Portraying Urban Functional Zones by Coupling Remote Sensing Imagery and Human Sensing Data

20 Pages·2017·9.52 MB·English

remote sensing
Article
Portraying Urban Functional Zones by Coupling
Remote Sensing Imagery and Human Sensing Data
WeiTu1,2,3 ID,ZhongwenHu2,4,*,LefeiLi5,JinzhouCao5 ID,JinchengJiang1,2,QiupingLi6
andQingquanLi2,5
1 ShenzhenKeyLaboratoryofSpatialInformationSmartSensingandServices&ResearchInstitutefor
SmartCities,SchoolofArchitectureandUrbanPlanning,ShenzhenUniversity,Shenzhen518060,China;
[email protected](W.T.);[email protected](J.J.)
2 KeyLaboratoryforGeo-EnvironmentalMonitoringofCoastalZoneoftheNationalAdministrationof
Surveying,MappingandGeoinformation,ShenzhenUniversity,Shenzhen518060,China;[email protected]
3 KeyLaboratoryofUrbanLandResourcesMonitoringandSimulation,MinistryofLandandResources,
Shenzhen518034,China
4 CollegeofLifeScienceandOceanography,ShenzhenUniversity,Shenzhen518060,China
5 StateKeyLaboratoryofInformationEngineeringinSurveying,Mapping,andRemoteSensing,
WuhanUniversity,Wuhan430079,China;[email protected](L.L.);[email protected](J.C.)
6 CenterofIntegratedGeographicInformationAnalysis,SchoolofGeographyandPlanning,
SunYat-senUniversity,Guangzhou510275,China;[email protected]
* Correspondence:[email protected];Tel.:[phone]
Received:15December2017;Accepted:16January2018;Published:18January2018
Abstract: Portrayingurbanfunctionalzonesprovidesusefulinsightsintounderstandingcomplex
urbansystemsandestablishingrationalurbanplanning. Althoughseveralstudieshaveconfirmed
theefficacyofremotesensingimageryinurbanstudies,couplingremotesensingandnewhuman
sensing data like mobile phone positioning data to identify urban functional zones has still not
beeninvestigated. Inthisstudy,anewframeworkintegratingremotesensingimageryandmobile
phonepositioningdatawasdevelopedtoanalyzeurbanfunctionalzoneswithlandscapeandhuman
activitymetrics.Landscapesmetricswerecalculatedbasedonlandcoverfromremotesensingimages.
Humanactivitieswereextractedfrommassivemobilephonepositioningdata. Byintegratingthem,
urbanfunctionalzones(urbancenter,sub-center,suburbs,urbanbuffer,transitregionandecological
area)wereidentifiedbyahierarchicalclustering. Finally,gradientanalysisinthreetypicaltransects
wasconductedtoinvestigatethepatternoflandscapesandhumanactivities.TakingShenzhen,China,
asanexample,theconductedexperimentshowsthatthepatternoflandscapesandhumanactivities
intheurbanfunctionalzonesinShenzhendoesnottotallyconformtotheclassicalurbantheories.
Itdemonstratesthatthefusionofremotesensingimageryandhumansensingdatacancharacterize
thecomplexurbanspatialstructureinShenzhenwell. Urbanfunctionalzoneshavethepotentialto
actasbridgesbetweentheurbanstructure,humanactivityandurbanplanningpolicy,providing
scientificsupportforrationalurbanplanningandsustainableurbandevelopmentpolicymaking.
Keywords: urbanfunctions;humanactivity;datafusion;mobilephone;SPOT-5;Shenzhen
1. Introduction
Citiesarehumansettlementswherepeopleengageindifferentactivitiesandinteractwiththe
man-madespaceandnaturalenvironment. Globalurbanareaoccupieslessthan2%oftheEarth’sland
surface,butconsistsofmorethan50percentoftheworld’spopulation[1]. InChina,theurbanization
rate has increased from 26.4 percent in the year 1992 to 57.4 percent in 2016 as 400 million people
havemitigatedtocities[2]. Fromaglobalview,itisestimatedthattotalurbanpopulationwillrise
RemoteSens.2018,10,141;doi:10.3390/rs10010141 www.mdpi.com/journal/remotesensing
RemoteSens.2018,10,141 2of20
to five billion in the year 2030. This transition has enormous economic, social and environmental
consequences[3,4]. Targetingtheaimofsustainablecities,remotesensinghasbeenwidelyusedto
monitorthespatialstructure,economyandenvironmentofcities[5–10].
