EcologicalApplications ,24(1),2014,pp.84Â…93 2014bytheEcologicalSocietyofAmericaLandscape-scalevariationinplantcommunitycomposition ofanAfricansavannafromairbornespeciesmappingC.A.BALDECK,1,3M.S.COLGAN,1J.-B.FE Â´ RET,1S.R.LEVICK,2R.E.MARTIN,1ANDG.P.ASNER11DepartmentofGlobalEcology,CarnegieInstitutionforScience,260PanamaStreet,Stanford,California94305USA2DepartmentofBiogeochemicalProcesses,MaxPlanckInstituteforBiogeochemistry,10Hans-Kno Â¨ llStrasse,07745Jena,GermanyAbstract .Informationonlandscape-scalepatternsinspeciesdistributionsandcommunitytypesisvitalforecologicalscienceandeffectiveconservationassessmentandplanning. However,detailedmapsofplantcommunitystructureatlandscapescalesseldomexistdueto theinabilityofÂ“eld-basedinventoriestomapasufÂ“cientnumberofindividualsoverlarge areas.TheCarnegieAirborneObservatory(CAO)collectedhyperspectralandlidardataover KrugerNationalPark,SouthAfrica,andthesedatawereusedtoremotelyidentify . 500000 treeandshrubcrownsovera144-km2landscapeusingstackedsupportvectormachines.Maps ofcommunitycompositionalvariationwereproducedbyordinationandclustering,andthe importanceofhillslope-scaletopo-edaphicvariationinshapingcommunitystructurewas evaluatedwithredundancyanalysis.ThisremotespeciesidentiÂ“cationapproachrevealed spatiallycomplexpatternsinwoodyplantcommunitiesthroughoutthelandscapethatcould notbedirectlyobservedusingÂ“eld-basedmethodsalone.Weestimatedthattopo-edaphic variablesrepresentingcatenalsequencesexplained21 % ofspeciescompositionalvariation, whilewealsouncoveredimportantcommunitypatternsthatwereunrelatedtocatenas, indicatingalargeroleforothersoil-relatedfactorsinshapingthesavannacommunity.Our resultsdemonstratetheabilityofairbornespeciesidentiÂ“cationtechniquestomapbiodiversity fortheevaluationofecologicalcontrolsoncommunitycompositionoverlargelandscapes.Keywords:biodiversity;CarnegieAirborneObservatory;catena;community;imagingspectroscopy; KrugerNationalPark;lidar;remotesensing;savanna;SouthAfrica;supportvectormachine .INTRODUCTIONSavannasarecomplexandheterogeneousecosystems covering40 % ofthelandsurfaceofAfrica(Scholesand Walker1993).Naturalsavannasarecomposedof spatiallypatchywoodyandherbaceousplantcommunitiesthatsupportadiverseassemblageofanimal species.Africansavannasarealsoheavilyimpactedby humanlanduseandmanagementpracticesdueto pressureontheseareasforuseinagricultureand livestockgrazing(Higginsetal.1999).Becausethe futureoftheseecosystemsreliesheavilyonmanagement decision-making,itiscriticaltounderstandspatial variationinvegetationcommunitycompositionand thefactorsthatgovernthisvariationacrossarangeof spatialscales. Thecommunitystructureofnaturalsavannavegetationishighlycomplex,andthedrivingenvironmental factorsvaryaccordingtospatialscale(Coughenourand Ellis1993).Atbroadregionalscales,thenatural distributionandmaintenanceofsavannasisdetermined bymeanannualprecipitation(Sankaranetal.2005).At landscapetoregionalscales,differencesinÂ“reregime andherbivorystronglyinÂ”uencecanopycover,vegetationheight,andabovegroundbiomass(Asneretal. 2009,Levicketal.2009).Withinalandscape,rainfall interactswithslightvariationsintopographyto inÂ”uencethesoildevelopmentbycarryingclayparticles downslope,leavinghillcrestswithsandy,nutrient-poor, andwell-drainedsoilsandlow-lyingareaswithmoist clayeysoils(ScholesandWalker1993,Khomoetal. 2011).Landscape-scalevariabilityinvegetationstructureandcompositionishighlyinÂ”uencedbythese catenalformationsandtheirassociatedsoiland hydrologicalcharacteristics(Milne1936,1947,Morison etal.1948,Fraseretal.1987,ScholesandWalker1993, Levicketal.2010,Colganetal.2012 a ). Althoughmuchisknownaboutthefactorsgoverning savannacommunitystructureatdifferentspatialscales, ourunderstandingislimitedbythefactthatÂ“eld-based vegetationsurveysaloneareoftenunabletoresolve spatialpatternsinplantspeciesturnoveratlandscape scales.Remotesensingoffersawaytoobtaindatawith continuousspatialcoverageoverlargegeographicareas, andthishasgreatpotentialforuseinmapping biodiversity(KerrandOstrovsky2003,Turneretal. 