利用人工智能深度學習網絡進行考古遺址檢測

SurfaceTopography:MetrologyandProperties

PAPER?OPENACCESS

Towardstheuseofartificialintelligencedeeplearningnetworksfordetectionofarchaeologicalsites

Tocitethisarticle:AlexandraKaramitrouetal2022Surf.Topogr.:Metrol.Prop.10044001

Viewthe

articleonline

forupdatesandenhancements.

Youmayalsolike

UnravellingtheroleofironandmanganeseoxidesincolouringLateAntiqueglassbymicro-XANESandmicro-XRFspectroscopies

FrancescaGherardi,ClémentHole,EwanCampbelletal.

GeomagneticandgeoelectricalprospectionforburiedarchaeologicalremainsontheUpperCityofAmorium,aByzantinecityinmidwesternTurkeyYunusLeventEkinci,?alayanBalkaya,Ayselerenetal.

UnmannedAerialVehicle(UAV)DataAcquisitionforArchaeologicalSiteIdentificationandMapping

WHandayani,EAAyuningtyas,FSCandraRetal.

ThiscontentwasdownloadedfromIPaddress11on22/04/2024at16:36

Surf.Topogr.:Metrol.Prop.10(2022)044001

/10.1088/2051-672X/ac9492

OPENACCESS

RECEIVED

5February2022

REVISED

25August2022

ACCEPTEDFORPUBLICATION

23September2022

PUBLISHED

3October2022

Originalcontentfromthisworkmaybeusedunderthetermsofthe

Creative

CommonsAttribution4.0

licence

.

Anyfurtherdistributionofthisworkmustmaintainattributiontotheauthor(s)andthetitleofthework,journalcitationandDOI.

PAPER

Towardstheuseofarti?cialintelligencedeeplearningnetworksfordetectionofarchaeologicalsites

AlexandraKaramitrou

1

,?

,FraserSturt

1

,PetrosBogiatzis

2

andDavidBeresford-Jones

3

1UniversityofSouthampton,DepartmentofArchaeology,UnitedKingdom

2OceanandEarthScience,NationalOceanographyCentreSouthampton,UniversityofSouthampton,UnitedKingdom

3UniversityofCambridge,DepartmentofArchaeology,UnitedKingdom

?Authortowhomanycorrespondenceshouldbeaddressed.

E-mail:

a.karamitrou@soton.ac.uk

Keywords:archaeology,machinelearning,arti?cialintelligence,convolutionalneuralnetworks,segnetSupplementarymaterialforthisarticleisavailable

online

AbstractWhileremotesensingdatahavelongbeenwidelyusedinarchaeologicalprospectionoverlargeareas,thetaskofexaminingsuchdataistimeconsumingandrequiresexperiencedandspecialistanalysts.However,recenttechnologicaladvancesinthe?eldofarti?cialintelligence(AI),andinparticulardeeplearningmethods,openpossibilitiesfortheautomatedanalysisoflargeareasofremotesensingdata.Thispaperexaminestheapplicabilityandpotentialofsuperviseddeeplearningmethodsforthedetectionandmappingofdifferentkindsofarchaeologicalsitescomprisingfeaturessuchaswallsandlinearorcurvilinearstructuresofdifferentdimensions,spectralandgeometricalproperties.Ourworkdeliberatelyusesopen-sourceimagerytodemonstratetheaccessibilityofthesetools.OneofthemainchallengesfacingAIapproacheshasbeenthattheyrequirelargeamountsoflabeleddatatoachievehighlevelsofaccuracysothatthetrainingstagerequiressigni?cantcomputationalresources.Ourresultsshow,however,thatevenwithrelativelylimitedamountsofdata,simpleeight-layer,fullyconvolutionalnetworkcanbetrainedef?cientlyusingminimalcomputationalresources,toidentifyandclassifyarchaeologicalsitesandsuccessfullydistinguishthemfromfeatureswithsimilarcharacteristics.Byincreasingthenumberoftrainingsetsandswitchingtotheuseofhigh-performancecomputingtheaccuracyoftheidenti?edareasincreases.Weconcludebydiscussingthefuturedirectionsandpotentialofsuchmethodsinarchaeologicalresearch.

