Assessing OpenStreetMap Completeness for Management of Natural Disaster by Means of Remote Sensing: A Case Study of Three Small Island States (Haiti, Dominica and St. Lucia)

Abstract

Over the last few decades, many countries, especially islands in the Caribbean, have been challenged by the devastating consequences of natural disasters, which pose a significant threat to human health and safety. Timely information related to the distribution of vulnerable population and critical infrastructure is key for effective disaster relief. OpenStreetMap (OSM) has repeatedly been shown to be highly suitable for disaster mapping and management. However, large portions of the world, including countries exposed to natural disasters, remain incompletely mapped. In this study, we propose a methodology that relies on remotely sensed measurements (e.g., Visible Infrared Imaging Radiometer Suite (VIIRS), Sentinel-2 and Sentinel-1) and derived classification schemes (e.g., forest and built-up land cover) to predict the completeness of OSM building footprints in three small island states (Haiti, Dominica and St. Lucia). We find that the combinatorial effects of these predictors explain up to 94% of the variation of the completeness of OSM building footprints. Our study extends the existing literature by demonstrating how remotely sensed measurements could be leveraged to evaluate the completeness of the OSM database, especially in countries with high risk of natural disasters. Identifying areas that lack coverage of OSM features could help prioritize mapping efforts, especially in areas vulnerable to natural hazards and where current data gaps pose an obstacle to timely and evidence-based disaster risk management.

Full text



RemoteSens.2020,12,118;doi:10.3390/rs12010118www.mdpi.com/journal/remotesensing
Article
AssessingOpenStreetMapCompletenessfor
ManagementofNaturalDisasterbyMeansof
RemoteSensing:ACaseStudyofThreeSmallIsland
States(Haiti,DominicaandSt.Lucia)
RanGoldblatt1,*,NicholasJones2andJennyMannix3
1NewLightTechnologiesInc.,Washington,DC20005,USA
2WorldBankGroup,Washington,DC20433,USA;njones@worldbankgroup.org
3NewLightTechnologiesInc.,Washington,DC20005,USA;jennifer.mannix@nltgis.com
*Correspondence:ran.goldblatt@nltgis.com;Tel.:+1‐202‐630–0362
Received:26November2019;Accepted:25December2019;Published:1January2020
Abstract:Overthelastfewdecades,manycountries,especiallyislandsintheCaribbean,havebeen
challengedbythedevastatingconsequencesofnaturaldisasters,whichposeasignificantthreatto
humanhealthandsafety.Timelyinformationrelatedtothedistributionofvulnerablepopulation
andcriticalinfrastructureiskeyforeffectivedisasterrelief.OpenStreetMap(OSM)hasrepeatedly
beenshowntobehighlysuitablefordisastermappingandmanagement.However,largeportions
oftheworld,includingcountriesexposedtonaturaldisasters,remainincompletelymapped.Inthis
study,weproposeamethodologythatreliesonremotelysensedmeasurements(e.g.,Visible
InfraredImagingRadiometerSuite(VIIRS),Sentinel‐2andSentinel‐1)andderivedclassification
schemes(e.g.,forestandbuilt‐uplandcover)topredictthecompletenessofOSMbuildingfootprints
inthreesmallislandstates(Haiti,DominicaandSt.Lucia).Wefindthatthecombinatorialeffectsof
thesepredictorsexplainupto94%ofthevariationofthecompletenessofOSMbuildingfootprints.
Ourstudyextendstheexistingliteraturebydemonstratinghowremotelysensedmeasurements
couldbeleveragedtoevaluatethecompletenessoftheOSMdatabase,especiallyincountrieswith
highriskofnaturaldisasters.IdentifyingareasthatlackcoverageofOSMfeaturescouldhelp
prioritizemappingefforts,especiallyinareasvulnerabletonaturalhazardsandwherecurrentdata
gapsposeanobstacletotimelyandevidence‐baseddisasterriskmanagement.
Keywords:OpenStreetMap;OSM;OpenStreetMapcoverage;disastermanagement;remotesensing

1.Introduction
Overthelastfewdecades,manycountrieshavebeenchallengedbythedevastating
consequencesofnaturaldisasterswhichposeasignificantthreattohumanhealthandsafetyand
impactvulnerablecommunitiesandcriticalinfrastructureglobally.Everyyear,naturaldisasters
impactcloseto160millionpeopleworldwide[1],causingdestructionofthephysical,biologicaland
socialenvironments,impactingfoodsecurity,andcausinggloballossesthatamounttoover100
billiondollars[2].Thefrequencyofnaturaldisastershasbeensteadilyincreasingsince1940[3]and
overthenextcentury,climatechangewilllikelyamplifythenumberandseverityofsuchdisasters
[4].
Whiletheimpactsofnaturaldisastersareworldwide,somecountrieshavebeenmorevulnerable
todifferenttypesofdisastersthanothers[5].Forexample,in2017,PuertoRico,SriLankaand
Dominicawereatthetopofthelistofthemostaffectedcountriestonaturaldisasterssuchas
significantprecipitation,floodsandlandslides.Caribbeanislandcountriesareespeciallyexposedto

RemoteSens.2020,12,1182of27
awiderangeofnaturaldisasters[6]andsmallislanddevelopingstates—whicharefrequently
characterizedbycoastalcommunities,geographicisolation,andlimitedtechnicalcapacity—are
amongthemostvulnerablecountriestonaturaldisastersandclimatechange[7].
Recognizingthesetrends,thereisanincreasingneedforefficientandwell‐planneddisaster
managementanddisasterreliefoperations.Thetermdisasterriskmanagementreferstothefull
lifecycleofactionsaimingtoprevent,preparefor,respondto,andrecoverfromdisasters.Generally,
disasterriskmanagementconsistsoffourmainphases:(1)Mitigation,i.e.,activitiesthatreducethe
likelihoodandexpectedadverseimpactsofanaturaldisasterevent;(2)Preparedness,i.e.,plansor
preparationstostrengthenemergencyresponsecapabilities;(3)Response,i.e.,actionstakentosave
livesandpreventpropertydamageinanemergencysituation;and(4)Recovery,i.e.,interventions
aimedatreturningcommunitiesandinfrastructuretoaproperleveloffunctionalityfollowinga
disaster.
Timelygeospatialinformationindicatingthedistributionofvulnerablepopulationandthe
location,availabilityandfunctionalityofcriticalinfrastructureiskeyforeffectivedisasterrelief
operations.Untilrecently,governmentalagenciesandthecommercialsectorweretheprimary
sourcesforgeospatialdatafordisastermanagement.Inthepastdecade,however,thepublichasbeen
increasinglyrecognizedasavaluablesourceforgeospatialinformationfordisastermanagement[8].
Recentdevelopmentsinwebmappingtechnologieshaveledtodisastermanagementoperationsthat
aremoredynamic,transparent,anddecentralized,withanincreasedcontributionbyindividualsand
organizationsfrombothinsideandoutsidetheimpactedarea[9],includingbymeansofgeospatial
informationthatiscontributedbyvolunteers.ThetermVolunteeredGeographicInformation(VGI)
referstogeographicinformationcollectedbyindividuals,oftenonavoluntarybasis[8].Thisdatais
madeopen,freelyaccessible[10]andfillsthedeficienciesoftraditionalmappingtechnologiesand
sourcesofdata[11,12].Governmentsindevelopingcountriesincreasinglyrecognizetheeconomic
andsocialvalueofVGIsandtheirpotentialtoprovidenewwaystointeractwiththecommunityand
forstrengtheningcivilsociety[13].
1.1.OpenStreetMap(OSM)forDisasterManagement
Createdin2004,OpenStreetMap(OSM)isacollaborativeuser‐generatedmappingproject
aimingtoprovideafreelyavailablegeographicalinformationdatabaseoftheworld[14].Duringthe
firstyearoftheproject,mostmappingeffortsfocusedonroadandtransportationnetworks.Today,
however,avarietyofgeographicalfeaturesareconstantlyaddedtoOSM’sdatabase,including
buildingsandtheirfunctionality,landuseandpublictransportationinformation[15].Thisdata
allowslocalgovernmentsandcommunitiestobetterperformriskassessmentandemergency
planning[16–18]andisroutinelyutilizedforvariousdisasterriskmanagementapplications[19,20].
Asoftoday,therearemorethan5.5millionOSMusersandonemillioncontributorswhogenerate
morethan3millionchangeseveryday.Inthecontextofnaturaldisasters,thecoordinationof
volunteers’mappingeffortsisoperatedbytheHumanitarianOpenStreetMapTeam(HOT),
originallyformedaftertheHaitiearthquake,whichconductsactivitiesaimedatenrichingOSMdata
tosupportemergencyreliefoperations(https://wiki.openstreetmap.org/wiki/Stats).Often,when
disastersoccur,thereisalackofthisessentialinformation,whichresultsinmappingcampaigns,
includingMapathons[21],whicharedesignedtomaptheimpactedareas.
OSMdataiscollectedbythreemainmeans[22]:(1)usingGPSrecords,whichcanbeuploaded
tothedatabase;(2)relyingonorthophotosandhigh‐resolutionsatelliteimagerytotraceanddigitize
features;or(3)importingdatasetsfromexternalsourcessuchasadministrativecensusdata.Recently,
largecorporations,includingApple,Microsoft,andFacebookhavebeenhiringeditorstocontribute
totheOSMdatabase[23].SeveraltoolshavealsobeendevelopedtosupportOSM’smappingefforts,
oneofthemistheMapSwipeapp(https://mapswipe.org/),whichenablesvolunteerstomapandtag
geographicalfeaturesonmobilephonesbasedonsatelliteimagery[24].Otherinitiatives,suchas
MissingMaps(https://www.missingmaps.org/)allowvolunteerstracefeaturesbasedonsatellite
imagesbysplittingmappingintosmalltasks,allowingremotevolunteerstoworksimultaneouslyon