2003,Gillespieetal.2008).Mostworkonremote sensingofbiodiversityreliesonindirectmapping methods,wheresomeaspectofbiodiversityderived fromÂ“eldinventoriesisrelatedtostructuralorspectral variables,andthisrelationshipisextrapolatedoverthe studyarea(Nagendra2001).ThisapproachhasbeenManuscriptreceived14February2013;revised19June2013; accepted20June2013.CorrespondingEditor:D.S.Schimel.3E-mail:email@example.com 84
appliedinmappingdistributionsofindividualspecies, variationinalphadiversity(usuallyexpressedasspecies richness),andbetadiversity(theturnoverinspecies compositionamongsites)overlargeareasusingsatellite data(30Â…1000mresolution;e.g.,Tuomistoetal.2003, FoodyandCutler2006,Saatchietal.2008,Feilhauer andSchmidtlein2009,Rocchinietal.2009).However, theindirectmappingapproachcanonlymaptheportion ofvariationintheresponsevariable(s)thatisrelatedto theremotelymappedsurrogatevariables. Airborneremotesensingcanprovidehigh-resolution dataoncanopythree-dimensionalstructureandtheir detailedspectralsignatures,whichsupportsthedirect mappingofindividualplantspecies(Asner, inpress ). Thetechnologyandanalyticalmethodstoidentify individualcrownstospeciesisrapidlyadvancing(e.g., Clarketal.2005,Fe Â´ retandAsner2012).InAfrican savannas,therehasbeenmuchworkfocusingon identifyingindividualtreeandshrubcrownstospecies withavarietyofclassiÂ“cationtechniques(Choetal. 2010,2012,Naidooetal.2012).Arecentadvancein mappingsavannaspeciesusesstackedsupportvector machines(SVM)andintegratespixel-scalespectraldata withcrown-levelstructuraldata,toachieveclassiÂ“cation accuraciesofapproaching80 % amongcommonspecies (Colganetal.2012 b ). In2008,theCarnegieAirborneObservatorycollected high-resolutiondataoncanopythree-dimensionalstructureandhyperspectralsignaturesoverlandscapesin KrugerNationalPark(KNP),SouthAfrica.Weapplied anupdatedversionofthecrownclassiÂ“cationmodelof Colganetal.(2012 b )toindividualtreesthroughouta 144-km2landscapeinKNP.Wethenusedthespecies mapstoexaminebetadiv ersityinthissavanna landscapeandtoevaluatetherelationshipbetweenthe mappedplantcommunityandhillslope-scaletopoedaphicvariation.Weask:(1)Howimportantare hillslope-scaletopo-edaphiccontrols(representingcatenaprocesses)overplantcommunitycomposition throughoutthelandscape?(2)Whataspectsofcommunitycompositionalvariationareunexplainedbycatenas? METHODSStudyarea Thestudylandscapeislocatedinthesouthwestern portionofKrugerNationalPark,SouthAfrica,along theNwaswitshakaRiver(Fig.1).Thearealieson granitesubstrate,withtopographycharacterizedby undulatinghillswithdrainagetotheNwaswitshaka.The hillslopesformcatenasequences,withsandy,welldrained,andnutrient-poorsoilsonhillslopecrestsand clayey,water-loggedsoilsinlow-lyingareas(Gertenbach1983).Themeanannualtemperatureis22 8 Cand themeanannualprecipitationis550mm/yr.Woody vegetationisamixofshrubsandtrees,withmanyofthe commonspeciesbelongingtothegenera Combretum and Acacia . Airbornedata TheCarnegieAirborneObservatory(CAO)Alpha system(Asneretal.2007)wasoperatedoverseveral landscapeswithinKNP,includingtheNwaswitshaka landscape,inAprilÂ…May2008.Thissystemcombines threeinstrumentsubsystemsintoasingleairborne package:(1)ahigh-Â“delityimagingspectrometer(HiFIS),(2)alightdetectionandranging(lidar)scanner, and(3)aglobalpositioningsystemÂ…inertialmeasurementunit(GPSÂ…IMU).Forthisproject,theCAO HiFISsubsystemprovidedspectroscopicimagesconsistingof72bandsinthevisible-nearinfraredspectral regionbetween384.8and1054.3nm.Thelidar subsystemwasoperatedindiscrete-returnmode,with uptofourreturnsperlasershot.Laserbeamdivergence wasdesignedtomatchtheÂ“eld-of-viewoftheimaging