Manystudieshavebeenconductedonurbanmorphologytoportrayurbanspatialstructure[11–15].
Severalsignificanttheorieshavebeendeveloped,suchastheconcentriczonetheory,thesectortheory,
themultiplenucleitheoryandthepolycentrictheory.Theseadvancedtheoriescapturetheurbanization
processandbenefitassociatedlandmanagementandurbanplanning.Onestandofurbanmorphology
studyistoinvestigatethefunctionprovidedbyurbanspace.Urbanfunctionalzoneisamixtureofurban
functionsandcharacterizedbytheroleofurbanspaceinthewholecity,likeurbancenter,sub-center,
suburbs,ecologicalarea,etc.[16,17].Theidentificationofurbanfunctionalzonesprovidesusefulinsights
forurbanplannerstocapturetheurbangrowthandmakesustainabledevelopmentpolicy.
Urbanfunctionalzoneanalysistraditionallyreliesonlanduseandlandcover(LULC),which
can be acquired by labor- and cost-intensive land survey. Remote sensing is another fast and
efficientapproachtocapturelandcoverandlandusedatatofacilitaterelatedstudies. Forexample,
AubrechtandLeónTorres[15]classifiedmixedorresidentialareasfromnighttimelight(NTL)images.
YangandLo[18]usedtimeseriesLandsatTMimagestoextractlanduse/coverchangedataofthe
Atlanta,Georgia,metropolitanareaintheUnitedStates. Thelandscapegradientfromtheurbancenter
totheruralareahasbeenobservedtoillustrateurbangrowth[19–25]. Usingseverallandscapemetrics,
Linetal.[17]extractedlandusefromPleiadesimagestoinvestigatetheurbanfunctionallandscape
patterninXiamen,China. YuandNg[24]classifiedlandusefromLandsatTMimagesandperformed
gradientanalysistoanalyzespatialandtemporalurbansprawldynamicsinthiscity. Thesestudies
focusonthespatialfeaturesinthecity,butignoretheeffectofhumanactivity. However,inthehighly
urbanizedcitiesinAsia,suchasSingapore,HongKong,BeijingandShenzhen,mostlandparcelsare
coveredbyman-madeinfrastructuresthatareamixbetweenresidential,businessesandworkfunction.
Suchcomplexurbanenvironmentsraiseagreatchallengeinunderstandingurbanstructureusingonly
remotesensingimagery.
Acityisacomplexsystemthatincludeshumanbeingsandthenaturalenvironment. Human
activityhasasignificantimpactonurbanmorphologybecauseoftheinteractionofurbanspaceand
humanbeings. Humanbeingsareun-ignorablecomponentsofthecity. Soarethehumanisticaspects
involved[4,26–29]. However,humanactivitieshavenotbeenwellintegratedwithremotesensing,due
tothelackofmassivehumanactivitiesdata. Ubiquitouslocationawarenesstechnologiessuchasthe
GlobalNavigationSatelliteSystem(GNSS),mobilephonepositioningandWi-Fipositioningallow
humanstoactassensorstoperceivethesurroundingenvironment[30,31]. Massivehumansensing
dataareavailable,suchasvehicleGPSdata[32–34],mobilephonerecords[35–40]andsocialmedia
data [41–44]. These large-volume human sensing data record the time and the position of people;
therefore,theyprovidemuchusefulinformationabouthumanactivitiesinthecity[35,39].
Human sensing data provide us with unprecedented opportunities to reveal human activity
distributionandtheimpliedurbanfunction.Theyenableustoimagethecityinalternativeapproaches.
Ratti et al. [35] mapped the cell phone usage at different times of the day. Their results provided
a graphic representation of city-wide human activities and the evolution through space and time.