Introduction

Analysisofaerialimageryrevolutionizedarchaeologyintheearlytwentiethcentury,exponentiallyincreas-ingthenumberofknownsites,allowinglargeareastoberapidlysurveyedandgivingaccesstoremoteregions(Reeves

1936

,BewleyandRaczkowski

2002

;Mossunetal

2013

;Lambers

2018

).Forexample,asearchforscienti?cpublicationsrelatedwithArchae-ologyandRemoteSensingusingtheDimensionsscienti?cresearchdatabasereturns2,732articleson2013,5,172on2018and14,323in2021(

https://app.

dimensions.ai

;accessedinMay2022).

Withtheintroductionofawiderrangeofairborne(i.e.,mannedaircraftanddrones)andspace-baseddata,includingpassivehighspatialresolutionopticalsensors,multispectralandhyperspectralsensors,light

detectionandranging(LIDAR),Syntheticapertureradar(SAR),thermalsensorsandgeophysicalimages,theamountofdataavailabletoarchaeologistshasalsoincreasedexponentiallyinrecentyears(e.g.,Chietal

2016

;Tamiminiaetal

2020

).Thesedataholdsig-ni?cantpotentialtotransformourunderstandingofthearchaeologicalrecord,butalsoposeasigni?cantchallengewithregardstotheamountoftimeanalysiswouldtakeusingtraditionalhuman-ledimageanaly-sismethods.

Arti?cialIntelligence(AI)offersapotentialbypasstothisbottleneckandthereforesubstantiallyreducetherequiredhumanlabor.AIdescribestheabilityofcomputerstoperformtasksandreachingdecisionsthroughlearningeitherdirectlyfromthedata(unsu-pervisedmethods)orfrompastexperiencewherethecorrectoutcomeisprovided(supervisedmethods),

?2022TheAuthor(s).PublishedbyIOPPublishingLtd

Surf.Topogr.:Metrol.Prop.10(2022)044001

AKaramitrouetal

PAGE

10

imitatinghumanintelligence(e.g.,Dey,

2016

;Copeland

2020

).

Overthepastthreedecades,applicationsofmachinelearning(ML)methodshaveseensigni?cantincreaseinArchaeology.MLalgorithmssuchassup-portvectormachine(CortesandVapnik

1995

;Kaoetal

2004

)randomforests(Ho

1995

;Ho

1998

),K-means(Caoetal

2009

;JinandHan,

2011

;Qietal

2017

)andothersimilarapproacheshavebeenwidelyadoptedwithconsiderablesuccessindetectingorclas-sifyingarchaeologicalsites,andartifacts(e.g.,KintighandAmmerman

1982

;Baxter

2009

;MenzeandUr

2012

;Floresetal

2019

;Orengoetal

2020

).Thesemethods,oftenreferredtoastraditionalMLalgo-rithms,requirethecarefulselectionofinputfeatures(e.g.,variousspectralindicesinsatelliteimaging)byhuman-experts,thatareimportantfortheoutcome.Thenthroughaniterativeoptimizationprocessbytheinputofexemplardatathealgorithmistrainedbaseduponmultivariatestatisticsandprogressivelyimprovesitsperformance.Sinceitrequiresthedeter-minationandthepriorcalculationofarangeofpossi-blestatisticallysigni?cantinputfeatures,itinevitablysuffersfromalevelofbiasasalthoughthetrainingprocedurecanpointoutwhichfromthefeaturesarestatisticallyinsigni?cant,itcannotsuggest,orextractfeaturesdifferentthantheprovidedones.Also,therelativelylimitednumberofthefeaturesinmostappli-cationsoftencannotfullydescribethetargetsatdifferentsituationsorenvironmentalconditions.Therefore,theapplicabilityofthesealgorithmsisoftenlimitedtospeci?ccasesandrestrictstheidenti?cationtofeatureswithlimitedspectralandgeometricvariations.