RemoteSens.2020,12,1183of27
thesameoverallarea(asof2018,therewerenearly60,000mapperscontributingtoMissingMaps)
[25].
1.2.AssessingOSMCompletenessandAccuracy
AlthoughOSMroadnetworkdataisestimatedtoexceed80%completenessinrelationtothe
world’sroadsandstreets[26],ingeneral,thecoverageandcompletenessofOSMfeatures(including
buildingfootprints)varysignificantly—notonlybetweencountries,butalsowithincountries.For
example,completenessofcoverageofremoteandruralareasisoftenlowerthanthatofhighly
populatedurbanareas[27],andthecoverageofdevelopedcountriestendstobelowerthanthatof
developingcountries[28–32].Thesedifferencesareinpartduetosocietalfactors,suchaspopulation
distributionandpopulationdensity,distancetomajorcitiesandthelocationofcontributingusers
[29,33–37].
WiththeincreasedutilizationofVGIs—includingOSM—fordisasterpreparednessand
response,variousmethodologieshavebeenproposedtoassessthequalityandtheaccuracyofthe
collecteddata[12];forexample,intermsofdatacompleteness,logicalconsistency,positional,
thematic,semanticandspatialaccuracy,temporalqualityandusability[28,38–42].Several
approacheshavebeenproposedtoassessthecompletenessoftheOSMdatabaseandthe
completenessofthestreetnetworks[32,43],thelanduseandthebuildingfootprints[44].The
completenessofthecoveragecanbeassessedbycomparingtheOSMmappedfeatureswithexternal
datasets,forexample,nationaladministrativedata[28,44–47].Suchdatavariesbycountryandisnot
alwaysmadeavailable—especiallyindevelopingcountries.
Inthisstudy,weproposeamethodologythatutilizesremotelysensedobservationstoestimate
thecoverageofOSMmappedfeatures,specificallytoidentifygapsinthecompletenessofOSM
buildingfootprints.Inthepast,expensivesatelliteimageryandlimitedcomputationalpoweronly
allowedanalysisofsmallgeographicalcontexts.Thismodelisbeingreplacedthankstothe
accessibilityofpubliclyavailableandfreesatellitedatathatcaptureeverylocationoneartheveryfew
days.Theavailabilityofdaytime(e.g.,Sentinel‐2,Landsat)andnighttime(e.g.,DMSPOperational
LinescanSystem(OLS)ortheVisibleInfraredImagingRadiometerSuite(VIIRS))satelliteimagery,
togetherwithadvancementsinthecapabilitiesofcloud‐basedcomputationalplatforms,nowallows
foranalyzingLandUseandLandCover(LULC)characteristicsofEarthacrossagreatergeographic
andtemporalscale.LandcoverreferstotheattributesoftheEarthlandsurfaceanditsimmediate
subsurface(e.g.,biota,soil,typography,surface,groundwaterandhumanstructure).Landuserefers
tothepurposeforwhichhumansexploitthelandcover[48].Becauseremotelysensedobservations
typicallycapturetheuniquereflectancecharacteristicsofphysicalobjectsonEarth,mostremote
sensingapplicationsfocusondetectionandclassificationofEarth`slandcovercharacteristics.
DifferentiationbetweendifferenttypesofLandUse(whichtypicallydonotholduniquephysical
characteristics)remainschallenging.InrespecttoOSM,mappedfeaturescanbetaggedaccordingto
both,theirlanduseandlandcover.AlthoughOSMcontributorsarefreetousetheirowntags,there
isaquasi‐officialcollectionoftagsthathasbeenestablishedandagreedupon(forexample,“landuse”
and“landcover”keysorothermorespecifickeyssuchas“building”or“highway”)[49].Previous
studiesdemonstratedthepotentialuseofthesetagstocreatedetailedLULCmaps[50].
Themethodologyweproposeinthisstudyreliesonremotelysensedmeasurementstoestimate
thecoverageofOSMbuildingfootprintsandtoidentify“mappinggaps”(i.e.,areasthathavenotyet
beenmapped).PreviousstudieshaveutilizedOSMdatafordifferentremotesensingapplications,
forexample,forclassificationofurbanareas[51]orforsemanticlabelingofaerialandsatelliteimages
[52].Despitesignificantprogressinthefieldofmachinelearningandtheincreasingavailabilityof
satelliteimagery,thereisstillascarcityofstudiesaimingtoutilizeremotelysensedobservationsto
estimatethecompletenessofOSMbuildingfootprintsatagivenpointintime.Identifyingareasthat
lackcoverageofOSMfeaturescouldhelpplanandprioritizemappingefforts,especiallyinareasthat
arevulnerabletonaturalhazardsandwherecurrentdatagapsposeanobstacletotimelyand
evidence‐baseddisasterriskmanagementactions.