Considering the relationship between human activities and land use, Pei et al. [37] developed
a clustering approach to classify land use with time series aggregated mobile phone data. Using
NTLimagesastheproxyofhumanactivity,Chenetal.[36]identifiedtheurbancenterorsub-center
and the surface slope to indicate the urban land use intensity gradient by considering human
activities implicitly. Recently, Cai et al. [44] fused NTL images and social media check-in data to
identifythepolycentricstructureinmegacities,includingBeijing,ChongqingandShanghai,China.
Thesepioneeringstudiessupportthepotentialofhumansensingdatainurbanstudies. However,
human sensing data have still not been integrated with remote sensing imagery to portray urban
functionalzones[45].
RemoteSens.2018,10,141 3of20
In this study, we present a novel data fusion framework integrating remote sensing imagery
and the highly penetrating mobile phone positioning data to analyze urban functional zones
comprehensively.LandscapemetricswerecalculatedbasedonlandcoverfromSPOT5images.Human
activitiesdatawereextractedfrommobilephonepositioningdata. Bycouplingthemwithurbancells,
urbanfunctionalzoneswereidentifiedusinghierarchicalclustering. Gradientanalysis[17–20,46]in
threetypicaltransectswasappliedtoportraythelandscapeandhumanactivitypatternfromurban
centertourbanborder. AnexperimentinShenzhen,China,wasconductedtovalidatetheproposed
framework. Theresultswereexploredtoaddressthefollowingquestions: (1)Whatarethegeneral
patternsoflandscapesindifferenturbanfunctionzones(takingShenzhenasacase)? (2)Whatare
thegeneralpatternsofhumanactivityindifferenturbanfunctionzones(takingShenzhenasacase)?
(3)Whatisthecompositeeffectofremotesensingimageryandhumansensingdatainportraying
urbanfunctionalzones? Theanswerstothesequestionswilldemonstratethespatialdynamicsofboth
landscapeandhumanactivityinthecity. Theydeepentheunderstandingofthecitygrowthandhelp
urbanplannersmakesustainabledevelopmentpolicies.
2. StudyAreaandDataset
2.1. StudyArea
Shenzhen is a coastal city in the south of China, located at the east of the Pearl River Delta,
withanareaof1996km2 (Figure1). AsthefirstSpecialEconomicZone(SEC)ofChina,Shenzhen
has experienced rapid urbanization since 1983 [47]. Shenzhen has expanded from a small village
with 0.6 million people to a megacity with a population of 16 million in 2015 [48]. It consists of
10administrativedistricts,withtheurbancenter(includingFutianandLuohu)insouthernShenzhen,
whereShenzhencitybegan. Inrecentyears,theurbanizedareaofShenzhengraduallyexpandedfrom
theurbancentertothewest(NanshanandBaoan),thenorth(LonghuaandGuangming)andtheeast
(Yantian,Longgang,PingshanandDapeng).
Figure1.ThestudyareainChina.
2.2. Dataset
2.2.1. RemoteSensingImages
ThemultispectralandthepanchromaticSPOT-5imagescollectedon30November2013wereused
tomapthelandcoversofShenzheninthisstudy. Thesemultispectralimagesandthepanchromatic
imageswerefusedusingapansharpeningmethodembeddedinPCIGeomaticaandthenmosaicked
using the ENVI software. The obtained image has 37,368 × 19,440 pixels, 4 spectral channels and
RemoteSens.2018,10,141 4of20
2.5 m/pixel. The final images were obtained by cropping the mosaicked image using the vector
coveringShenzhencityasshowninFigure2. ItimpliesthatmostofShenzheniscoveredbybuilt-up
areasandgreenland. Mostofthebuilt-upareasareaggregatedatthesouth,west,northandincentral
Shenzhen. East Shenzhen, which is reserved as a natural park, has many forests, mountains and
beaches. WeusedtheseremotesensingimagestoextractlandcoverinShenzhen.
Figure2.TheSPOT-5images.