Intheearly2000sanewmachinelearningtechnol-ogyemergedknownasDeepLearning(DL)basedonArti?cialNeuralNetworks(ANN),andinthecaseofimageapplications,ConvolutionalNeuralNetworks(CNNs).ThisnewtechnologywaslargelybasedontheseminalworkofFukushima(

1980

)aswellasHubelandWiesel(

1959

)thatintroducedthe‘neocognitron’(Fukushima

1980

;

1983

;2003)andestablishedtheuseofconvolutionalanddown-samplinglayers.In1986,RinaDecherwasoneofthe?rsttousetheterm‘deeplearning’tothemachinelearningcommunity,inwhich‘deep’wasusedtodescribetheuseofmultiplelayersinanetwork.Later,Waibel(

1987

)proposedthetimedelayneuralnetwork(TDNN),oneofthe?rstconvolutionalnetworksfollowedbyLeCunetal(

1989

)whoappliedthatinahandwrittencharacterrecognitionproblemusinga7-levelConvolutionalNeuralNetowork(CNN),calledLeNet-5(LeCunetal

1998

).Asigni?cantadvantageofdeeplearningmeth-odsisthatthefeatureextractionandselectionstageisperformedbythelearningalgorithmautomaticallyandnotbyaperson.Yet,thisusuallyrequiressig-ni?cantamountsoflabeleddataandconsiderablecomputationalresourcesforthetrainingprocess.TheutilizationofGPUsinthetrainingprocesswasthe

turningpointforusingCNNsinimagerecognition.Inthe2012ImageNetcompetition,the?rstCNNeversubmitted,namedAlexNet(Krizhevskyetal

2012

),wonthecompetition.ThetrainingofAlexNetusedoveronemillionlabeledimagesabout～1000objectcategoriesandtook～6daysusing2GPUs(Krizhevskyetal

2012

).Sincethen,deepneuralnetworkshavewon

manyinternationalpatternrecognitioncompetitionsandhaveattractedbroadattention,byoutperforminglegacymachinelearningmethodsandhandlingbetterlargeamountsofdatawithminimumuserinterven-tion(Schmidhuber

2015

).Assuch,theyoffercon-siderablepotentialforarchaeology.

Amongthecommontasksassignedtodeeplearn-ingCNNnetworksareimageclassi?cation,objectdetection,andsemanticsegmentation.Classi?cationisabasicprocessroutinelyperformedinarchaeologywiththeobjectiveofclassifyinggroupsofimagesthatsharesomecommonfeatures,orobjectsintooneofanumberofprede?nedclasses.Forexample,AImeth-odshavebeenusedtoanalyzeuse-wearonlithictools(e.g.,VandenDries

1998

)andtoclassifyandidentifytypesofpottery(e.g.,H?rretal

2008

;Anichinietal,

2021

;PawlowiczandDownum

2021

).CaspariandCrespo(

2019

),usedanobject-detectionbasedmethodtoidentifyIronAgeburialmoundsinaerialimagery.Morerecently,Agapiouetal(

2021

)appliedtheobjectdetectionmethodtodetectsurfaceceramicsindroneimages.Finally,semanticsegmentationalgorithmsattempttoanalyzeimagesfurther,bypartitioningthemintosemanticallymeaningfulpartsandafter-wardsbyclassifyingeachpartintooneofthe‘X’pre-determinedclassesi.e.,interpretableimageregionsforinstance,archaeologicalsites,regionsofvegetation,modernstructuresandothers(e.g.,Garcia-Garciaetal

2018

;Minaeeetal

2020

).Semanticsegmentationoperatesatpixel-levelinthesensethateachpixelofanimageislabeledaccordingtotheclassitbelongsto.Thismakessemanticsegmentationamuchmorecomplicatedandcomputationallyintensivetask,yetitcanproducemoreinformativeanddetailedresultscomparedtoclassi?cationandobjectidenti?cation(e.g.,Kendalletal

2015

;Garcia-Garciaetal

2018

;Minaeeetal

2020

).ThevalueofthisapproachforgeophysicalanalysishasbeendemonstratedintheworkofKü?ükdemirciandSarris’s(

2020

)usingground-penetratingradarimages.