RemoteSens.2020,12,1184of27
Byitsnature,theOSMdatabaseisdynamicandisupdateddailywiththousandsofnewentries.
However,asdiscussedabove,thefrequencyandextentofupdatesvarylargelybygeographical
areas.Someregionsarebeingupdatedmorefrequentlythanothers,andespeciallydeveloping
countriesarenotfullymapped,whichareoftenthemostvulnerabletotheimpactsofnatural
disasters.Theobjectiveofthisstudyistoproposeamethodologytoestimatethecompletenessof
OSMbuildingfootprintsbasedonremotelysensedmeasurementsthatareavailableataglobalscale
andareupdatedfrequently.Wedemonstrateourmethodologyinthecasestudyofthreesmallisland
states:Haiti,DominicaandSt.Lucia.
Theremainderofthisarticleisorganizedasfollows.InSection2,wediscussthemethodology,
thestudyareaandthedataweusetopredictthecoverageofOSMbuildingfootprints.InSection3,
wepresentandevaluatetheresultsinthecaseofHaitiandinSection3.3,weillustratethe
applicabilityofourapproachinthecaseofDominicaandSt.Lucia.InSection4,weofferaconcluding
discussion.
2.MaterialsandMethods
2.1.StudyAreas
Wedemonstrateourmethodologyinthecaseofthreesmallislandstates:Haiti,Dominicaand
St.Lucia(Figure1).
2.1.1.Haiti
LocatedonthewesternsideofHispaniolaIsland,Haiti(27,750km2insize,withapopulationof
approximately11.5million)isthepoorestcountryintheWesternHemisphere,withaGrossDomestic
Product(GDP)percapitaofUS$870[53].Haitiishighlyvulnerabletonaturaldisasters;morethan
96%ofitspopulationisexposedtodifferenttypesofnaturalhazards,particularlyhurricane,coastal
andriverineflood,andearthquake[53].Morethanhalfofthepopulationlivesincitiesandtowns,a
majorshiftfromthe1950swhenapproximately90%ofHaitianslivedinthecountryside[54].Almost
allofHaiti‘s30majorwatershedsexperiencesignificantfloodevents,duetointenseseasonalrainfall,
stormsurgeinthecoastalzones,deforestationanderosion,andsediment‐ladenriverchannels[55].
Furthermore,largeportionsofthecountry`spopulation(e.g.,inthecapital,Port‐au‐Prince)livein
shantytownsbuiltuponsteepandexposedhillsides[56].In2018alone,some2.8millionpeoplewere
consideredtobeinneedofhumanitarianassistancevaluedatUS$252.2million[57].
2.1.2.St.Lucia
AsmallwindwardislandstatelocatedintheCaribbeanSeaandtheNorthAtlanticOcean,St.
Lucia(616km2insize)hasapopulationofapproximately165,000[58]andaGDPpercapitaofUS$
10,315[59].St.Luciaissusceptibletonumerousnaturalhazards,includinghurricanes,landslides,
flooding,andvolcaniceruptions.Itsterrainconsistsmainlyofmountainsandsteepslopesinthe
centerofthecountryduetoitsvolcanicoriginswithlow‐lyingareasalongthecoasts[60].Asof2018,
approximately19%ofthepopulationresidesintheselow‐lyingareas[61].Inaddition,St.Lucia’s
economyishighlydependentontwosources:theexportofbananasandincomefromtourism.Both
havebeennegativelyimpactedrecentlybynaturalhazardssuchasin2016whenHurricaneMatthew
caused70%oftheislandtolosepoweranddamaged80%ofthecountry’sbananaplantations[60].
2.1.3.Dominica
Dominica(approximately74,000people[62])islocatedinLeewardIslandschainintheLesser
AntillesoftheCaribbeanSea,approximately1,200kmsoutheastofHaiti,withlargeportionsofits
populationresidinginthecapitalRoseau(population14,700)andPortsmouth(population5,200)[62].
Dominicaisvulnerabletoawiderangeofnaturalhazards,includinghurricanes,intenserainfall,
slopeinstability,volcaniceruptions,seismicactivities,andtsunamis[63].Reflectingarugged
physicaltopography,mostofthepopulationandinfrastructurearelocatedonthecoast,makingthe

RemoteSens.2020,12,1185of27
countryparticularlyvulnerabletostrongwindsandhighseas[64].InSeptember2017,aCategory5
hurricaneMariahitthecountry,causinglossesanddamagesworth226percentofGDP[65].

Figure1.Locationsandsizecomparisonsofthethreestudyareas:Haiti,Dominica,andSt.Lucia.
2.2.AnalyticalFramework
TheobjectiveofthisstudyistoidentifygapsinthecompletenessofOSMbuildingfootprintsin
threesmallislandstates(Haiti,St.LuciaandDominica)basedonremotelysensedmeasurementsand
othergeospatialfeatures.Theprocedureinvolvessevensteps.
2.2.1.Step1:ConstructanArtificialTessellation
Weconstructanartificialtessellatedgridofcellsthatspaneachofthecountries;eachcellis0.25
squarekminsize(atotalof136,747gridcellsoverHaiti,2,796gridcellsoverSt.Luciaand3,861grid
cellsoverDominica).Eachgridcellwastreatedanindependentunitofanalysis.
2.2.2.Step2:DownloadtheCurrentOSMBuildingFootprints
Wedownloadedthemostup‐to‐dateOSMdataforthethreecountries(datadownloadedinJuly
2019).ForHaiti,wedownloadedthedata(inaShapefileformat)fromGeofabrik

(https://www.geofabrik.de/data/download.html).Atthetimeoftheanalysis,Geofabrikdidnothave
dataforDominicaandSt.Lucia;thus,wedownloadedthedataforthesecountriesfromoverpass
turbo(https://overpass‐turbo.eu/)inaKMLformat(thisdatarequiresadditionalpre‐processingand
weselectedOSMfeaturesthatarelabeledas“building=Yes”).Atthetimeoftheanalysis,therewere
930,000mappedbuildingsinHaiti,38,619mappedbuildingsinDominicaand29,412mapped
buildingsinSt.Lucia.
2.2.3.Step3:CalculateTotalAreaofOSMBuildingFootprintsinaGridCell
WecalculatedthetotalareaofOSMbuildingfootprintsineachgridcell.Thisisthevaluetobe
predictedbytheexplanatoryvariables(theremotelysensedandgeospatialmeasurements).
2.2.4.Step4:PreprocessandAggregatetheRemotelySensedandGeospatialData
Wereliedonseveralpredictors(explanatoryvariables)toestimatethecoverageofOSMbuilding
footprintsinagridcellandtoidentifygapsinOSMcoverage.Wepreprocessedthedataand
aggregatedittothelevelofagridcells(Table1providesadescriptionoftheevaluatedexplanatory
variablesandtheaggregationmeasures).Thepreprocessing,analysisandaggregationofthe
remotelysenseddatawerecompletedbyusingGoogleEarthEngine(GEE).GEEisaplatformthat
leveragescloud‐computingservicestoachieveplanetary‐scaleutilityandhasbeenpreviouslyused
forawiderangeofapplications[66],includingmappingpopulation[67,68]andurbanareas[69,70].