spectrometerforaccuratealignmentofspectroscopic andlaserreturndata(Asneretal.2007).TheGPSÂ…IMU subsystemprovidedthree-dimensionalpositioningand altitudedatafortheCAO-Alphasystem,allowingfor accurateprojectionofHiFISandlidardataontothe landsurface.TheHiFISsystemisapushbroomimaging arraywith1500cross-trackpixels,andwasÂ”ownatan altitudeof2kmproviding1.12mgroundsampling distance. FIG.1.Studysitelocationandtopographicmapofthe Nwaswitshaka(KrugerNationalPark,KNP,SouthAfrica) landscapewiththemainrive randcontributingstreams depicted. January2014 85 REMOTEANALYSISOFSAVANNACOMMUNITIES
Digitalelevationmodels(DEMs)wereÂ“ttothelaser pointclouddatatoestimatetop-of-canopyandground surfaces.Canopyheightwascalculatedasthedifference betweenthecanopyandgroundDEMs.Radiancedata fromtheimagingspectrometerwereconvertedto surfacereÂ”ectanceusingACORN5BatchLi(Imspec LLC,Palmdale,California,USA)withaMODTRAN look-uptabletocompensateforRayleighscatteringand aerosolopticals.Tocorrectforcross-trackreÂ”ectance gradientsduetodifferencesinviewandillumination angles(orbidirectionalreÂ”ectancedistributionfunction [BRDF]effects),thereÂ”ectancedatawereadjustedusing akernel-basedBRDFmodel(Colganetal.2012 b ). CrownclassiÂ“cationmodel WithintheoverÂ”ightareas,over1000individualtree andshrubcrownswereidentiÂ“edtospeciesandhad theirlocationrecordedusingasurvey-gradehandheld globalpositioningsystem(GPS)withdifferentialcorrectionduringpost-processing(TrimbleGeoXT;Trimble,Sunnyvale,California,USA).Thesecrownswere locatedintheimages,andtheircorrespondingpixels wereextractedtoconstru ctalibraryofspecies reÂ”ectancespectraandstructuralcharacteristics.Prior tomodelconstruction,thespectraldatawereÂ“lteredto containonlypixelswithNDVI 0.5andmeanNIR reÂ”ectance 20 % (deÂ“nedhereasthemeanreÂ”ectance between850and1054nm),andcrownsoccupyingat leastthreesuchpixels,resultinginasetof742crowns. AclassiÂ“cationmodelwasbuilttoidentifycrownsto oneof15speciesclasses: Acacianigrescens / Acacia burkei , Acaciatortilis , Combretumapiculatum , Combretumcollinum , Combretumhereoense , Combretumimberbe , Crotonmegalobotrys , Colophospermummopane , Diospyrosmespiliformis , Eucleadivinorum , Philenoptera violacea , Spirostachysafricana , Salvadoraaustralis , Sclerocaryabirrea / Lanneaschweinfurthii ,and Terminaliasericea (Appendix:TableA1).Anotherclasswas includedinthemodel,referredtohereasÂÂother,ÂÂ consistingofallotherspeciesidentiÂ“edintheÂ“eldto avoidconstrainingtheclassiÂ“cationtooneofthe15 namedspeciesclasses. ThebasicbuildingblockofthecrownclassiÂ“cation modelwasthesupportvectormachine.SVMisa classiÂ“cationtechniqueknownforitsexcellentperformancehandlingsmalltrainingdatasetswithhigh dimensionalitycomparedtootherapproaches(e.g., lineardiscriminantanalysis,maximumlikelihood,and neuralnetworks;Camps-Vallsetal.2004,Mountrakiset al.2011).ThecrownclassiÂ“cationmodelalsousedan automatedcrownsegmentationalgorithmthatcreates polygonsofindividualtreeorshrubcrownsfromthe canopyheightDEM.CrownsegmentationwasperformedintheeCognitionsoftwarepackage(Developer 8.7;DeÂ“niens,Carlsbad,California,USA)usingthe algorithmdescribedinColganetal.(2012 b ). ThefullcrownclassiÂ“cationmodelwasastacked modelcomposedoftwoSVMs,inwhichtheoutputof theÂ“rstpixel-levelSVMservedastheinputforthe second,crown-levelSVM.TheÂ“rstSVMclassiÂ“ed pixelsbasedontheirreÂ”ectancespectrumandoutputted theprobabilityofbelongingtoeachclass.Theseclass probabilitieswereaveragedoverallpixelswithina crown(asdeterminedbythecrownsegmentation algorithm),andthemaximumheightandareaofeach crownwerecalculatedfromthecanopyheightDEM andthecrownpolygons.Theaverageclassprobabilities, maximumheight,andcrownareaservedastheinputfor thesecond,crown-levelSVM.Cross-validationtests performedatthecrownlevelindicatedthatthestacked SVMmodelclassiÂ“edcrownswith ; 76 % accuracy(see Appendix). Analysisofcommunitystructure Wemappedthespeciesidentityof . 