2.2.2. MobilePhonePositioningData
Themobilephonepositioningdatausedwereprovidedbyadominantmobilecommunication
companyinShenzhen.Theywererecordedbythe5349celltowersofthiscompany,coveringthewhole
city. Thisdatasetcontainstimeseriespositionsof9.2millionmobilephoneusersduringoneworkday
inMarch2012. Thepositionsofmobilephoneuserswererecordedatahalf-hourinterval;thus,there
are48recordsforeachuseriftheuser’smobilephonedoesnotturnoff. Forauser,eachpositioning
record has four fields, including a user ID (i), a time stamp (t), longitude (x ) and latitude (y ).
it it
Thespatialresolutionoflocationisrestrictedatthecelltowerlevel,whichisapproximately100–500m.
Intotal,thereare245millionrecordsinthisdataset. Figure3displaysthespatialdistributionofcell
towersandmobilephonepositioningdata. Itdemonstratesthatmanycelltowersaredistributedin
theurbanizedarea. AfewcelltowersarelocatedintheforestalongroadsineastShenzhen. Italso
indicatesthatmoremobilephonepositioningdataareaggregatedattheurbancenterinthesouth,
butlessinthenorth,whichimpliesthedifferenceofhumanactivityintensity. Usingthisusefuldata,
dailyhumanactivities(in-home,workingandsocialactivity)ofmobilephoneuserswereextractedto
portrayurbanfunctionzoneswithlandcoverdatafromremotesensingimagery. Itshouldbenoted
herethathumanactivitiesarehaveregularity;therefore,itisreasonabletofusetheseone-daymobile
phonepositioningdatawithremotesensingimagery.
Figure3. Spatialdistributionofmobilephonepositioningdata: (a)5349celltowersinShenzhen;
(b)245millionmobilephonepositioningrecords.
RemoteSens.2018,10,141 5of20
3. Methodology
Adatafusionframeworkcouplingremotesensingimageryandhumansensingdataispresented
to portray urban functional zones. The workflow is illustrated in Figure 4. Urban landscapes are
extracted from remote sensing imagery; city-wide human activities are recognized from massive
mobilephonepositioningdata. Couplinglandscapesandhumanactivitymetricsintheurbancell,
a hierarchal clustering method is used to identify urban functional zones. Finally, the pattern of
landscapesandhumanactivitiesindifferenturbanfunctionalzonesandtypicaltransectsisanalyzed.
Figure4.Theworkflowofthepresentedframework.
3.1. Landscapes
3.1.1. ImageSegmentation
Anobject-basedimageanalysis(OBIA)approachwasusedtoquantifythelandscapesfromSPOT
5images. TheOBIAapproachisusuallyusedforclassifyinghighspatialresolutionremotesensing
images,becauseofthehighaccuracybyemployingspectral,textural,spatialandcontextualfeatures.
Inthisapproach,theSPOT-5imagewasfirstpartitionedintonon-overlappingregions,whichwere
usedasbasicunitsforclassification,insteadofindividualpixels. Manyimagesegmentationmethods
couldbeusedforthistask,suchaswatershedtransformation[49]andthegraph-basedmethod[50].
However,themostpopularmethodusedinremotesensingapplicationsisthemulti-resolution/scale
imagesegmentationmethod[51],becauseitcanproducemulti-resolutionandhigh-qualityresults.
Inthisstudy,theregionmerging-basedmulti-scaleimagesegmentationmethodwasused[52],and
the multi-scale regions were recorded using the bi-level scale-sets model (BSM) [53]. The BSM
recordsthehierarchicalregionsusingascale-indexedtreestructureandisveryefficienttoretrievethe
segmentationresultsofdifferentscalesandtoestimateanoptimalscaleparameter.
IntheimplementationoftheBSM,thecriterionproposedbyBaatzandSchäpe[54]wasused,
whichwasthemostwidely-usedcriterioninmulti-resolutionimagesegmentation. Therearetwokey
parametersthatcontrolthequalityofthesegmentationresults: theweightofcolorfeaturesw and
color
theweightofcompactnessw . Ahighw willresultinasegmentationresultwithhighspectral
compt color
homogeneity, however, with irregular shapes. A high w will make the regions more compact.
compt
Fromanumberofexperiments,twoempiricalparametersw =0.3,w =0.5wereset.