RemoteSens.2020,12,1186of27
NighttimeLights(VIIRS):TheVisibleInfraredImagingRadiometerSuite(VIIRS)isone
ofthekeyinstrumentsonboardtheSuomiNationalPolar‐OrbitingPartnership(SuomiNPP)
spacecraft(launchedin2011).VIIRSinstrumentcollectsvisibleandinfraredimageryandglobal
observationsofland,atmosphere,cryosphereandoceans.Thisinstrumenthassignificant
improvementsoverthecapabilitiesoftheformerDMSP‐OLS[71],notablyitsavailabilityonadaily
basisandhigherspatialresolution(upto500mattheequator).TheVIIRSDNBprovidesglobal
coveragewith12‐hourrevisittime.First,werecordforeachpixelthemaximumvalueofall
overlappingpixels(inthesamelocation)inastackofsevenmonthlycomposites(Jan–July)of2019.
Then,foreachgridcell,wecalculatedaSumofLight(SOL)measure(calculatedasthesumofthe
digitalnumbervaluesofalloverlappingpixelsineachcell).
Sentinel‐2‐DerivedSpectralIndices:TheCopernicusSentinel‐2missioncomprisesa
constellationoftwopolar‐orbitingsatellitesthatcollectmultispectraldatain13spectralbands,with
fourbandsataspatialresolutionof10mand6bandsataspatialresolutionof20m.Therevisitperiod
ofSentinel‐2is5daysattheequator.Wecalculatedfourremotelysensedmeasuressensitiveto
vegetationandbuilt‐uplandcover:NormalizedDifferenceVegetationIndex(NDVI)[72],Soil
AdjustedVegetationIndex(SAVI)[73],NormalizedDifferenceBuilt‐upIndex(NDBI)[74]andUrban
Index(UI)[75].Foreachgridcell,wecalculatedaper‐indexsumvalueofallpixelsoverlappingwith
thegridcell.
Sentinel‐1SAR:Sentinel‐1missioncomprisesaconstellationoftwopolar‐orbitingsatellites,
performingC‐bandsyntheticapertureradarimaging,enablingthemtoacquireimageryindayand
nightconditionsregardlessoftheweather.Sentinel‐1hasa12‐dayrepeatcycle,withaspatial
resolutiondownto5m.Similarlyto[70],wecapturedthetextureofthesurfacebyutilizingSentinel‐
1’sC‐band(singleco‐polarizationverticaltransmitandverticalreceive(VV)acquisitionmodewith
anInterferometricWideSwath(IW)instrumentmode,a250kmswathat5mby20mspatial
resolution(singlelook)).Fromeachscene,weremovedspecklenoiseandperformedradiometric
calibrationandterraincorrection.Tocreatetheannualcomposites,wecalculatedforeachlocation
(pixel)themedianvalueofalloverlappingpixelsinanentirestackofallscenescapturedin2019.For
eachgridcell,wecalculatedtheaveragevalueofallpixelsincorporatedwithintheareaofthegrid
cell.
Slope:Tocapturethetopographyofthesurface,weusedtheGlobalSRTMmTPIdataset
(availableinGEEinaspatialresolutionof270m),wherealocalgradientiscalculatedforeachpixel
basedontheglobalSRTMDEMelevationdata(30mresolution).ThemTPIdistinguishesridgefrom
valleyformsandiscalculatedusingelevationdataforeachlocationsubtractedbythemeanelevation
withinaneighborhood[76].Foreachgridcell,wecalculatedtheaveragevalueofallpixelsinthe
gridcell.
ForestCover:Weestimatedtheextentofforestcoverin2018basedontheHansenGlobal
ForestChangev1.6(2000–2018)[77].First,wedefinedapixelas“forest”intheyear2000ifmorethan
20%ofitwascoveredin2000withforest.Werecordedpixelsthatexperiencedamajoreventofforest
coverlossbetween2000and2018andestimatethetotalareaofforestcoverin2018pergridcell.
UrbanFootprints:Wereliedontworemotelysensedderivedproductssignifyingurbanand
ruralsettlementsthatwereproducedbytheEarthObservationCenteratDLR:TheGlobalUrban
Footprint(GUF)(inaspatialresolutionof~12m)andtheWorldSettlementFootprint(WSF)(ina
spatialresolutionof~10m)[78–80].
OSMTransportationNetworkFeatures:WecalculatedthetotallengthofOSMroads
inacellandthetotalnumberofjunctionsinacellasadditionalpotentialpredictorsofOSM‐building
footprints.
Table1.Thepredictorsusedtopredictper‐cellareaofOpenStreetMap(OSM)buildingfootprints.
PredictorSourceNumber
ofscenes
Per‐cellstatistics

RemoteSens.2020,12,1187of27
NighttimelightsVIIRS7SumofLight(SOL):ThesumofDNmaxvalueof
allpixelsincell,where𝐷𝑁isthemaximum
digitalnumber(DN)valueofpixelinlocationi
over7monthlycompositesin2019.
NDVI
(NIR‐RED)/
(NIR+RED)
Sentinel‐2~42ThesumNDVIvalueofallpixelsinagridcell
SAVI
(NIR‐RED)/
(NIR+RED+L)*
(1+L)
Sentinel‐2~42ThesumSAVIvalueofallpixelsinagridcell
NDBI
(MIR‐NIR)/
(MIR+NIR)
Sentinel‐2~42ThesumNDBIvalueofallpixelsinagridcell
UI
(SWIR2‐NIR)/
(SWIR2+NIR)
Sentinel‐2~42ThesumUIvalueofallpixelsinagridcell
deforestationHansenGlobal
ForestChange
v1.6(2000‐2018)
1Totalforestcoverinagridcell(2018)
Built‐uparea GUF1Totalbuiltupareainagridcell
Built‐upareaWSF1Totalbuiltupareainagridcell
Topography
(slope)
SRTM1Averageslopepergridcell
SurfacetextureSentinel‐1~70Averagetexturepergridcell
RoadsOSM‐ Totallengthofroadsinagridcell
RoadsjunctionsOSM‐ Numberofjunctionsinagridcell

2.2.5.Step5:IdentifyMappedGridCells
WeadoptedavisualinterpretationmethodtovisuallyassessthecompletenessofOSMbuilding
footprintsinthegridcellsinHaitiandSt.Lucia.WeachievedthisbyoverlayingtheOSMbuilding
footprintdatasetwiththemostrecenthigh‐resolutionbasemapimage(providedbyESRI,updated
asof2019[81]).WeidentifiedgridcellsinHaitiandSt.Luciawhereweassessedthatatleast75%of
thebuildingsthatarevisibleinthesatelliteimagehavebeenmapped(weidentified835gridcellsin
Haitiand179gridcellsinSt.Lucia).BecausethemajorityareaofDominicahasbeenmapped,we
skippedthisstepinthecaseofthiscountry.
2.2.6.Step6:PerformCorrelationAnalysisandPrediction
Weevaluatedthecorrelationbetweentheremotelysensedandthegeospatialmeasures(the
explanatoryvariables)andtheareaofOSMbuildingfootprintinagridcellusingaPearson
CorrelationTest,andperformedanOrdinaryLeastSquares(OLS)regressiontoestimatethepotential
ofthevariables,combined,toexplaintheobservedvariationintheareaofOSMbuildingfootprints
inagridcell.Additionally,weevaluatedthepotentialoftheexplanatoryvariablestopredictthearea
ofOSMbuildingfootprintsinagridcellusingaregressionwithRandomForests.RandomForests
[82]aretree‐basedmodelsthatincludekdecisiontreesandprandomlychosenpredictorsforeach
recursion.Whenpredicting,foranexample,itsvariablesarerunthrougheachofthektrees,andthe
kpredictionsareaveragedthroughanarithmeticmean.Eachtreeistrainedusingasubsetof
examplesfromthetrainingset,drawnrandomlywithreplacement,witheachnodeʹsbinaryquestion
determinedusingarandomsubsetofpinputvariables.Weperformedtheregressionwiththe835
gridcellsthatwerevisuallyassessedasbeingrelativelyfullymapped(i.e.,morethan75%ofthe
buildingsinagridcellareassessedasmapped).Toevaluatetheaccuracyoftheprediction,we

RemoteSens.2020,12,1188of27
adoptedafivefoldcross‐validationmethod.Ineachexperiment,theexamplesinoneofthedatafolds
wereleftoutfortestingandtheexamplesintheremainingfourfoldswereusedtotrainthemodel.
Theperformancequalityofthetrainedmodelwastestedontheexamplesintheleft‐outfold,andthe
overallperformancemeasureisthenaveragedoverthefivefolds.Weassessedtheclassification
accuracywithadifferentnumberofdecisiontrees:2,4,8,16,32,64,128,256and512,withminimum
sizeofterminalnodessetto5.
2.2.7.Step7:PredicttheCoverageofOSM‐BuildingFootprintsinEachEntireCountry
Weusedeitherthegridcellsthatarevisuallyassessedasrelativelyfullymapped(inthecaseof
HaitiandSt.Lucia)orallthegridcells(inthecaseofDominica)asreferencesforthetrainingof
RandomForestRegressionandtopredicttheareaofOSMbuildingfootprintsovertheentiregrid
cellsineachcountry.WeidentifiedthegridcellsthatwerepredictedtoincorporateOSMbuilding
footprints,butwerenotyetmapped.
3.Results
Anexaminationofthe136,747cellsspanningHaitishowsthatonly25.1%ofthecellshaveat
leastonemappedbuilding,andonly512ofthe136,747cellshavemorethan10%oftheirareacovered
withbuildingfootprints(Figure2showsahistogramofthedistributionofOSMbuildingfootprints
percell).Onaverage,thereare27.5buildingsinacell(Std=83.4);1530ofthecells(i.e.,only1.1%of
thecellsspanningHaiti)incorporatemorethan100mappedbuildings.Incomparison,8.15%and
6.84%ofthecellsincorporatebuilt‐uplandcoveraccordingtoWSFandGUF,respectively.