500000crowns withintheNwaswitshakastudyregion.Tocreatethe site-by-speciesmatrix,thecrownswerepartitionedby theirgeographiccoordinatesintosquarequadrats0.25 hainsize.CrownsbelongingtotheÂÂotherÂÂclasswere eliminatedfromthecommunityanalyses,foratotalof ; 51000quadratsthatcontainedatleastonecrownand anaverageof ; 5.6crownsperquadrat.Eliminationof theÂÂotherÂÂclasswasnecessarybecausewecannotknow whethertwocrownsofthisclassbelongtothesameor differentspecies,whichiscriticalforthecalculationof betadiversityandtheanalysisofspeciesdistributions. Weperformedbothnonmetricmultidimensional scaling(NMDS)andhierarchicalclusteringofthesiteby-speciesmatrixtomapthecommunitystructureofthe Nwaswitshakalandscape.NMDSrepresentsvariation incommunitystructuremorerealisticallyascontinuous gradientsofspeciesturnover,whileclusteringsitesinto discretecommunitytypesfacilitatesinterpretationofthe mapsthroughexaminationofthespeciescompositionof thecommunitytypes.Forbothanalyses,weusedBrayCurtisdissimilaritytomeasurethedifferenceinspecies compositionamongquadrats.Calculationofacomplete dissimilaritymatrixfor ; 51000sampleunitsispracticallyimpossible;therefore,mappingthecommunity structureofallquadratswasperformedbycombining NMDSandhierarchicalclusteringwiththek-nearest neighbors(knn)method(Ferrieretal.2007).Forthe clusteranalysis,theoptimalcuttingpointforthecluster dendrogramwaschosenbyperformingindicatorspecies analysisatvariouslevelsoftheclusterdendrogram (DufreneandLegendre1997).DetailsoftheNMDS ordinationwithknn,andtheclusteranalysiswith indicatorspeciesanalysisandknnareprovidedinthe Appendix. TheNMDSresultsweredisplayedasacolorimageby assigningthescoresofeachofthethreeaxesto intensitiesofred,green,andblue.Whentranslating theNMDSaxisscorestocolorintensities,thesame scalingwasappliedtoallthreeaxessimultaneously, preservingtheamountofvariationexplainedbyeach axis.Inthisimage,quadr atswithsimilarspecies composition(lowBray-Curtisdissimilarity)areshown insimilarcolors,whiletheabsolutecolorofaquadratisC.A.BALDECKETAL. 86EcologicalApplications Vol.24,No.1
irrelevant.Theresultingred-green-blue(RGB)image displaysthebetadiversityofthelandscape,showing boththeoverallspatialvariationincommunitycompositionwithinthelandscapeandtheturnoverinspecies compositionbetweenanytwosites.Thecommunity typesidentiÂ“edbyhierarchicalclusteranalysiswerealso mapped.ForthesakeofcomparisonwiththeNMDS betadiversitymap,eachcommunitytypewasassigned theaveragecolorofitsquadratsintheNMDSmap(by takingtheaveragealongeachofthered,green,andblue axes).Thecolorsofthecommunitytypesthusgivean indicationoftheircompositionalsimilarity. Analysisofenvironmentalcontrols Theroleoftopographyinshapingcommunity compositionwasinvestigatedwithredundancyanalysis (RDA;Rao1964).RDAisanextensionofmultiple regressionfortheanalysisofmultivariateresponsedata, whichcanquantifytheamountofcommunitycompositionalvariation(variationinthesite-by-speciesmatrix) explainedbytheenvironmentalvariables.RDA,combinedwithanordinationoftheresidualvariationvia PCA,allowsonetoexaminethestructuresinthe communitydatathatareexplainedbytheenvironmentalvariablesetaswellasthosethatarenotexplainedby theenvironmentalvariables(Borcardetal.2011). Eighteentopographicvariablesrepresentinghillslope morphologyandwateravailabilityweregeneratedfrom thegroundDEMoftheNwaswitshakalandscape. Alongwithelevation,ashapeÂ“leoftheriverand streamswasusedtogeneraterelativeelevationabove stream.SixteenothertopographicindicesweregeneratedinSAGAGIS(version2.1)usingtheterrainanalysis, morphometry,andhydrologymodules:slope,aspect, curvature,plancurvature,proÂ“lecurvature,convergenceindex,catchmentarea,catchmentslope,modiÂ“ed catchmentarea,LS-factor,topographicpositionindex, topographicruggednessindex,multi-resolutionindexof