Forallthissuccess,onlyrecentlytherehavebeenlimitedyetincreasingworkadoptingCNNapproachesfortheautomateddetectionofarchaeologicalsites(Trieretal

2018

;CaspariandCrespo,

2019

;Kazimietal

2019

;Lambersetal

2019

;Rayneetal

2020

;Somraketal

2020

;Soroushetal

2020

;Bonhageetal

2021

;Verschoof-vanderVaartandLandauer

2021

)fromEarthobservation(EO)data.Inpart,thisisduetotheneedforanabundanceoflabeleddatatoenabletheCNNtoaccuratelyidentifydifferentsignatures.Forexample,ImageNet,anopenlyavailablevisualdatabasedesignedforuseineverydaycontemporary

Figure1.Demonstrationoftheconvolutionofanimagewithanedgedetection?lter.Ontheleftistheinitialimage,inthemiddleisanedgedetection?lterandontherightistheresultedimage,whichshowstheedgesoftheinitialimage.

objectrecognitioncomprises14,197,122images(Rus-sakovskyetal

2015

).Itisthisvolumeoflabelleddata,whichhasenabledrapidadvancesintheuseofCNNinday-to-daytasks.Inarchaeologyhowever,similarlytoother?elds,theamountoffreelyavailable,properlylabeleddataiscurrentlylimited.Furthermore,onlinesharingofsuchdataisoftenrestrictedbycon-?dentialityissuesthatariseoftenfromlocallegisla-tion,relatedwiththeefforttoprotectthesesitesfromlooting.

Inthispaper,weofferarouteforwardbyusingopenlyavailablesatellitehighspatialresolutionima-geryandthroughexaminingtwoneuralnetworkarchitectures:TheSegNet(Kendalletal

2015

),adeepconvolutionalencoder-decoderarchitectureforrobustsemanticpixel-wiselabeling;andacustom8-layerCNNdesignedforthisresearch(SimpleNet).Wealsoopen-upaccesstothesetoolsthroughprovid-ingapackagedapplication(supplementaryinforma-tion)allowingreaderstoruntheirownanalysis,helpingthemtoevaluatethestrengthsandweaknessesofthisnetworkandbeginamoreopenandinclusiveconversationabouttheiruseinarchaeology.

Convolutionalneuralnetworks(CNN)

Inthissectionwebrie?yintroducethefundamentalconceptsofCNNs.Althoughamoreextensivepre-sentationofCNNsisbeyondthescopeofthiswork,theinterestedreadercan?nddetailedintroductionsfocusingonvariousaspectsofCNNsinseveralworksincluding,Nielsen(

2015

);Wu(

2017

);Alzubaidietal(

2021

);Lietal(

2021

);andUlkuandAkagün-düz(

2022

).

DeeplearningalgorithmsareatypeofmachinelearningtechniquethatusesANNofseverallayersinahierarchicalarchitecturetoenablemachinestopro-cessdatainanonlinearmanner.Arti?cialneuralnet-worksconsistofcircuitsofsimple,yethighlyinterconnected,nodestoselectivelytransmitsignalsinaprocessthatmimicsthebiologicalneurons(Hop?eld

1982

),therebysimulatingthewaybiologicalneuralnetworkswork.Thesenodesareorganizedin

layerswhichprocessinformationbyoutputtingdynamicstateresponsestoexternalinputs(commonlyaresponsefromapreviouslayer).DataareintroducedtotheANNthroughaninputlayerandresultsdeliv-eredwitha?naloutputlayer.Allintermediatelayersaretermedhiddenlayers,whichcarryoutallthepro-cessing.Thelargerthenumberofhiddenlayers,the‘deeper’thenetwork,enablingtheidenti?cationpro-gressivelyofmorecomplexpatternsanddetails.Forexample,the?rstlayermaylearnrecognizingedgesinanimage,thesecondshapes,thethirdobjectsandsoon.

Informationispassedbetweenlayersthroughcon-

nectionsthatarecharacterizedbyweightsandbiases,sothatthereceivedtotaloutputcorrespondstoaweightedsumofindividualnode-inputs,plussomebias.Theresultoutputmayormaynotexceedathresholdde?nedbyapre-setactivationfunctionsuchasasigmoidormostcommonlyarecti?edlinearacti-vationfunction(ReLU;seebelow),essentiallydecidingifthisinformationshouldbetransmittedtothenextlayer(forwardpassed),asitisorinamodulatedform,orrather?lteredout.Theoptimalvaluesofeachweightandbiasarede?nedbythetrainingofthenet-work:anon-linearoptimizationprocesswherebyacostfunctionrepresentingthedistancebetweentrain-inglabeleddataandthatpredictedbynetworkresultsisminimized.