Figure2.Thedistribution(histogram)ofOSMbuildingfootprintsarea(squaremeters)pergridcell.
Asdiscussedabove,avisualexaminationofthecompletenessofOSMbuildingfootprintsover
Haitisuggeststhatlargeportionsoftheislandremainunmapped(Figure3a).Figure3b,cshow,as
anillustration,thecoverageofOSMbuildingfootprintsinthecapitalofHaiti,Port‐au‐Princeand
Carrefour,andintheadjacentCarrefourcommune.Whilebuildingsinmanyareaswithinthesecities
havebeenmapped,largeportionsarestillnotfullymapped.Weobservethatdenselymappedzones
ofPort‐au‐Princeco‐existalongsidezonesthatremainentirelyunmapped(Figure3c),avisualpattern
thatmayresultfromtheepisodicengagementofcommunitymappingvolunteersandthedefinition
ofmapping‘tasks’onaneighborhoodscalethroughOSMeditingtools.Moreover,significantparts
innorthernHaitiarenotmapped(Figure4),including,forexample,thecitiesGonaïvesandCap‐
Haitien.
0
2000
4000
6000
8000
10000
12000
200
800
1400
2000
2600
3200
3800
4400
5000
5600
6200
6800
7400
8000
8600
9200
9800
10400
11000
11600
12200
12800
13400
14000
14600
15200
15800
16400
17000
17600
18200
18800
19400
Numberofcells
OSMareapercell(Sqm)

RemoteSens.2020,12,1189of27

Figure3.(a)OSMbuildingfootprintscoverageinHaiti,(b)inthecapitalofHaiti,Port‐au‐Prince,and
(c)intheadjacentCarrefourcommune.YellowindicatesOSMbuildingfootprints.

(a) (b)
Figure4.OSMbuildingfootprintscoveragein(a)thecityofCap‐Haitienand(b)Gonaïvesinnorthern
Haiti.
APearsoncorrelationtestindicatedasignificant(p<0.01)correlationbetweenthetotalareaof
OSMbuildingfootprintsinagridcellandseveraloftheexaminedexplanatoryvariables.As
expected,therewasapositiveandsignificantcorrelationbetweentheareaofOSMbuilding
footprintsinagridcellandthetotalareaofbuilt‐uplandcover,accordingtoWSFandGUF(r=0.73
and0.71,respectively,p<0.01)aswellaswithnighttimelights(VIIRSSOL)(r=0.63,p<0.01).Wefind
asignificant(p<0.01)correlationbetweenOSMbuildingfootprintsareainagridcellwiththefour
Sentinel‐2spectralindices,indicatedbyapositivecorrelationwithUIandNDBI(r=0.59andr=0.47)
andanegativecorrelationwithbothSAVIandNDVI(r=–0.53).
Weidentified835gridcellswhere,accordingtoavisualassessment,atleast75%ofthebuildings
thatwerevisibleinthesatelliteimagearemappedinOSM(Figure5showsexamplesofgridcells
wheremorethan75%ofthestructuresaremapped).ThecorrelationbetweentheareaofOSM
buildingfootprintsinagridcellandtheexaminedpredictorswashighercomparedtotheprevious
experiment,whereallthegridcells(i.e.,136,747gridcells)wereconsidered(forexample,r=0.78and
r=0.65withWSFandVIIRSandr=0.61andr=–0.55withUIandSAVI,respectively)(Table2),which
islikelyduetothefactthatlargeportionsofthecountryarenotmapped(i.e.,therearegridcellsthat
lackOSMcoveragewhileactuallypopulatedandexhibitLULCcharacteristicsofpopulatedareas).
Asexpected,therewerealsosimilaritiesandcorrelationsbetweensomeoftheexplanatoryvariables.

RemoteSens.2020,12,11810of27
Figure6apresentspairwisecorrelationcoefficientsbetweentheexplanatoryvariable(variablesare
orderedaccordingtoahierarchicalclustering).Theexplanatoryvariablesformseveralsimilarity
clusters:aclustercomposedoutofvegetationspectralindices(NDVI,SAVI)andforestcover(which
arepositivelycorrelatedwitheachother),andaclustercomposedoutofbuilt‐uplandcoverspectral
indices(NDBI,UI),togetherwithVIIRS,WSF,GUF,andOSMroadnetworkfeatures.Asexpected,
thereisanegativeandsignificantcorrelationbetweenthevegetationandthebuilt‐uplandcover
spectralindices.Thedendrogramshowninthefigurefurtherhighlightshierarchicalclustersformed
betweenthevariables,notably,OSMareaandVIIRS,UIandNDBI,NDVIandSAVI,androadlength
andnumberofjunctionsinagridcell.

Figure5.Examplesofgridcellsthathavemorethan75%oftheirareamappedwithOSMbuilding
footprints.
Table2.PearsoncorrelationtestbetweentheareaofOSMbuildingfootprintsinagridcellandthe
evaluatedpredictors(thiscorrelationtestincludesonlygridcellsthatwereassessedasmapped,
N=835).
VIIRSGUFWSFNDVINDBISAVI
r0.654*0.76*0.78*–0.551*0.486*–0.551*
UIForestCoverSE1Slope RoadlengthOSMjunctions
r0.614*–0.388*0.16–0.110.69*0.60*
Note:*p<0.01

(a)


RemoteSens.2020,12,11811of27
(b)

(c)



Figure6.Pairwisecorrelationcoefficientsbetweentheexplanatoryvariablesin(a)Haiti,(b)St.Lucia
(calculatedwithinvisuallyassessedgridcells)and(c)Dominica(calculatedwithinallgridcells).
Variablesareorderedaccordingtohierarchicalclustering,whichisalsorepresentedbythe
dendrogram.ThebluelineinthelegendisahistogramofthedistributionofthePearsoncorrelation
coefficients.
AnOrdinaryLeastSquares(OLS)regressionshowsthatnineofthevariablestogetherexplain
upto82%ofthevariationofOSMbuildingfootprintsareainagridcell(R2=0.82,F(12,822)=323.20,
p<0.01)(Table3).Weevaluatedthecontributionoffourtypes(groups)offeaturestothemodelfit
usingastepwiseregressionanalysis:(1)onlyGUFandWSF;(2)withtheadditionofnighttimelights
(VIIRS);(3)withtheadditionoffurtherremotelysensedmeasuresandderivedproducts;(4)withthe
additionofOSMroadnetworkfeatures.Theresultsshowanimprovementofthemodelfitwiththe
additionofeachofthepredictivevariablesgroups(Table4).WhileGUFandWSFtogetherexplain
66%ofthefit,theadditionofnighttimelightsimprovesthefitofthemodel(indicatedbyexplanation
ofupto76%ofthevariation).Theadditionoffurtherremotelysensedmeasures(i.e.,Sentinel‐2‐
derivedspectralindices,slope,textureandforestcover)improvesthemodelfitbyafurther5%(up

RemoteSens.2020,12,11812of27
to81%ofthevariation).WiththeadditionofOSMtransportationnetworkfeaturesthefitofthe
modelimprovesmarginallytoaround82%.
Table3.Modelfitforthestepwiseregressionanalysisofnineevaluatedpredictors.
StepVariableR2AdjustedR2C(p)AICRMSE
1WSF0.6140.613984.918235.213332.1
2UI0.7050.704559.918013.311666.7
3GUF0.7640.763282.917828.110435.7
4VIIRS0.7990.798120.217696.09636.3
5Roadlength0.8140.81349.517631.29264.0
6FCarea0.8200.81923.417605.89118.9
7NDBI0.8220.82117.017599.59078.8
8Numberofjunctions0.8230.82213.117595.69052.3
9Medianslope0.8240.82211.917594.39040.2
Table4.Fourregressionmodeloutputsshowinganimprovementofmodelfitwiththeinclusionof
fourgroupsofvariables:(1)onlyGUFandWSF;(2)additionofnighttimelights(VIIRS);(3)addition
ofremotelysensedmeasuresandderivedproducts;(4)additionofOSMroadnetworkfeatures.
Step(1)(2)(3)(4)
GUF0.115***0.124***0.127***0.138***
(0.010)(0.009)(0.008)(0.008)
WSF0.141***0.081***0.050***0.034***
(0.010)(0.009)(0.009)(0.009)
VIIRS2,214.671***1,276.821**1,073.838***
(124.738)(127.805)(126.619)
NDBI ‐37.152***‐17.790
(11.258)(11.409)
NDVI 72,452.840**64,046.550**
(30,894.100)(29,957.530)
SAVI ‐48,312.220**‐42,708.300**
(20,600.640)(19,976.140)
UI 46.857***25.415**
(9.751)(10.191)
Forestcover  0.060***0.051***
(0.012)(0.011)
Slope 179.453274.377
(192.545)(187.222)
Sentinel‐1  ‐696.229‐295.342
(453.517)(442.270)
Roadlength   1.470***
 (0.543)
No.ofjunctions   60.057**
 (29.311)
Constant0.716‐699.98830,909.080***17,403.480***
(672.122)(573.990)(3,897.596)(4,351.436)

