015. Robust inferences of travel paths fromGPS trajectories. Int. J. Geogr. Inf. Sci. 29 (12), 2194e2222.

Lin, B., Tan, R., 2017. Are people willing to pay more for new energy bus fares?Energy 130, 365e372.

Lin, X., Tang, D., Ding, Y., Yin, H., Ji, Z., 2009. Study on the distribution of vehiclemileage traveled in China (In Chinese). Res. Environ. Sci. 22 (3), 377e380.

Liu, H., Wu, B., Liu, S., Shao, P., Liu, X., Zhu, C., Wang, Y., Wu, Y., Xue, Y., Gao, J., Hao, Y.,Tian, H., 2018. A regional high-resolution emission inventory of primary airpollutants in 2012 for Beijing and the surrounding five provinces of NorthChina. Atmos. Environ. 181, 20e33.

Liu, W., Teng, J., 2014. Analysis and optimization of the capacity bottlenecks ofelevated bus rapid transit system. Proced. Soc.. Behav. Sci. 138, 386e393.

Luo, X., Dong, L., Dou, Y., Zhang, N., Ren, J., Li, Y., Sun, L., Yao, S., 2017. Analysis onspatial-temporal features of taxis' emissions from big data informed travelpatterns: a case of Shanghai, China. J. Clean. Prod. 142, 926e935.

Ma, Y., Ke, R.-Y., Han, R., Tang, B.-J., 2017. The analysis of the battery electric vehicle'spotentiality of environmental effect: a case study of Beijing from 2016 to 2020.J. Clean. Prod. 145, 395e406.

Mamakos, A., Dardiotis, C., Martini, G., 2012. Assessment of particle number limitsfor petrol vehicles. European Commission Joint Research Centre, Institute forEnergy and Transport 2e47.

Mao, L., Yin, L., Song, X., Mei, S., 2016. Mapping intra-urban transmission risk ofdengue fever with big hourly cellphone data. Acta Trop. 162, 188e195.

MEE, M.o.E.a.E., P. R. China., SAMR, S.A.f.M.R., P. R. China., 2001. Limits and

Measurement Methods for Emissions from Light-duty Vehicles (I).MEE, M.o.E.a.E., P. R. China, SAMR, S.A.f.M.R., P. R. China, 2005a. Limits and Mea-

surement Methods for Crankcase Pollutants from Heavy-duty Vehicles Equip-ped with P.I Engines.

MEE, M.o.E.a.E., P. R. China, SAMR, S.A.f.M.R., P. R. China, 2005b. Limits and Mea-surement Methods for Emissions from Light-duty Vehicles (III, IV).

MEE, M.o.E.a.E., P. R. China, SAMR, S.A.f.M.R., P. R. China, 2005c. Limits and Mea-surement Methods for Exhaust Pollutants from Diesel Engines of Tri-wheel &Low-speed Goods Vehicles.

MEE, M.o.E.a.E., P. R. China, SAMR, S.A.f.M.R., P. R. China, 2013. Limits and Mea-surement Methods for Emissions from Light-duty Vehicles China 5 (in Chinese).

MEE, M.o.E.a.E., P. R. China, SAMR, S.A.f.M.R., P. R. China, 2016. Limits and Mea-surement Methods for Emissions from Light-duty Vehicles(CHINA 6.

MOHURD. P. R. China, M.o.H.a.U.-R.D.o.t.P.s.R.o.C., 2016. Economy Policy Suggestionsabout Developing Urban Public Transport System in Priority.

Munoz-Organero, M., Ruiz-Blaquez, R., S�anchez-Fern�andez, L., 2018. Automaticdetection of traffic lights, street crossings and urban roundabouts combiningoutlier detection and deep learning classification techniques based on GPStraces while driving. Comput. Environ. Urban Syst. 68, 1e8.

Ntziachristos, L., Samaras, Z., 2010. MEP/EEA Emission Inventory Guidebook 2009.Nyhan, M., Sobolevsky, S., Kang, C., Robinson, P., Corti, A., Szell, M., Streets, D., Lu, Z.,

Britter, R., Barrett, S.R.H., Ratti, C., 2016. Predicting vehicular emissions in highspatial resolution using pervasively measured transportation data and micro-scopic emissions model. Atmos. Environ. 140, 352e363.

Osorio, C., Nanduri, K., 2015. Urban transportation emissions mitigation: couplinghigh-resolution vehicular emissions and traffic models for traffic signal opti-mization. Transp. Res. Part B Methodol. 81, 520e538.

Pallavidino, L., Prandi, R., Bertello, A., Bracco, E., Pavone, F., 2014. Compilation of aroad transport emission inventory for the Province of Turin: advantages andkey factors of a bottomeup approach. Atmos. Pollut. Res. 5 (4), 648e655php.weather.sina.com, 2012. Historic weather in Beijing.

Poku-Boansi, M., Cobbinah, P.B., 2017. Land Use and Urban Travel in Kumasi. Geo-Journal, Ghana.