Thenumberofrequireddeeplayerswithinthe

network,andthereforeindirectlythenumberofunknowns(i.e.,parametersthataretobetunnedthroughthetraining),dependsonthecomplexityofthepatternstobeidenti?edandtheamountoflabeleddata.Atpresent,alimitednumberoflabelledimagesforarchaeologyimposesarequirementforcarefuldesignoflearningnetworks,keepingthenumberoflayersandconnectionslowenoughtoensurethattheoptimizationproblemofnetworktrainingisnotunder-determinedi.e.,thenumberofunknownpara-metersexceedthenumberofdataandpriorcon-straintsthatareusedtoregularize/stabilizethetrainingandreducethegeneralizationerror(over-?tting)(e.g.,Goodfellowetal

2016

).

Figure2.Architectureofthe8-layerconvolutionalneuralnetwork.

Table1.ArchaeologicalsitesinPeruusedtotrainthealgorithm.

Archaeologicalareas&sites CoordinatesWGS84(centrepoint) Period

LaCentinela(ChinchaValley)

?13.450385,?76.171092

Inca(AD1476–1532)LateIntermediate(AD1000–1476AD)

Cahuachi(NazcaValley)

?14.818241,?75.117462

EarlyIntermediate(c.200BC–AD600)

Caral(SupeValley)

?10.890938,?77.521858

LatePreceramic(c.3000–1800BC)

TamboColorado(PiscoValley)

?13.704619,?75.829431

Inca(AD1476–1532)

Table2.Additionalarchaeologicalsites(areas)inPerutofurthertrainthealgorithm.

Archaeologicalareas&sites CoordinatesWGS84(centrepoint) Periods

Various(LowerIcaValley)

?14.614319,?75.614994

Various(1800BC–AD1534)

NazcaGeoglyphs(PampadeSanJosé,NascaValley)

?14.696486,?75.178422

EarlyIntermediate(200BC–AD600AD)

CerroSechín(CasmaValley)

?9.480703,?78.258997

InitialPeriod(1600BC)

Hereweexaminetwodifferent,supervised,fullyconvolutional,neuralnetworks:onebasedonthearchitectureofSemanticSegmentationcalledSegNet(Kendalletal

2015

;Badrinarayananetal

2017

);andtheotheracustom8-layernetworkdevelopedbytheauthorscalledSimpleNet.Botharefullyconvolutionalneuralnetworks,acategoryofnetworkconsistingoflocallyconnectedlayerssothateachneurononlyreceivesinputfromasmalllocalsubgroupofthepixelsintheinputimage.SuchLayerscanperformconvolu-tion/deconvolution,pooling(i.e.,asample-baseddis-cretizationprocessthateffectivelydown-samplestheimage)andup-sampling,butnotcontainingfullycon-nectedlayers,andthusrequiringsigni?cantlylessmemoryandcomputationalpower(Longetal

2015

).SemanticsegmentationalgorithmshavebeenusedwidelyinclassifyingfeaturesinvariousremotesensingimagesincludinghighresolutionGoogleEarthimages(Yuetal

2021

).Additionally,thecustom8-layernet-workwasdesignedtobeimplementedforthelownumberoflabeleddatausedinthiswork.Inthefol-lowingsections,wedescribethearchitectureandfunc-tionalityofthesetwonetworks.

SegNet

SegNetisadeepfullyconvolutionalneuralnetworkthatsegmentstheimagebyclassifyingeachpixelindependently.Itconsistsofanencodernetworkwith13layers,eachdesignedforobjectclassi?cation.Eachlayerisconvolvedusingasetof2D?lterstoproduceasetoffeaturemapsofincreasingcomplexityasdescribedpreviously.Thesemapsarelaterbatchnormalizedi.e.,tohaveameanoutputcloseto0andtheoutputstandarddeviationcloseto1.Next,aReLUactivationfunctionisappliedfollowedbydown-samplingusingamaxpoolinglayerwitha2×2nonoverlappingwindow(Kendalletal