valleybottomÂ”atness,multi-resolutionindexofridge topÂ”atness,topographicwetnessindex,andSAGA wetnessindex( availableonline ).4Priortoanalysis, histogramsofeachvariablewereexaminedandoutliers werediscarded,resultinginasamplesizeof49813 quadrats.ToincreasemodelÂ”exibility,threeorthogonal polynomials(Â“rst,second,andthirddegree)were generatedforeachvariable,withtheexceptionof aspect.Thesineandcosineofaspectwereusedasthe twoaspectvariables,foratotalof53topographic variables. Thevariablesweresubjectedtoforwardselectionto produceamoreparsimoniousmodelthatretainsmost oftheexplanatorypowerofthefullmodel(Blanchetet al.2008,Borcardetal.2011).Forwardselectionofthe variableswasperformedusingthedoublestopping criteriaofBlanchetetal.(2008),inwhichnewvariables addedtothemodelhadtoachievea0.05 a levelandthe cumulativeadjusted R2ofthevariablesetcouldnot exceedtheadjusted R2ofthefullmodel(all53 variables).RDAwasthenperformedusingthereduced variableset,followedbyaPCAoftheresidualvariation. Thespatialstructureoftheexplainedandunexplained variationinthecommunitydatawasexaminedby mappingtheconstrainedandunconstrainedaxes. CrownclassiÂ“cationusingSVM,forwardselectionof topographicvariables,RDA,andhandlingofraster dataweredoneusingtheÂÂe1071ÂÂ(Dimitriadouetal. 2011),ÂÂpackforÂÂ(Drayetal.2011),ÂÂveganÂÂ(Oksanen etal.2012),andÂÂrasterÂÂ(HijmansandvanEtten2012) packages,respectively,oftheRprogramminglanguage (RDevelopmentCoreTeam2012). RESULTSThespeciespredictedmostoftenbythecrown classiÂ“cationmodelwere C.apiculatum (20.1 % ), S. birrea / L.schweinfurthii (8.1 % ), C.hereoense (6.6 % ), A. nigrescens / A.burkei (5.3 % ),and E.divinorum (3.9 % ). Thesespeciesareknowntobeprevalentinthe Nwaswitshakalandscape(Gertenbach1983)andwere commoninourÂ“eldinventorydata(Appendix).The mostcommonlypredictedclasswastheÂÂotherÂÂclass (49.7 % ),whichwasalsofoundtobethemostcommon classwhenÂ“eldinventorydataweregroupedaccording tothemodelclasses(Appendix:Fig.A1).Themodel included15speciesclassesselectedforbroadrepresentationthroughoutKNP,includingafewspecies( C. megalobotrys , C.mopane ,and S.australis )thathavenot beenencounteredwithintheNwaswitshakalandscape (Appendix:Fig.A1).Thesespecieswerepredictedwith theleastfrequency,at0.002 % ,0.060 % ,and0.008 % of crowns,respectively.Ap ortionofthelandscape showingindividualcrownsandtheirpredictedspecies identitiesisshowninFig.2.ThisÂ“guredisplaysmanyof thespeciesencounteredinthelandscapeandtheir distributionwithrespecttocatenaposition:theoccurrenceof C.apiculatum , C.collinum ,and S.birrea / L. schweinfurthii onhillcrests; A.nigrescens / A.burkei in thelow-lyingareas;patchesof E.divinorum occurring adjacenttostreams;andthinbandsof T.sericea along thehillcrestmargins,atmid-elevationsabovestream. ThebetadiversitymapoftheNwaswitshakalandscape(Fig.3a)revealsalargeamountofcompositional turnovercorrespondingtotheriver,streams,and surroundinghillslopes(Fig.1).The11groupschosen torepresentthecommunitytypeshadaone-to-one correspondencewiththeirindicatorspecies:eachgroup hadonesigniÂ“cantindicatorspeciesandeachspecies wasonlysigniÂ“cantforonegroup(Table1).For convenience,werefertoeachcommunitytypebythe nameofitsindicatorspecies.Thevastmajority(91 % )of quadratsbelongedtooneoftheÂ“vemostcommon communitytypes(Table1).Themapdisplayingthese Â“vecommunitytypes(Fig.3b)appearssimilartothe betadiversitymapproducedbyordination(Fig.3a), indicatingthattheseÂ“vegroupscapturemuchofthe communitycompositionalvariationofthelandscape. Theindicatorspeciesforeachofthesecommunitytypes4http://www.s aga-gis.org January2014 87 REMOTEANALYSISOFSAVANNACOMMUNITIES