Puliafito, S.E., Allende, D., Pinto, S., Castesana, P., 2015. High resolution inventory ofGHG emissions of the road transport sector in Argentina. Atmos. Environ. 101,303e311.

Seo, T., Bayen, A.M., Kusakabe, T., Asakura, Y., 2017. Traffic state estimation onhighway: a comprehensive survey. Annu. Rev. Contr. 43, 128e151.

Serves, C.D., 2005. Emissions from Flexible Fuel Vehicles with Different EthanolBlends.

State Council P. R. China, 2012. Planning for the Development of the Energy-savingand New Energy Automobile Industry (2012e2020).

State Council P.R. China, 2013. China's Air Pollution Action Plan Pollution PreventionAction Plan.

Sun, L., Ling, X., He, K., Tan, Q., 2016. Community structure in traffic zones based ontravel demand. Phys. Stat. Mech. Appl. 457, 356e363.

Tseng, P.-H., Lin, D.-Y., Chien, S., 2014. Investigating the impact of highway elec-tronic toll collection to the external cost: a case study in Taiwan. Technol.Forecast. Soc. Change 86, 265e272.

USEPA, 2012. MOBILE Model Documentation. U.S.E.P.A.Wang, H., Fu, L., Chen, J., 2010. Developing a high-resolution vehicular emission

inventory by integrating an emission model and a traffic model: Part 2da casestudy in Beijing. J. Air Waste Manag. Assoc. 60 (12), 1471e1475.

Wu, Y., Wang, R., Zhou, Y., Lin, B., Fu, L., He, K., Hao, J., 2011. On-road vehicleemission control in Beijing: past, present, and future. Environ. Sci. Technol. 45(1), 147.

Wu, Y., Zhang, S., Hao, J., Liu, H., Wu, X., Hu, J., Walsh, M.P., Wallington, T.J.,Zhang, K.M., Stevanovic, S., 2017. On-road vehicle emissions and their control inChina: a review and outlook. Sci. Total Environ. 574, 332e349.

Xu, F., Liu, J., Lin, S., Yuan, J., 2017. A VIKOR-based approach for assessing the serviceperformance of electric vehicle sharing programs: a case study in Beijing.J. Clean. Prod. 148, 254e267.

Xu, J., Hilker, N., Turchet, M., Al-Rijleh, M.-K., Tu, R., Wang, A., Fallahshorshani, M.,Evans, G., Hatzopoulou, M., 2018. Contrasting the direct use of data from trafficradars and video-cameras with traffic simulation in the estimation of roademissions and PM hotspot analysis. Transport. Res. Transport Environ. 62,90e101.

Yang, L., 2018. Real-world Emissions in China: a Mets-study of PEMS Emissions Datafrom China 0 to CHina 5/V Light-and-heavy-duty Vehicles.

Yu, B., Kong, L., Sun, Y., Yao, B., Gao, Z., 2015. A bi-level programming for bus lanenetwork design. Transport. Res. C Emerg. Technol. 55, 310e327.

Yu, B., Ma, Y., Xue, M., Tang, B., Wang, B., Yan, J., Wei, Y.-M., 2017. Environmentalbenefits from ridesharing: a case of Beijing. Appl. Energy 191, 141e152.

Yuan, J., Zheng, Y., Xie, X., 2012. Discovering regions of different functions in a cityusing human mobility and POIs. In: ACM SIGKDD International Conference onKnowledge Discovery and Data Mining, pp. 186e194.

Yue, X., Wu, Y., Hao, J., Pang, Y., Ma, Y., Li, Y., Li, B., Bao, X., 2015. Fuel qualitymanagement versus vehicle emission control in China, status quo and futureperspectives. Energy Pol. 79, 87e98.

Zhan, X., Zheng, Y., Yi, X., Ukkusuri, S.V., 2017. Citywide traffic volume estimationusing trajectory data. IEEE Trans. Knowl. Data Eng. 29 (2), 272e285.

Zhang, L., Zhang, J., Duan, Z.Y., Bryde, D., 2015. Sustainable bike-sharing systems:characteristics and commonalities across cases in urban China. J. Clean. Prod. 97,124e133.

Zhang, S., 2014. Characteristics and Control Strategies of Vehicle Emissions in Typical

T. Li et al. / Journal of Cleaner Production 206 (2019) 688e700700

Cities of China Environment Science and Engineering. Tsinghua university.Zhang, S., Wu, Y., Wu, X., Li, M., Ge, Y., Liang, B., Xu, Y., Zhou, Y., Liu, H., Fu, L., Hao, J.,

2014. Historic and future trends of vehicle emissions in Beijing, 1998e2020: apolicy assessment for the most stringent vehicle emission control program in

China. Atmos. Environ. 89, 216e229.Zhu, W., Zhang, J., 2017. An original traffic additional emission model and numerical

simulation on a signalized road. Phys. Stat. Mech. Appl. 467, 107e119.