color compt
AnotherkeyparameterforanOBIAapproachisthescaleparameteraforthemulti-scaleimage
segmentation. IntheBSMwork,segmentationresultsofdifferentscalescouldberetrievedefficiently,
andtheoptimalonecouldbechosenusingsomeunsupervisedmethods,suchastheestimationofscale
parameter(ESP)method[55],theglobalscore(GS)method[56]andtheoverallgoodnessF-measure
(OGF)[57]. Inthisstudy,differentsegmentationresultsfromtheminimumscaletothemaximum
scalewereretrieved,withanintervalof1. Then,theOGFmethodwasusedtoestimatetheoptimal
scaleparameter,whichwasusedtocontrolthefinenessofthefinalresult. Alargerawillresultin
a finer result. To avoid the irreparable negative impact of under-segmentation from a number of
RemoteSens.2018,10,141 6of20
experiments, weusedalargeparametera=3, wherethefinalresultwasslightlyover-segmented.
Then,thissegmentationresultwasusedforfeatureextractionandclassification.
3.1.2. Object-BasedImageClassification
Theclassificationisusedtoenrichthelandcoverinformationforsegmentedobjects. Landcover
isrecognizedwithasupervisedclassificationmethod. Foreverysegmentedobject,spectral,shapeand
texturalfeaturesareextractedfortheclassification.
The minimum, maximum, mean value, the standard deviation of the original four spectral
channelsandthemeanofnormalizeddifferencevegetationindex(NDVI)wereusedasthespectral
features. Theareaindexandtheperimetershapeindex[58]wereusedastheshapefeature. Sincethe
focusofthisstudyistoanalyzethestructureofthecity,aspecifictextureindexnamedPANTEX[59]
wasused.ThePANTEXprocedureemploysgrey-levelco-occurrencematrix(GLCM)contrastmeasures
tocalculatearotation-invariantisotropictextureindex.Itissensitivetodenseresidentialbuildings[60].
Thus,itisusuallyusedtoextractbuilt-upareasfromhighspatialresolutionremotesensingimages
inahigh-densitycitylikeShenzhen. Inthisstudy,themaximum,minimumandmeanvalueofthe
PANTEXindexofeachsegmentwereusedasthetexturefeatures. Basedonthesefeatures,allthe
segmentedobjectswereclassifiedusingthesupport-vector-machine(SVM)classifier[61]. Sixhundred
andtwenty-nineobjectsdistributedaroundthecitywereselectedasthesamplestotraintheSVM
classifier,andthentheclassifierwasappliedtolabeltheclassforeachsegmentedobject. Finally,all
segmentedobjectswereclassifiedintofivetypelandcovers,includingbuilt-uparea,road,greenland,
wateranddevelopingareas. Taking600samplesrandomlydistributedinthecity,theaccuracyofland
coverwasevaluated. Ourresultsindicatethatthelandcoverresultsachievedanoverallaccuracy
95.33%andaKappacoefficient0.9433. Figure5displaystheobtainedlandcovermap. Itsuggested
thatShenzhenhasbeenhighlyurbanizedasmanyareasarecoveredbybuilt-upareaorgreenland.
Figure5.LandcoverfromSPOT-5images.
3.1.3. LandscapeMetrics
Landscape metrics quantify the specific spatial characteristics of LULC. A suite of landscape
metrics has been developed for landscape analysis, such as class area (CA), patch density (PD),
thenumberofpatches(NP),patchshapeindex(PSI)andShannon’sdiversityindex(SHDI)[61–63].
Table1liststhelandscapemetricsusedinthisstudy. Theclass-levelmetricscontaintotalclassarea
(CA)andthepatchdensity(PD).Thelandscape-levelmetricsincludeNPandSHDI.Thisiscalculated
byEquation(1):
∑m
SHDI = − p ln(p ), (1)
i=1 i i
wherep istheproportionofthelandscapeoccupiedbypatchtypesiandmisthenumberoftypes.
i
RemoteSens.2018,10,141 7of20
Table1.Spatiallandscapemetricsselectedinthisstudy.
