transferring multiscale map styles using generative adversarial networks yuhao kang资料.pdf

该【transferring multiscale map styles using generative adversarial networks yuhao kang资料 】是由【宝钗文档】上传分享，文档一共【28】页，该文档可以免费在线阅读，需要了解更多关于【transferring multiscale map styles using generative adversarial networks yuhao kang资料 】的内容，可以使用淘豆网的站内搜索功能，选择自己适合的文档，以下文字是截取该文章内的部分文字，如需要获得完整电子版，请下载此文档到您的设备，方便您编辑和打印。:..InternationalJournalofCartographyISSN:2372-9333(Print)2372-9341(Online)Journalhomepage:https:///tica20TransferringmultiscalemapstylesusingworksYuhaoKang,SongGao&:YuhaoKang,SongGao&(2019):Transferringmultiscaleworks,InternationalJournalofCartography,DOI:.1615729Tolinktothisarticle:https:///.1615729Publishedonline:&essandusecanbefoundathttps://ion/journalInformation?journalCode=tica20:..INTERNATIONALJOURNALOFCARTOGRAPHYhttps:///.1615729TransferringmultiscalemapstylesusinggenerativeworksYuhaoKanga,,DepartmentofGeography,UniversityofWisconsin,Madison,WI,USA;bCartographyLab,DepartmentofGeography,UniversityofWisconsin,Madison,WI,USAABSTRACTARTICLEHISTORYTheadvancementoftheArti?cialIntelligence(AI),;generativeadversarialnetwork;styletransfer;Speci?cally,weidentifyandtransferthestylisticelementsfromaconvolutionalneuraltargetgroupofvisualexamples,work;mapdesignOpenStreetMap,andartisticpaintings,tounstylizedGISvectorwork(GAN)?,éSUMéLesavancéesenintelligencearti?cielle(IA)permettentd’apprendreàunemachinelescritè,nousproposonsunnouveaucadrelogicielpourtransféédiésauxcartographiestellesquecellesdeGoogleMaps,OpenStreetMapoumêmelesstylesutilisésparlesartistespeintrespeuventêtreapprisettransférésàdesdonnéesSIGvectoriellesgraceàdeuxréseauxantagonistesgénératifs(GANs).Unclassi?eurbinairebasésurunréseaudeneuronesconvolutifprofondestentrainépourévaluersilesimagesdesstylestransféréspréserventlescaractéésultatsexpérimentauxmontrentquelesGANsontunfortpotentielpourletransfertdesstylescartographiquesmaisqu’ilrestedenombreuxdé?sà-istics(Kent&Vujakovic,2009).Themapstylesetstheaesthetictoneofthemap,evokingavisceral,emotionalreactionfromtheaudiencebasedontheinterplayofform,color,type,andtexture(Gao,Janowicz,&Zhang,2017).Twomapscanhaveaverydi?erentlookandfeelbasedontheirmapstyle,evenifdepictingthesameinformationorregion(Figure1;CONTACTSongGaosong.******@?2019InternationalCartographicAssociation:..(left)andOpenStreetMap(right)stylesforMadison,Wisconsin(USA).GoogleMapshasa?attervisualhierarchytoemphasizelabelsandpointsofinterestsaswellasenablevectoroverlays,&Vujakovic,2009;Stoter,parisonsofin-housestylesofnationalmappingagencies).Arguably,mapstyling–andthemyriaddesigndecisionstherein–isaprimarywaythatthecartographerexercisesagency,authorship,andsubjectivityduringthemappingprocess(seeBuckley&Jenny,2012forrecentdiscussionsonaesthetics,style,andtaste).Increasingly,webcartographersneedtodevelopacoherentanddistinctmapstylethatworksconsistentlyacrossmultiplemapscalestoenableinteractivepanningandzoomingofa‘mapofeverywhere’(Roth,Brewer,&Stryker,2011).