RemoteSens.2020,12,11813of27
Observations835835835835
R20.6630.7560.8130.825
AdjustedR20.6620.7550.8110.823
ResidualStd.Error12,460.3910,615.9409,329.4209,029.806
FStatistic818.245***

856.591***

357.937***

323.203***

Note:  
*p<0.1;**p<0.05;***p<0.01
3.2.PredictionofOSMBuildingFootprintCoverage
TheresultsaboveindicatethattheareaofOSMbuildingfootprintsinagridcellcanbeexplained
byseveraloftheremotelysensedandgeospatialexplanatoryvariables.Toevaluatethepotentialof
thesevariablestopredicttheareaofOSMbuildingfootprintsinacell,weperformedaregression
withRandomForests.
RandomForestregressionpredictsupto89%ofthevariationofOSMbuildingfootprintsina
gridcell.Performanceimproveswiththeadditionofdecisiontreesupto64trees,forexample,from
81%to89%ofthepredictedarea(with2and64decisiontrees,respectively,Figure7).Figure8
presentsacomparisonbetweentheactualandthepredictedareaofOSMbuildingfootprintsinagrid
cell(regressionwith64decisiontrees)(Figure8a).Thetwomostimportantvariablestothemodel
areWSFandGUF,followedbyOSMroadnetworkfeaturesandSentinel‐2derivedspectralindices
(indicatedbyvariableimportancesensitivity(lncNodePurity),Figure8b).

Figure7.TheimprovementofR2withtheincreasefrom2to64decisiontreesintheRandomForest
model.
0.76
0.78
0.8
0.82
0.84
0.86
0.88
0.9
2 4 8 16 32 64 128 256 512
R2
NumberofTrees

RemoteSens.2020,12,11814of27
(a) (b)

Figure8.Comparisonbetweenthe(a)actualandpredictedareaofOSMbuildingcoverageineach
gridcell(using64decisiontrees)and(b)variableimportancesensitivity.
WeusetheRandomForestmodeltopredicttheareaofOSMbuildingfootprintsoverallthegrid
cellsinHaiti(i.e.,wetrainthemodelwiththe835gridcellsthatwereassessedasrelativelyfully
mappedandpredictforthecoverageovertheentiredataset).Figure10bshowsthepredictedareaof
OSMbuildingfootprintspergridcelloverHaiti.TheresultshighlightlargeportionsinHaitithat
havenotyetbeenmapped(e.g.,forexample,thenortherncitiesGonaivesandCap‐Haitien)aswell
aspatchesofunmappedgridcellsaroundmajorcities(e.g.,Port‐au‐Prince).Thisanalysisallowsus
toidentifyareas(gridcells)thatarepredictedtoincorporatelargeareasofbuildingfootprintsbut
arenotmapped(Figure9).
AvisualexaminationshowsthatthepredictedcoverageofOSMbuildingfootprints(Figure10b)
correspondsmorecloselywiththedistributionofbuilt‐uplandcover(accordingtoGUF,for
example)andnighttimelights(VIIRS)(Figure10c,d,respectively)thancomparedtothecurrent
distributionofOSMbuildingfootprints(Figure10a).
Tofurtherevaluatetheaccuracyofthemodel,weperformanOLSregressionanalysisusingthe
entireHaitidataset(i.e.,136,747gridcells).Wefindthattheremotelysensedmeasurementsexplain
upto89%ofthevariationofthepredictedareaofOSMbuildingfootprintsinallthegridcells
spanningHaiti(R2=0.89,F(12,136,734)=90,690,p<0.01).Incomparison,theseindicatorsexplain
only48%ofthevariationofthecurrentareaofOSMbuildingfootprintsintheHaitidataset(R2=0.48,
F(12,136,734)=10,330,p<0.01).
Finally,inordertoidentifygridcellsthatarepredictedtoincorporatebuildingfootprintsbut
areactuallynotmapped,wecalculatedtheratiobetweentheactualandthepredictedareaofOSM
buildingfootprintinagridcell(calculatedasthepredictedareaofOSMbuildingfootprintsinagrid
celldividedbytheactualareaofOSMbuildingfootprintsinagridcell)(Figure11a,c).Thisanalysis
allowsustoidentifygridcellswheretheratiobetweenthepredictedandtheactualareaofOSM
buildingfootprintinagridcellislow(Figure11b,d),highlightinggridcellsthatrequiremapping.

RemoteSens.2020,12,11815of27

(a)
(b)

Figure9.Acomparisonbetween(a)existingOSMbuildingfootprintsand(b)predictedareaofOSM
buildingfootprintsinagridcell(Carrefour,Haiti).Areasshownindarkpurplearepredictedto
includealargeareaofbuildingfootprints.

RemoteSens.2020,12,11816of27


Figure10.(a)TheareaofOSMbuildingfootprintsinagridcellcomparedto(b)theareapredicted
byRandomForestRegression.Thepatternsobservedinthepredictedmapalignwiththeobserved
distributionofthebuilt‐uplandcoveraccordingto(c)GUFandto(d)theintensityofnighttimelights
accordingtoVIIRS.

Figure11.(a,c)TheratiobetweentheactualandthepredictedareaofOSMbuildingfootprintina
gridcell(calculatedasthepredictedareaofOSMbuildingfootprintsinagridcelldividedbythe
actualareaofOSMbuildingfootprintsinagridcell);Gridcellswiththelowestratio(lowerthan20).
(b,d)GridcellsshowninpurplearegridcellswheretheactualareaofOSMbuildingfootprintsis
muchlowerthanthepredictedarea.

RemoteSens.2020,12,11817of27
3.3.EvaluationoftheMethodintheCaseofDominicaandSt.Lucia
Theresultsabovesuggestthepotentialofseveralremotelysensedindicatorstopredictthe
coverageofOSMbuildingfootprints,atleastinthecaseofHaiti.Inordertoassessthevalidityofthe
method,weperformedfurtheranalysisinthecaseoftwoadditionalsmallislandstates:Dominica
andSt.Lucia.AvisualexaminationsuggeststhatthecoverageofOSMbuildingfootprintsin
Dominicaisrelativelycomplete,whilelargeportionsofSt.Luciaremainunmapped.Areaslacking
OSMbuildingfootprintsincludepartsofthecapital,Castries,andthesecond‐largesttown,Vieux
Fort(Figure12).Thesetwoareasaccountforapproximately49%ofthepopulationofSt.Lucia(64,654
and16,624people,respectively[83]).