Name Description
Totalclassarea(CA) Totalareaofoneclasslandscape
Patchdensity(PD) Patchcountpersquarekmofoneclasslandscape
Numberofpatches(NP) Totalnumberofpatches
Equalsminusthesum,acrossallpatchtypes,oftheproportional
Shannon’sdiversityindex(SHDI)
abundanceofeachpatchtypemultipliedbythatproportion
ThevectordataofthelandscapeswereconvertedtoarasterformatatapixelsizeusingArcGIS
10.2. Landscapemetricswerethencalculatedattheclassandlandscapelevelsusingtherasterversion
oftheFRAGSTATSprogram(Version4.2)[64].
3.2. HumanActivity
3.2.1. ActivityDetection
Therawmobilephonepositioningdatalackactivitysemanticinformationsuchas“in-home”,
“working” or “social activity”. Because of the natural rhythm of human beings, human activities
displayregularityinbothspaceandtime[65]. Therefore,itisreasonabletorecognizehumanactivities
fromsequentialmobilephonepositioningdata. Itshouldbenotedherethatdailyhumanactivities
refertoin-home,workingandsocialactivity(i.e.,shopping,education,sports,etc.) thattakemore
thanonehour.
City-widedailyhumanactivitieswereextractedfrommassivemobilephonepositioningdata.
Potential human activities without type information were first detected. A user’s mobile phone
recordsweresortedbytimeandconnectedasaspatial-temporaltrajectory(Figure6a). Then,iftwo
consecutive records are at the same location, in other words, the user does not move, a potential
activityisidentified,suchasp –p ,p –p ,p –p inFigure6b;suchthat,thetrajectorywassplitinto
1 2 5 6 7 8
twotypesofsegments: theverticalsegmentsrepresentinghumanactivitiesatfixedplacesandthe
slopesegmentsindicatingamovebetweenplaces.
Figure6.Humanactivitydetectionfrommobilephonerecords.(a)Thetrajectoryofamobilephone
user;(b)trajectorysegmentation. Verticalsegmentsdenotehumanactivitiesatfixedplaces. Slope
segmentsdenotemovesbetweenplaces.
Itshouldbenotedthatthepositioningjumpbetweenneighbortowersexistsbecauseofsignal
noises,i.e.,thepointp . Toovercomethisissue,followingTuetal.[39],adistancethresholddwas
3
usedtofilterfalsemoves: ifthedistancefromapointtothenextpointislessthand,themovecan
beomitted,andthispointcanbemergedintocurrentpointsofapotentialactivity. Consideringthe
RemoteSens.2018,10,141 8of20
positioningerrorofmobilephonesensingdata,wesetdto500m. Afterprocessingallmobilephone
positioningdatapersonbyperson,city-widehumanactivitieswithouttypeinformationwereobtained.
3.2.2. Rule-BasedActivityLabeling
Dailyhumanactivitieshaveanaturalregularitybecauseoftherhythmofhumanbeings[60].
Regardingmobilephonepositioningdata,thespatial-temporalcharacteristicsoftrajectoryindicate
the information about the home and workplace of a user [38]. Considering the diurnal rhythm of
humanbeings[33],thepotentialhumanactivitieswerelabeledwith“in-home”,“working”and“social
activity”accordingtotimewindows. Thetimewindowforin-homeactivitywassetas[0:00,6:00],
while the time window for working activity was set as [9:00, 17:00]. The semantic information of
potentialhumanactivitieswasenrichedasfollows:
• In-homeactivitylabeling: Forauser,ifthetotaldurationinafixedplaceismorethanhalfof
theearlymorningperiod[0:00–6:00],thisplacewillbedefinedashome. Allpotentialactivities
locatedatthehomeofthispersonarerecognizedasin-homeactivities.
• Workingactivitylabeling: Ifthetotaldurationinaplaceismorethanhalfofthedailyworking
period[9:00–12:00]and[14:00–17:00],thisplacewillbedefinedastheworkplace. Considering
the living style, the time [12:00, 14:00] for lunch is eliminated from working time to avoid
biases. Finally, all potential activities located in the workplace of this user are defined as the
workingactivities.
• Social activity labeling: The remaining non-in-home or non-working activities are labeled as
socialactivities.