( C.apiculatum , S.birrea / L.schweinfurthii , C.hereoense , A.nigrescens / A.burkei ,and E.divinorum )arealsothe mostcommonspeciesclassesovertheentirelandscape (notincludingÂÂotherÂÂ;Fig.3c). Forwardselectionreducedtheexpandedsetof53 topographicvariablesdownto16variables(Table2), andthisreducedvariablesetcontainedversionsofnine oftheoriginalvariables.Together,thesevariables explainedatotalof21.2 % ofthevariationincommunity compositionamongquadrats.Therelativeelevation abovestreamexplainedthegreatestamountofvariation (14.6 % )ofanysinglevariable(Table2).Nearlyallof theexplainedvariationwascontainedinonecanonical (constrained)axis,withtheÂ“rstcanonicalaxisexplaining19.5 % ofthecommunitycompositionalvariation andallothercanonicalaxesexplaininglessthan1 % .The unexplainedvariationwasmoreevenlydistributed amongtheunconstrainedaxes,withtheÂ“rstÂ“ve unconstrainedaxesrepresenting41.7 % ,9.9 % ,8.5 % , 7.3 % ,and5.6 % ofthecommunitycompositional variation,andallotherunconstrainedaxesrepresenting , 2 % .MapsofthequadratscoresfortheÂ“rstcanonical andtheÂ“rstÂ“veunconstrainedaxesaredisplayedinFig. 4.TheÂ“rstcanonicalaxiswasweaklycorrelatedwiththe FIG.2.Aportionofthemappedlandscapeshowingindividualtreeandshrubcrownscolor-codedbytheirpredictedspecies identity.StreamsareapparentasdensecorridorsofcrownsbelongingtotheÂÂotherÂÂclass.Thetwoclose-upshighlightthe distributionalpatternsof Terminaliasericea and Eucleadivinorum . C.A.BALDECKETAL. 88EcologicalApplications Vol.24,No.1
abundancesofmanyspecies,whiletheunconstrained axeseachcorrespondedmorestronglytothedistributionofaparticularspecies(Table3).TheÂ“rst unconstrainedaxiswasstronglyrelatedtothedistributionof C.apiculatum ,thesecondwith E.divinorum ,the thirdwith C.hereoense ,thefourthwith S.birrea / L. schweinfurthii ,andtheÂ“fthto A.nigrescens / A.burkei . DISCUSSIONOurresultsdemonstratetheabilityofairbornespecies identiÂ“cationtechniquestomapbiodiversityforthe evaluationofecologicaldriversofcommunitycompositionoverlargelandscapes.Thepatternsamongmore thanhalfamillionmappedcrownsrevealspatially complexvariationinwoodyspeciescommunities(Fig. FIG.3.(a)Resultsofthethree-dimensionalNMDSordinationofcommunitycomposition,displayedasanRGBimage;(b) mapoftheÂ“vemostcommoncommunitytypes,withcolorscorrespondingtotheaveragevaluesfromtheNMDSmap;(c) proportionalabundanceofthesixmostabundantspeciesclassesÂ„theÂÂotherÂÂclass, C.apiculatum ( C.api ), S.birrea / L. Schweinfurthii ( S.bir / L.sch ), C.hereoense ( C.her ), Acacia ,and E.divinorum ( E.div )Â„fortheentirecommunity(gray)andtheÂ“ve mostcommoncommunitytypesfromTable1(coloredbars). January2014 89 REMOTEANALYSISOFSAVANNACOMMUNITIES
3),whichwouldnotbevisiblethroughground-based samplingmethodsalone.Astrongrelationshipbetween communitycompositionandcatenalpatternswas found;21 % ofvariationexplainedbyanenvironmental variablesetisquitelargeforthistypeofanalysis (Baldecketal.2013).Thisresultisconsistentwiththe viewofcatenasequencesasaprimarydeterminantof savannavegetationpatternswithinalandscape(Milne 1947,Morisonetal.1948,Fraseretal.1987).The ecologicalimportanceofcatenasinshapingthese communitiesisapparentinthemapofcommunity compositionalvariation(Fig.3a),andtheshapeofthis explainedvariationisdisplayedinthemapoftheÂ“rst canonicalaxisfromtheRDA(Fig.4a).Themost importanttopographicvariableinexplainingchangesin communitycomposition(elevationabovethenearest stream)wasalsofoundtobeofprimaryimportanceina studyofvegetationpatternsinripariancorridorsof KNP(vanColleretal.2000). However,agreatdeal(79 % )ofthecommunity compositionalvariationwasunexplainedbythetopoedaphicvariables,anditisusefultoexplorehowthis variationmayhavearisen.Thisportionmaybethe productofavarietyofprocesses,includingspecies responsestoenvironmentaldriversthatareunmeasured inthisstudy,spatialautocorrelationofthecommunity compositionduetolimiteddispersal,randomstochasticityinspeciesdistributions(Legendreetal.2005, Andersonetal.2011,Drayetal.2012),lackofmodelto-dataÂ“tinconstrainedordinationmethods(Ã˜kland 