(a)VieuxFort(b)Castries
Figure12.ExamplesofareasmissingOSMbuildingfootprintsinSt.Lucia((a)VieuxFortand(b)the
Castries).
SimilartothemethodologydescribedinthecaseofHaiti,wecreatedafishnetofgridcells,0.25
km2insize,spanningthetwoislands.InthecaseofDominica,wefoundahighpositiveand
significantcorrelationbetweentheareaofOSMbuildingfootprintsinagridcellandseveral
explanatoryvariables.WefoundahighpositivecorrelationwithGUF,numberofjunctionsandWSF
(betweenr=0.90andr=0.91,p<0.01forboth).ThecorrelationbetweentheareaofOSMbuilding
footprintsandVIIRSisabitlower(r=0.75,p<0.01).WithSentinel‐2‐derivedspectralindices,this
correlationrangesbetweenr=0.35andr=0.38(withUIandNDBI,respectively,p<0.01forboth).An
OLSregressionanalysisrevealsthattogether,thesevariablesexplain92%ofthevariationofOSM
buildingfootprintareainagridcell(R2=0.92,F(12,3846)=3848,p<0.01).RandomForestregression
(with64decisiontrees)resultsinsimilartrends,indicatedbyahighaccuracyrateof88%(regression
accuracyassessedusingfivefoldcross‐validation).
InthecaseofSt.Lucia,whentheanalysiswasdonewiththeentiredatasetofgridcellsoverthe
country(i.e.,N=2781),thecorrelationbetweentheareaofOSMbuildingfootprints,GUFandWSF
rangesbetween0.70and0.75(p<0.01)andthereisalowercorrelationbetweenOSMbuilding
footprintsareaandVIIRS(r=0.58,p<0.01).Together,thepredictorsexplainonly66%ofthevariation
isOSMbuildingfootprints(R2=0.66,F(12,2783)=464.6,p<0.01).Werelatethelowerfitofthemodel
tothefactthatlargeportionsofthecountryhavenotbeenmapped.Thus,wevisuallyassessedthe
completenessofOSMbuildingfootprintsinSt.Luciagridcellsandidentified179gridcellsinwhich
wecouldassessthatmorethan75%oftheirareaismapped.Withthesevisuallyassessedgridcells,
thefitofthemodelimproved,andtogether,thepredictorsexplain92%ofthevariationofOSM
buildingfootprintarea(R2=92%,F(12,166)=166.4,p<0.01).RandomForestregression(with64
decisiontrees)resultsinasimilaraccuracyrate(R2=92%)(Table5).
Finally,weuseRandomForesttopredicttheareaofOSMbuildingfootprintsineachofthe
countries.Figure13presentsthepredictedareaofOSMbuildingfootprintinSt.Lucia.Unmapped
areasinclude,forexample,areassurroundingthecapitalofCastries.Thecentralpartofthecityis

RemoteSens.2020,12,11818of27
denselymappedwhileitsadjacentneighborhoodslackcoverage.Largeswathsofthesurrounding
areasofCharlotte,Vigie,andBiseecompletelylackOSMbuildingfootprints.
Table5.PearsonCorrelationTest,OLSRegression,andRandomForestRegressionforDominicaand
St.Lucia.
DominicaSt.Lucia

Fulldataset
(N=3861)
Full
(N=2781)
Visuallyassessedcells*
(N=179)
(I)PearsonCorrelationTest
GUFr=0.91* r=0.75* r=0.89*
NumofJuncr=0.90*r=0.76*r=0.88
WSFr=0.90*r=0.70*r=0.78
RoadLengthr=0.81* r=0.68*r=0.84*
VIIRSr=0.75*r=0.58*r=0.72*
NDBIr=0.38* r=0.30r=0.3
UIr=0.35* r=0.26r=0.35
Sentinel1r=0.03*r=0.00r=0.20
NDVIr=–0.20* r=–0.19 r=–0.29
SAVIr=–0.20*r=–0.19r=–0.29
Sloper=–0.16r=–0.14*r=0.10*
ForestCoverr=–0.07r=–0.35r=–0.43
(II)OLS
R2=92%
F(12,3846)=3848,p=0.00
R2=66%
F(12,2783)=464.6,p=0.00
R2=92%
F(12,166)=166.4,p=0.00
(III)RandomForest
R2=88%R2=94%
Note:*p<0.01



RemoteSens.2020,12,11819of27

Figure13.(a)ExistingOSMbuildingfootprintsversus(b)predictedareaofOSMbuildingfootprints
inagridcell.Areasshownindarkpurplearepredictedtoincludealargeareaofbuildingfootprints.
4.Discussion
Inrecentdecades,naturaldisastershavebeenresponsibleforanestimated0.1%ofglobaldeaths,
killingonaverage60,000peopleperyear[84].Inthelasttwodecadesalone,developingcountries
haveaccountedformorethanhalfofallreportedcasualties[85].Naturaldisastersoftencause
significantdamagetocommunities,infrastructureandtheenvironment,andrequireimmediate
interventionandimplementationofappropriatemeasuresaimingtosavelives.
Accurateandeasilyaccessiblegeospatialinformationiskeyforaneffectivedisasterrisk
managementcycleandforinformeddecision‐makingduringhumanitarianresponse[86].The
increasingavailabilityofgeospatialinformationisrevolutionizingdisasterresearchandemergency
management.Untilrecently,muchofthisessentialgeospatialinformationwasproprietary,scarce
andinmanycases,unavailableduringsignificantdisasters.
Inthelastdecade,OSMhasrepeatedlybeenshowntobehighlysuitablefordisastermapping
andmanagement.DespitethecontinuouseffortstoimprovecompletenessoftheOSMdatabase,large
portionsoftheworldremainunmapped,especiallyincountriesthatarepronetonaturalhazards,
reflectinglimitedinternetspeedconnectivity,limitedavailabilityofGPSdevices,lackof
technologicallyskilledvolunteersandlimitedawarenessofVGItechnologies[18].
SeveralmethodshavebeenproposedtoassessthecompletenessoftheOSMdatabase,including
evaluationofthecompletenessofstreetnetworks,landuseandbuildings;manyofthemrelyon
externaldatasetsforaccuracyandcompletenessestimation.Theincreasedavailabilityoffreeand
open‐sourceremotelysenseddatacanbeutilizedtoidentifymappinggapsinOSMdatasets,find
locationsthatrequiremapping,andhelpprioritizeandplanmappingcampaignsandefforts.