Suchmultiscalemapstylingtapsintoarichbodyofresearchongeneralizationandmultiplerepresentationdatabasesincartography(seeMackaness,Ruas,&Sarjakoski,pendiumdevelopedbymissiononGeneralization).Alargenumberofgeneralizationtaxonomiesnowexisttoinformthemultiscalemapdesignprocess(.,2016;DeLucia&Black,1987;Foerster,Stoter,&K?bben,2007;McMaster&Shea,1992;Raposo,2017;Regnauld&McMaster,2007;Stanislawski,Butten?eld,Bereuter,Savino,&Brewer,2014;Shen,Ai,Wang,&Zhou,2018),eometryoperationsformeaningfullyremovingdetailingeographicinformation(,smooth,aggregate,collapse,merge).BrewerandButten?eld(2007),Rothetal.(2011)discusshowcartographerscanmanipulatethevisualvariables,orfundamentalbuildingblocksofgraphicsymbols(,size,orientation,dimensionsofcolorlikehue,value,saturation,andtransparency),-vicesandtechnologiesnowexisttodevelopandrendersuchmultiscalemapstylerulesasinterlockingtilesets,suchasCartoCSS,1MapboxStudio,2TileMill,(seeMuehlenhaus,2012forareview),thesetoolsenablemultiscalewebmapstylingthatisexploratory,playful,andevensubversive(forinstance,seeChristopheandHoarau,2012,forexamplesofmultiscalemapstylingusingPopArtasinspiration).Despitetheseadvances,establishingamapstylethat:..INTERNATIONALJOURNALOFCARTOGRAPHY3worksacrossregionsandscalesremainsafundamentalchallengeforwebcartography,hoicesavailabletothecartographerandthelimitedgui-danceforintegratingcreative,artisticstylesintomultiscalemapslikeGoogleMaps5andOpenStreetMap(OSM).6Here,weaskifarti?cialintelligence(AI)canhelpilluminate,transfer,andultimatelyimprovemultiscalemapstylingforcartography,(Armstrong&Xiao,2018),inwhich‘theproductionofmapsswitchesfromasequenceofactionstakenbyamapmakertoaprocessofspecifyingcriteriathatareusedtocreatemapsusingintelligentagents’.Speci?cally,whetherAIcanlearnmapdesigncriteriafromexistingmapexamples(orworksofart)andthentransferthesecriteriatonewmulti-?cation,segmentation,objectionlocalization,styletransfer,naturallanguageprocessing,andsoforth(Gatys,Ecker,&Bethge,2016;Goodfellow,Bengio,Courville,&Bengio,2016;LeCun,Bengio,&Hinton,2015).Recently,GIScientistsandcartographers,puterscientistshavebeeninvestigatingvariousAIanddeeplearningapplicationssuchasgeographicknowledgediscovery(Hu,Gao,Newsam,&Lunga,2018;Mao,Hu,Kar,Gao,&McKenzie,2017),map-typeclassi?cation(Zhou,Li,Arundel,&Liu,2018),sceneclassi?cation(Law,Seresinhe,Shen,&Gutierrez-Roig,2018;Srivastava,Vargas-Mu?oz,Swinkels,&Tuia,2018;Zhangetal.,2018;Zou,Ni,Zhang,&Wang,2015;Zhang,Wu,Zhu,&Liu,2019),scenegeneration(Deng,Zhu,&Newsam,2018),automatedterrainfeatureidenti?cationfromremotesensingimagery(Li&Hsu,2018),automaticalignmentofgeographicfeaturesincontemporaryvectordataandhistoricalmaps(al.,2017),satelliteimageryspoo?ng(Xu&Zhao,2018),spatialinterpolation(Zhuetal.,2019),andenvironmentalepi-demiology(VoPham,Hart,Laden,&Chiang,2018).Relevanttoourworkonmultiscalemapstyle,works(GANs)havebeendevelopedtogeneratesyntheticphotographsthatmimicrealones(Goodfellowetal.,2014).TheGANsinputrealphotographstotrainthemodel,andtheresultingoutputphotographslookatleastsuper?ciallyauthentictohumanobservers,suggestingapoten-tialapplicationformultiscalemapstyling.