1999),anderrorinspeciesidentitiesfromthecrown classiÂ“cationmodel.Thepresenceofspatialstructurein communitydataindicatesanecologicalorigin,speciÂ“callyenvironmentalcontrolordispersalprocesses(Dray etal.2012).Themapsoftheunconstrainedcommunity axesdisplayahighdegreeofspatialstructure,suggestingthatthereareimportantenvironmentaldrivers shapingthecommunitystructureofthislandscapethat wereunidentiÂ“edbystatictopographicmodelsalone. Thedistributionsofmanyspeciesandcommunity typesarestronglyrelatedtotheunconstrainedaxes.For example,theÂ“rstunconstrainedaxiscorrespondsclosely tothedistributionof C.apiculatum (Table3).The E. divinorum communitiesthatoccuralongtheriversand streams(Fig.3b)correspondtothesecondunconstrainedaxisofcommunitycompositionalvariation (Fig.4c,Table3).Thespatialpatchinessandthe adjacencyofthe E.divinorum communitytypetothe riverandstreamssuggest thatthesecommunities correspondtosodiczones,extremelysalinesoilconditionsthatoccurinpatchesÂ”ankingriversatthissite (Gertenbach1983).The A.nigrescens / A.burkei community(Fig.3b)correspondstotheÂ“fthunconstrained communityaxis(Fig.4f,Table3).Thiscommunity exhibitsadifferentspatialstructurewithamuchlarger patchsize,suggestingthatitmaybeassociatedwith broadvariationinsoilconditionsorgeologicsubstrate. Althoughhillslopepositionpartiallyexplainsthedistributionsofthesespeciesandcommunities,correctly placingthemoneitherthehillcrests( C.apiculatum )or low-lyingareas( E.divinorum and A.nigrescens / A. burkei ),theyexhibitlarge-scalespatialpatchinessthat stronglysuggestsaninteractionwithotherimportant environmentalvariables. ThecrownclassiÂ“cationmodelwasdesignedtobe generalizabletoallofKNPbyincludingcommon speciesfromdifferentareasoftheentireregion.In additiontothehighdegreeofaccuracyattainedfor individualcrowns,thelandscape-levelreliabilityofthis modelisdemonstratedbythefactthattheresults conÂ“rmtheextentofcurrentunderstandingofKruger TABLE1.Summaryofthe11communitytypesidentiÂ“edby hierarchicalclusteranalysiswithindicatorspeciesanalysis. Communitytypeandspecies Indicator valueArea( % ) 1) Combretumapiculatum 60.635.6 2) Sclerocaryabirrea / Lanneaschweinfurthii 43.321.5 3) Acacianigrescens / Acaciaburkei 55.012.3 4) Combretumhereoense 38.712.2 5) Eucleadivinorum 76.09.1 6) Combretumimberbe 40.14.7 7) Terminaliasericea 62.41.8 8) Acaciatortilis 75.61.0 9) Spirostachysafricana 63.10.9 10) Philenopteraviolacea 49.40.6 11) Diospyrosmespiliformis 86.80.1 Notes: Communitytypesarelistedinorderofdecreasing abundancewithinthelandscape,withtheproportional coveragegivenbypercentageofarea.Onlyoneindicator speciesclasswassigniÂ“cantforeachcommunitytype.The indicatorvalueforaspeciesrangesbetween0and100,100 indicatingaperfectindicatorwithperfectfaithfulnessand exclusivitytothatgroup.IndicatorvaluesweresigniÂ“cantat P , 0.001,except C.imberbe ,whichwassigniÂ“cantat P , 0.003. TABLE2.Resultsoftheforwardselectionofthetopo-edaphic variables. OrderVariableCum. R2adj 1Relativeelevation0.146 2SAGAwetnessindex0.169 3Topographicwetnessindex20.176 4Elevation0.182 5Topographicwetnessindex30.187 6SAGAwetnessindex30.191 7ModiÂ“edcatchmentarea0.194 8Relativeelevation20.197 9Planecurvature0.199 10Topographicruggednessindex30.201 11SAGAwetnessindex20.204 12ProÂ“lecurvature0.206 13Sineofaspect0.208 14Elevation30.209 15Relativeelevation30.210 16Topographicwetnessindex0.212 Notes: Variablesarelistedindecreasingorderoftheir additionalcontributiontotheoverallproportionofvariation explained(theadjusted R2).Cum. R2adjgivesthecumulative proportionofvariationexplainedbythemodelthatincludes thatvariableandallpreviousvariables. Theadditionalcontributionofeachofthe16selected variablestothetotalvariationexplainedwassigniÂ“cantat P , 0.001. C.A.BALDECKETAL. 90EcologicalApplications Vol.24,No.1