RemoteSens.2020,12,11820of27
AlthoughseveralapplicationshaverecentlybeenproposedtopredictthecoverageofOSMbymeans
ofremotelysensedderivedproducts(suchasWorldPop)[87],tothebestofourknowledge,nostudy
hasyetevaluatedthepotentialuseofremotelysensedmeasurementstopredictthecompletenessof
OSMfeatures.
Inthisstudy,wedemonstrateamethodologytoidentifyareaswherebuildingfootprintshave
notyetbeenmappedinOSMdataset.Themethodologyreliesonremotelysensedmeasurementsand
derivedproductsandgeospatialinformationrelatedtotheroadnetworktopredictthecompleteness
ofOSMbuildingfootprintsinthreesmallislandstates(Haiti,DominicaandSt.Lucia).
InthecaseofHaiti,theresultsshowthatlargeportionsofthecountryarestillunmapped,
despitethecontinuedmappingeffortstomaintainafullandup‐to‐datemapwhilealsokeepingpace
withthechangingsocio‐physicalcharacteristicsofthecountryandtoaidresponseandrecoveryin
futuredisasters[88].Wefindthatinthecaseofthethreecountries,thecoverageofOSMbuilding
footprintsissignificantlycorrelatedwithseveralremotelysensedmeasuresandindicators.As
expected,thecoverageispositivelycorrelatedwiththedistributionofbuilt‐uplandcover(indicated
byaPearsoncorrelationcoefficientofbetweenr=0.78andr=0.91inthecaseofHaitiandDominica,
respectively).Tosomeextent,thisisnotsurprising,andpreviousstudieshavealreadyshownthe
potentialofremotelysensedderivedproductstopredictthecoverageofOSMbuildingfootprints
[87].However,alimitingfactorinutilizationofremotelysensedderivedproductsisthattheir
availabilityvariesinspaceandtimeandtheseproductsarenotalwaysupdatedonaregularbasis.
Inthisstudy,wedemonstratethepotentialuseoffreeandopen‐sourceremotelysensedindicators
toestimatethecoverageofOSMbuildingfootprints.Weshowthattheintensityofnighttimelights
luminosity(measuredbyVIIRS)ishighlycorrelatedwiththeareaofOSMbuildingfootprints
(indicatedbyaPearsoncorrelationcoefficientofbetweenr=0.65andr=0.75inthecaseofHaitiand
Dominica,respectively).Thisfindingalignswithpreviousstudiesshowingthattheintensityof
nighttimelightsiscloselycorrelatedwithanthropogenicactivitiesandwithchangesinthe
distributionofbuilt‐uplandcover[89,90],aswellaswiththedistributionofgeographicalfeatures,
suchasthedensityoftheroadnetwork[91].Further,weshowthatremotelysensedspectralindices
signifyingthedistributionofvegetation(NDVI,SAVI)andbuilt‐uplandcover(NDBI,UI)arealso
correlatedwiththedistributionofOSMbuildingfootprints.Whilepreviousstudiesproposetoutilize
vegetationspectralindices(e.g.,NDVI)toestimatethecompletenessofurbangreenspacesmapping
inOSM[92],weshowthattheseindicesarealsosignificantlycorrelatedwiththecoverageofOSM
buildingfootprints.
Becausedifferentsensorsrecorddistinctcharacteristicsoftheland(e.g.,brightness,temperature,
height,density,texture),datafusiontechniquesthatexploitthebestcharacteristicsofeachtypeof
sensorhavebecomeavaluableprocedureinremote‐sensinganalysis,includingforurban
applications[70,93]andinmappingthebuilt‐uplandcover[69].Inthisstudy,weshowthatthe
combinedeffectoftheexplanatoryvariablesexplainsbetween92%and94%ofthevariationinthe
areaofOSMbuildingfootprints,exceedingtheeffectofeachpredictorbyitself.Weshowthatinthe
caseofHaiti,theadditionofnighttimelightstobuilt‐uplandcoverclassificationproducts(GUF,
WSF)improvesthefitofthemodelby10%,andthatwiththeadditionofremotelysensedmeasures,
thefitofthemodelimprovesby5%.Althoughlandcovercharacteristicsareoftenrelatedtonighttime
lightluminosity(forexample,vegetationdensitydecreaseswhileluminosityincreasesfromtherural
areatotheurbancore),fusingnighttimeanddaytimeremotelysensedmeasuresallowsforan
increaseintheseparabilitybetweenurbanandnonurbanland[94].Fusingdaytimeandnighttime
measurementsenablefeaturecomplementationandcompensationforthelimitationofsingledata
sourcesinextractingurbaninformation[95].
WealsoperformaRandomForestRegressiontopredicttheareaofOSMbuildingfootprintsin
acellandfindthattheregressionexplainsbetween88%and94%ofthevariation(inDominicaand
Haiti,respectively).PreviousstudiessuggestthatthenumberofdecisiontreesoftheRandomForest
isgenerallyproportionaltothemodel’saccuracy[96],althoughtheyshowmixedresultsforthe
optimalnumberoftreesinthedecisiontree.Thenumbervariesbetween10[97]and150trees[98].

RemoteSens.2020,12,11821of27
Here,wefindthatwithRandomForests,accuracyimprovesupto64decisiontreesandthen
moderatelydecreasesasthenumberoftreesincreases.
Finally,wedemonstratethepotentialofourapproachtoidentifyingmappinggaps,orareasthat
lackOSMcoverage.Wedothisbytrainingthemodelwiththegridcellsthatarevisuallyassessedas
relativelycompletelymappedandusethetrainedmodeltopredictthecoverageofOSMbuilding
footprintsthroughoutthecountries.WecapturegridcellsthatlackOSMbuildingfootprintscoverage
(i.e.,gridcellswherethepredictedcoverageofOSMbuildingfootprintslargelyexceedstheactual
coverage).Thesegridcellsrepresentlocationswheremappingeffortscouldpotentiallybetargeted.
Tosummarize,VGIcontributions,especiallyindevelopingcountries,tendtobemadeinspurts,
forexampleinresponsetoaspecifictrigger,suchasanaturaldisasterorhumanitariancrisis,rather
thanasaregular,continuousprocess[13].BecauseOSMdatasetsrelyonvolunteers,thecompleteness
andmappingeffortsvaryinspaceandtime.Thereisaneedforasystematictoolthatwouldguide
andprioritizemappingeffortsandmappingcampaigns.Asmoreandmoreremotelysenseddata
becomeavailabletotheresearchcommunity,ourstudyextendstheexistingliteratureby
demonstratinghowtheycouldbeleveragedtoconductnovel,andcritical,large‐scaleassessmentsof
thecompletenessoftheOSMdatabase,especiallyinareasathighrisktonaturaldisasters.
Wenoteafewlimitationstothisstudy.First,wedemonstrateourmethodologyinthecasestudy
ofthreesmallislandstates,which,atleasttosomeextent,arecharacterizedbyrelativelysimilar
geographicalconditionsandcharacteristics.Anextensionofthisstudywouldevaluateour
methodologyinadditionalcountriescharacterizedbydiversegeographicalconditions(topography,
landcover,landuse,etc.).Second,inordertocreatethetrainingexamplesforthemodelandto
evaluatetherelationbetweentheexplanatoryvariablesandtheareaofOSMbuildingfootprints,we
createdarelativelysmalldatasetofexamples(gridcells),whichwerevisuallyassessedusinga
subjectivevisualinterpretationmethodasrelativelyfullymapped.Byitsnature,visualinterpretation
maybesubjecttoidiosyncraticvariationacrossindividualsperformingthemanualclassification.An
extensiontothisstudywouldleveragethecrowdtocreateanextensivedatasetofgridcellsvisually
assessedandmappedandwouldaccountforanagreementbetweentheinterpreters.Third,the
analysispresentedinthisstudywasdoneatasinglepointintime.Byitsnature,theOSMdatabase
iscontinuouslyupdatedandisdynamic,whilesimultaneously,newremotelysensedmeasurements
arebeingcollected.Anextensiontothisstudywouldaccountforthesechangesintimeandevaluate
thecompletenessofOSMbuildingfootprintsonanongoingbasisasnewdatabecomesavailable.
ExtensionstoourapproachmayimprovetheidentificationofareasthatlackcompletedOSM
coveragebyaccountingforadditionalinputs;forexample,socioeconomicvariables(including
WorldPop,Facebook’sHighResolutionSettlementLayer(HRSL)andadditional
physical/geographicalcharacteristicsandspectralindices).Furtherextensionstoourapproachmay
alsoincludetheapplicationoflearningalgorithmsandevaluationwithvarioustuningparametersof
theclassifiersandthefitmodels.
5.Conclusions
Globally,therehasbeenanincreaseinthefrequencyandimpactsofmajornaturaldisaster
events;inthenextcentury,itislikelythatclimatechangewillamplifythenumberandseverityof
suchdisasters.Whileaccurateandtimelygeospatialinformationisvitalforthefullcycleofdisaster
riskmanagement,thisdataisnotalwaysavailableforthedisastermanagementcommunitywhen
disastersoccurs.AlthoughVGIplatforms,specificallyOpenStreetMap(OSM),showgreatpotential
tosupporthumanitarianmappingtasks,gapsinVGIdataremainsamajorconcern[99].Thereisan
increasingneedforafullyautomatictoolthatwouldallowtoidentifyareasthatlackacomplete
mappingofOSMfeatures—especiallyinareaspronetohazardevents.Whilepreviousstudieshave
utilizedOSMdataasreferenceforclassificationofbuilt‐uplandcoverwithsatelliteimagery[100–
103]hereweshowthepotentialuseofpubliclyavailable,remotelysenseddataaspredictorsofthe
spatialcoverageofOSMbuildingfootprints.Thetoolandmethodologywepresentherearetime‐
efficientandscalable.

RemoteSens.2020,12,11822of27
AnextensiontoourapproachmayimprovetheaccuracyofthepredictionofOSMbuilding
footprintsareabyaddingadditionalremotelysensedmeasures.Incorporatingadditionaldatasets,
suchasnewlydevelopedVIIRSnighttimelightproducts,