-binedwithmulti-workstotransferthestylesofexistingsatelliteimageryandvectorstreetmaps(Isola,Zhu,Zhou,&Efros,2017;Xu&Zhao,2018;Zhu,Park,Isola,&Efros,2017;Ganguli,Garzon,&Glaser,2019).However,,whichfeaturetypes,symbolstyling,andzoomlevelsbestenablemapstyletransfer?Second,binationsofalgorithmsworkbestformapstyletransfer?Finally,howusablearetheresultingmapsafterstyletransfer;dotheresultsappearasauthenticmapsornot?Tothisend,weproposeanovelframeworktotransferexistingstylecriteriatonewmul-tiscalemapsusingGANswithouttheinputofCartoCSSmapstyleconsheets.?gurationSpeci?cally,theGANslearn(1)whichvisualvariablesencode(2)whichmapfeaturesanddistributionsat(3)whichzoomlevels,,wethentraina:..)classi?,wedescribethemethodsframework,includingdatacollectionandpreprocessing,tiledmapgeneration,(USA)?cally,weprovidebothaqualitativevisualassessmentandquantitativeassessmentoftwodi?erentGANmodels,Pix2PixandCycleGAN,attwodi??,weprepareunstyledor‘raw’GISvectordatafromageospatialdatasourcethatwewishtostyle(hereOSMvectordata,whichisgivenaninitialsimplestylingforthepurposeofvisualdisplay;detailsbelow)aswellasexamplestyleddatasourceswewishtoreproduceandtransfer(hereGoogleMapstilesandpaintedvisualart).Second,wecon?guretwogen-workmethodstolearnthemultiscalemapstylingcriteria:Pix2Pix,whichusespairedtrainingdatabetweenthetargetandexampledatasources,andCycle-GAN,whichcanuseunpairedtrainingdata(detailsbelow).Third,weemployadeepcon-N)classi?er(describedasIsMapbelow)tojudgewhethertheoutputswithtransferredmapstylingdoordonotpreservemapcharacteristics(Evansetal.,2017;Krizhevsky,Sutskever,&Hinton,2012).::(1)datapreparation,(2)mapstyletransferusingGANs,(3)IsMapclassi?er.:..,butcontainstylingsymbolsthatrepresentdi?erenttypesofgeographicfeatures(,lakes,roads,andsoon).Tiledmapservicesareamongthemostpopularwebmappingtechnologiesforrepre-sentinggeographicalinformationatmultiplescales(Roth,Donohue,Sack,Wallace,&Buck-ingham,2015).Suchwebmaptilesetsinterlockusingamulti-resolution,,mapscaleisreferredtoaszoomlevelandexpressedin1–20notation,with1denotingthesmallestcartographicscale()and20thelargestcartographicscale().Whilethespatialres-olutiongetscoarserfromthetoptothebottomofthetilepyramid,thesizeofeachimagetileinthetilesetremainsacrosszoomlevels(Peterson,2011,2014),typicallycapturedat256×256pixels(8-bit).Therefore,servingpre--:OpenStreetMap,whichwedownloadedinrawvectorformatwithoutanymapstylesandservedasatiledwebserviceforthesimplestyledmapscaseusingGeoServer,7andGoogleMaps,,multipleclassesoffeaturesexistforeachgeometrytype(,line,andpolygon).Forthesimplestyledmaps,werenderedthesedi?erentclassesusingdi?(EPSG:900913),biningthegeo-(Goodfellowetal.,2014):,whichgeneratesfakeoutputsthatmimicrealexamplesusingtheupsamplingvectorsofrandomnoise,andthediscriminatorD,,Giteratesthroughapresentnumberofepochs(