2015

;Badrinaraya-nanetal

2017

).TheReLUactivationfunctionisalinearfunctionthatoutputstheinputifitispositive,orelse,outputszero(Haraetal

2015

).Themaxpoolingfunctioncalculatesthemaximum,ineachpatchofeachfeaturemap(Chollet

2017

).Inthe?nallayertheresultingoutput,fromthepreviousstep,issub-sampledbyafactorof2whiletheboundaryinforma-tionisalsostored.Thisiscrucialasduringthesuccessivedown-samplingoperationsthehighfre-quencydetailsoftheimagearelessenedresultingin

Figure3.Asampleofthe2000Trainingimages,ofsize256×256×3pixels(GoogleEarthimagery),fromvariousarchaeologicalareasaroundPeru.Thetoprowshowstheinitialimagesandthebottomrowthelabeledimages.

blurryandinaccurateboundaries.However,bound-ariesareimportantinsmallobjectsandstructuressuchasbuildings,cropmarksetcandbystoringthisinformationitcanberetrievedduringthedecodingstage.

Thenetworkconsistsof13decoderlayerseachonecorrespondingtoitsrespectiveencoderlayer.Theroleoftheencoderlayersistosemanticallyprojectthelowerresolutionfeaturesextracted(learnt)bytheencoder,ontothehigherresolutionimagespacetogetadenseclassi?cation,i.e.,aclassi?cationforeachpixelintheoriginalsizedimage.Eachdecoderlayerpro-ducesdensefeaturemaps(images)byup-samplingitsinputfeaturemaps(theoutputofthepreviouslayer)usingthememorizedmax-poolingindicesproducedonthepreviousstage.Thenconvolutionisappliedusingatraineddictionaryof?lterstoproducedensefeaturemaps.The?naldecoderoutputisfedintoaSoftMaxclassi?er,i.e.,alayerthatassignseachpixelindependentlytoaclassaccordingtoaprobabilityscoreamongthecandidateclasses(e.g.,Nielsen

2015

;Alzubaidietal

2021

).

2.2.Acustom8-layerconvolutionalneuralnetwork

(SimpleNet)

Sincetheamountoflabeleddataavailableforarchae-ologyislimited,weconstructedacustom8-layerconvolutionalneuralnetwork(SimpleNet),basedontheSegNetarchitecturewiththeaimofkeepingthenumberoflayersandtrainableparametersaslowaspossiblewhileachievingadequatelyaccurateresults.The?rstlayerisanimageinputlayerthatreceivesRGBimages.Thenextlayerisaconvolutionallayerwith32trainable?ltersappliedinanon-overlappingmovingwindowofsize5×5andwithstride1.Strideshows

howmuchthe?ltershiftsaroundtheinputvolume(inourcaseitshiftsbyoneunit)whilethe?lterapproximatestheLaplacian(i.e.,a2Dsecondspatialderivative)oftheGaussianoperatorandessentiallywhenconvolvedwithanimagederivesasanoutputanapproximationofitssecondspatialderivative.Thismeansthatinregionswheretheimagehasconstantintensitythe?lter’sresponsewillbezero.Inregionswheretheintensity(i.e.,pixelbrightness)changesrapidly,however,suchasattheedgesofanobject,the?lter’sresponseyieldshighamplitudes(?gure

1

).

The?lterscanbeconceivedofas2Dimageswhose

shapeandcolorareadjustedthroughthetrainingpro-cesstooptimallyexpressdifferentfeaturesofthedata(e.g.,?gure

2

).Next,arecti?edlinearunit(ReLU)isappliedfollowedbyamax-poolingwitha2×2nonoverlappingwindowwithstride2andapaddingwith0’s.Thisisthemostcommoncon?gurationasitdis-cardsthe75%oftheactivationsinaninputimageduetodown-samplingby2inbothwidthandheight.Fol-lowingthis,atransposeconvolutionisappliedwiththesamenumberof?ltersandawindowwith4×4sizeandstride2.Likewise,thisisacommoncon?g-uration,asthedivisibilityofthewindowsizebythestridemitigatestheproblemofcheckerboardartifactsintheup-sampledimage(e.g.,Odenaetal

2016

).Thesixthlayerisanotherconvolutionallayerof1×1windowsizeandstride1.Then,aSoftMaxclassi?erisapplied,tothe?naloutputfromthepreviouslayer,toassigneachpixelintoaclass.Finally,theimageisseg-mentedintotheassignedclassesbyaclassi?cationlayerthatcalculatestheclassweighedcross-entropyloss(e.g.,Bishop

2006

).The8-layerconvolutionalneuralnetworktechniqueisillustratedin?gure

2

.