Finally, 31,669,042 activities were obtained for the 9.2 million users. Table 2 reports the summary
of human activities. Among these, 14,470,460 potential human activities were labeled as in-home,
10,459,657asworkingand6,738,925associalactivities. Figure7displaysthedistributionofhuman
activities. ItindicatesthatthreetypesofhumanactivitiescovermostoftheareaofShenzhen,inline
withthebuilt-upareainFigure5generally. However, veryfewhumanactivitiesarefoundinthe
lakes, mountains and forests. To verify the accuracy of the results, they were compared with the
household travel investigation of 2010 in Table 2. This suggests that they are matched well as the
biggestgapis2.9%,appearingforin-homeactivities. Therefore,theyareacceptablefortheintegration
withlandscapesfromremotesensingimagery.
Table2.Comparisonofhumanactivitiesfromhumansensingdatawithhouseholdtravelinvestigation.
Name In-Home Working SocialActivities
Count 14,470,460 10,459,657 6,738,925
Ratio 45.7% 33.0% 21.3%
Ratiointhehouseholdtravelinvestigation2010dataset 42.8% 34.3% 22.9%
Figure7.Humanactivityfrommobilephonerecords.(a)Densityofin-home;(b)densityofworking;
(c)densityofsocialactivity.
RemoteSens.2018,10,141 9of20
3.2.3. HumanActivityMetrics
Asetofmetricssummarizingin-home,workingandsocialactivitiesisusedtoquantifyhuman
activities, including the mean µ , the standard variation σ and the ratio r of each type of activity
j j j
(j=in-home,workingandsocialactivity),thenumberofcelltowers,theaverageareaofcelltower
areaandthedensityofhumanactivity.
3.3. HierarchicalClusterAnalysis
The whole city was split into a grid with a resolution of 2000 m, following the approach of
Linetal.[17].Landscapemetricsandhumanactivitymetricswereintegratedinspatialcells.Therefore,
eachgridcellwasdescribedbyavectorv = {v , v }combiningbothlandscapemetricsv andhuman
s a s
activitymetricsv . Thelandscapemetricsv containsCA andPD (i=built-uparea,road,greenland,
a s i i
wateranddevelopingarea)attheclasslevelandNPandSHDIatthelandscapelevel. Thehuman
activitymetricsincludesthemeanµ ,thestandardvariationσ andtheratior ofeachtypeactivity
j j j
(j=in-home,workingandsocialactivity),thenumberofcelltowers(NCT),theaverageareaofthecell
towerservicearea(ACTSA)andthedensityofhumanactivities,d.
Thehierarchicalclusteranalysis[66]wasusedtogroupgridcellsfromthebottomtothetop.Before
theclustering,bothlandscapeandhumanactivitymetricswerenormalized.Wardclustering[67,68]for
minimizingtheincreaseddistanceofwithin-clusterdistancewasused. Thewithin-clusterdistanceis
definedasEquation(2).
D = ∑m (cid:107)i−m (cid:107)2 (2)
A i∈A A
whereiisanelementinClusterA,m isthecenterofClusterAandmisthenumberofelements.
A
TheincreaseddistancemergingtwoclustersisdefinedasEquation(3):
m+n m n
∆DAB = ∑ (cid:107)i−mA∪B (cid:107)2− ∑ (cid:107)i−mA (cid:107)2− ∑ (cid:107)i−mB (cid:107)2 (3)
i∈A∪B i∈A i∈B
whereiisanelementinClusterAorClusterB,m andm arethecentersofclustersandmandnare
A B
thenumberofelementsinClusterAandB.
Finally,gridcellsinthedifferentclusterswereidentifiedastheproperurbanfunctionalzones
withexpertknowledge. Gradientanalysisinthreetypicaltransectswasconductedtoinvestigatethe
patternoflandscapesandhumanactivities.
4. Results
4.1. UrbanFunctionalZones
Allurbancellsaredividedintosixcategoriesthatrepresenttypicalurbanfunctionzones: urban
center, sub-center, suburbs, transit region, urban buffer and the ecological area. The dendrogram
generatedbythehierarchalclusteranalysis(Figure8a)suggeststhesimilarityofgridcellswiththe
sameurbanfunctions. Thespatialdistributionofurbancellswithdifferentfunctionsisdisplayedin
Figure8b. Theurbancentercontains42cells,mostofwhichareinNanshan,FutianandLuohudistrict.