Figure4.Histogramillustratingthenumberofpixelsusedineachofthe4classesfortheD500datasetwithorangecolorandfortheD2000datasetwithbluecolor.

Trainingandoptimisation

Data

Weusedopenlyavailablehigh-resolutionimagesfromGoogleEarthofarchaeologicalsitesinPeruasatrainingsetforbothnetworks.Thisgeographicalregionwaschosenforitscontinueddiscoveryofnewsitesusingremotelysenseddata(Ruggles

2015

;Bikoulisetal

2018

;CignaandTapete

2018

)andtheavailabilityofdatafrompreviouslarge-scalearchae-ologicalterrestrialsurveysforevaluationpurposes.

Initially,welabeled500imagesfrom4differentarchaeologicalsites,(table

1

).AsmallpartoftheTamboColoradoarchaeologicalsitewasthenusedtotrainthealgorithmandalargerareaofthesamesitefortesting.

Later,weaugmentedtheoriginal500imageswithafurther1500fromwiderarchaeologicalareasandsitesacrossPerutofurthertrainthealgorithm(table

2

)andcheckitsperformanceasthenumberoflabeleddataincrease.Theseadditionalimagesconsistmostlyofgeoglyphs(usuallylinearfeatures)markedinopendesertpampaenvironments.Figure

3

showssomesamplesoflabeledimagesusedinthiswork.

OptimisationprocessThedatawerelabelledwiththeImageLabelerprograminMatlab9.6usingfourdifferentclasses:‘archaeologi-cal’,‘modern’,‘vegetation’and‘background’.As‘archaeological’weincludedeverytargetofarchae-ologicalinterest,regardlessofshape,condition,color,periodetcWelabeledlinear,rectilinear,andcircularfeaturesthatwereclearlyvisibleintheGoogleEarthimagery,correspondingtoalargevarietyofarchae-ologicalfeatures.Weusedsuchbroadterminologybecausethetargetofthisworkwastofurtherincreasethenumberoftrainingimagesavailabletousersin

Table3.OptimalsetofparametersforSegNet,8-layerD500andD2000networks.

Parametername Value

Gradientdecayfactor 0.9000

Squaredgradientdecayfactor 0.9990Epsilon 1e-08

Initiallearningrate 1e-04(D500)and1e-03(D2000)

Dropratefactor 0.4

Dropperiod 5

L2-regularizationparameter 1e-09Gradientthresholdmethod UsingtheL2-normMaxepochs 30

Minibatchsize 5(D500)and15(D2000)

Shuf?e Ateveryepoch

future,withsub-classi?cationopenasanoptiontothosewhowishtomakeuseofthedataset.As‘modern’welabeledmodernstructuressuchasmodernbuild-ingsandvehicles.‘Vegetation’incorporatesareasofgrass,plants,andtrees.Finally,as‘background’weclassi?edeverythingelse,suchassoil,non-pavedroads,and?eldswithoutvegetation.Imagesintheinitialsetof500weredenotedasD500,andinthelarger2000setasD2000.Imagesforthesites/areasofinterestwereextractedfromGoogleEarthinRGB(Red,Green,Blue)asjpg?les.Ourgoal,istotrainanalgorithmtousehighresolution,freelyavailableGoogleEarthimages.Unfortunately,GoogleEarthdoesnotproviderawimagesthereforewehavetorelyonthealreadyprocessedimagesthataremadeavailablethroughtheGoogleEarthapplication.Itshouldbenotedherethatatpresent,thehigh-resolutionimagesinGoogleEarthapplicationarenotavailableinGoogleEarthEngineandthereforeisnotpossibletousethisenvironmenttotraindataset.

Figure5.SegmentationofthearchaeologicalsiteofTamboColoradowiththe3trainednetworks,(a)GoogleEarth