Thesub-centerhas108cells,appearingatthewestandthenorthofShenzhen,andonlytwoappearin
eastShenzhen. Thesuburbhas78gridcells,mostgatheringatnortheastShenzhen. Theurbanbuffer
contains40cells,attheboundaryoftheurbanizedareaortheecologicalarea. Theecologicalarea
has119cellstoprovideecologicalserviceforthepublic. Thetransitregionhas89cells,connecting
differentfunctionalareas.
RemoteSens.2018,10,141 10of20
Figure8.Urbanfunctionalzones.(a)Thedendrogram.(b)Clusteringresult(withsixclusters).
Table 3 reports the composition of landscapes and human activities in urban functional
zones. It indicates that built-up area and green land are two main landscapes in Shenzhen city.
Fromacity-wideview,therankingoflandscapesis: greenland>built-uparea>developingarea>
water>road. Regardingurbanfunctionalzones,thecompositionsoflandscapesaredifferent. Urban
center,sub-centerandsuburbshaveasimilarranking: built-uparea>greenland>road,developing
areaorwater. However,sub-centerratherthanurbancenterhasthemostbuilt-uparea,whichdefies
commonsense. Thisimpliesthatthelandcoverisnotsufficienttoidentifyurbancenterandsub-center
inShenzhen. Transitregionshavethemostdevelopingareaat57.3haperurbancell. Urbanbuffer
nearlakeshaslessbuilt-uparea(46.2percell),butmorewater(77.6percell). Theremainingecological
areaisdominatedbygreenland,about330.9hapercell. Theseresultscharacterizedthestatusofthe
landscapeindifferenturbanfunctionalzones.
Table3.Compositionoflandscapeandhumanactivity.
LandscapeComposition(ha/cell) HumanActivity(n/ha)
Name No.ofCells Area(km2) Built-Up Developing Social
Water GreenLand Road In-Home Working
Area Area Activity
Global 476 1845 148.6 17.8 186.3 14.0 24.3 78.4 56.7 36.5
nal Urbancenter 42 164.2 206.7 15.0 146.1 12.7 10.4 322 260.7 195.2
Urbanfunctio zones ETUcrSoaruSlbnoubasgb-nitciucebrarneublgtfsaeiforreerna 118740198089 433142404652812.....88543 2125260268896.....28889 111780737.....64616 11137839368160.....27869 112211309....3.3134 511297427.....73898 11226238801.5.4 4181855.82..342 2573141..18.8.15
These results also suggest that human activities are differentiated in the six urban functional
zones. Therankingofthreekindsofhumanactivitiesinallfunctionalzonesis: in-home>working
>socialactivity. Althoughurbancenterandsuburbssharesimilarlandscapes,urbancenterhasthe
mostin-home(332perha),working(260.7perha)andsocialactivity(195.2perha). Sub-centerhasthe
secondmostin-home(128.4perha),working(82perha)andsocialactivity(51.5perha),whichareless
thanthatinurbancenter. Thesuburbshavemedianin-home(60perha),working(45.4perha)and
socialactivity(24.1perha). Theecologicalareahastheleastin-home(13.5perha),working(8.3perha)
andsocialactivity(3.8perha). Consideringtheintensityofhumanactivities,therankingis: urban
center>sub-center>suburbs>transitregion>urbanbuffer>ecologicalarea. Thesevariationsof
humanactivitiessuggestthedifferentiationoffunctions.
ThedendrograminFigure8afurtherdemonstratestheclusteringhierarchy,whichsuggeststhe
compositeeffectofbothlandscapeandhumanactivity. Itcanbeseenthat,iftheclusternumberis
reduced,thegroupofecologicalareaandurbanbufferwillbefirstlymergedbecauseofthesimilar
landscapeandhumanactivity. Thespatialdistributionofthecorrespondingcellsalsosuggeststhat
theysharespatialboundaries.Iftheclusternumberfurtherreduces,suburbsandtransitregionswillbe
merged. Afterthat,thethirdmergingwillappearatthesub-centerandurbancenter. Suchahierarchy
validatesthecompositeeffectoflandscapeandhumanactivityonrevealingtheurbanfunctionalzones.