savannas;speciesmostcommonlypredictedbythe modelmatchwiththosethatareknowntobecommon inthelandscape,andspeciesthatwerealmostnever predictedbythemodelmatchthosethatareknownto beabsentfromthelandscape.Furthermore,the distributionsofthespeciesandcommoncommunity typesidentiÂ“edbytheclusteranalysismatchwhatis knownaboutthelandscape(Figs.2and3b).SpeciÂ“cally,theidentiÂ“cationof C.apiculatum and S.birrea on hillcrests, T.sericea innarrowbandsbetweenhillcrests andvalleybottoms, A.nigrescens / A.burkei inthe lowlands,andpatchesof E.divinorum neartheriver andstreamsareconsistentwithpreviousknowledgeof thedistributionsofthesespeciesandÂ“elddataonthe speciescommonintheuplandsandriparianareas (Gertenbach1983,Fraseretal.1987;Appendix). TheremoteidentiÂ“cationofcrownsbringsup importantstatisticalissuesthatwarrantfurtherexploration.First,onlycrownsthathadsufÂ“cientspectral information(atleastthreepixelswithNDVI 0.5and meanNIRreÂ”ectance 20 % )wereclassiÂ“edtoensure propervegetationdetection.Second,nearlyhalfof classiÂ“edcrownswereclassiÂ“edasÂÂotherÂÂ;thesecrowns donotcontributetoourunderstandingofthespecies compositionalturnoveramongsites,andwerediscarded fromcommunityanalyses.Third,theassignmentof crownstothespeciesclassesoccurswithsomedegreeof error,whichhasbeenestimatedtobe24 % inourmodel. Theseissuesreducethenumberofindividualsand speciesusedintheanalysisandreducedataquality througherrorsinspeciesidentiÂ“cation.Itiscertainthat eachoftheseissuesdecreasesthepowertoresolve compositionalgradients;however,theywillalsoimproveasthetechnologyandstatisticalmethodsadvance. Inparticular,theadditionoftheshort-waveinfrared regiontothespectralmeasurementswouldlikely provideadditionalinformationrelevanttothediscriminationofspecies(Asner1998).Anexplorationofhow FIG.4.(a)MapoftheÂ“rstcanonicalaxis;(bÂ…f)mapsoftheÂ“rstÂ“veunconstrainedaxes.Negativescoresalongtheseaxesare lightgray,andpositivescoresaredarkgray.Thegrayscalehasbeenstretchedslightlytoincreasevisibilityofthepatterns. January2014 91 REMOTEANALYSISOFSAVANNACOMMUNITIES
thesefactorsinÂ”uencetheaccuracyofbetadiversity estimatesandthusthepowertodetectcompositional gradientsisbeyondthereachofthisdataset,but presentsanimportantavenueforfutureresearch. Forsavannas,moststudiesinvestigatingtherelationshipsbetweenvegetationcompositionandenvironmentalvariationhavebeenconductedoverrelativelysmall spatialextents,withincompletecoverageofthearea underinvestigation.Wethereforehavelimitedunderstandingofhowvegetationpatternschangeoverthe gradualclimaticgradients,especiallyprecipitation,that occursacrosssouthernAfrica(WitkowskiandOÂConnor1996).Wehavedemonstratedthatimagingspectroscopyhastheabilitytocreatereliablespeciesand biodiversitymapsoverlargeareas,whichcanfacilitate investigationsofthefactorsgoverningspeciescompositionalpatternsanduncovergapsinourunderstanding. Thisapproachshouldbeextendedtoadditionalsavanna landscapestoexploredifferencesinenvironmental controlsoverregionalenvironmentalgradients.These techniquescanalsobeemployedforlong-termmonitoringoftheselandscapesinthefuture,andto understandhowvegetationcommunitieschangein responsetolandmanagementpracticesandclimate change.ACKNOWLEDGMENTSWethanktheSANParksstafffortheiroutstandinglogistical andscientiÂ“csupport.ThisstudywasfundedbytheAndrew MellonFoundation.TheCarnegieAirborneObservatoryis madepossiblebytheGordonandBettyMooreFoundation, theJohnD.andCatherineT.MacArthurFoundation,the AvatarAllianceFoundation,theW.M.KeckFoundation,the MargaretA.CargillFoundation,theGranthamFoundation fortheProtectionoftheEnvironment,MaryAnneNyburg BakerandG.LeonardBakerJr.,andWilliamR.HearstIII. LITERATURECITEDAnderson,M.J.,etal.2011.Navigatingthemultiplemeanings of b diversity:aroadmapforthepracticingecologist. EcologyLetters14:19Â…28. Asner,G.P.1998.Biophysicalandbiochemicalsourcesof variabilityincanopyreÂ”ectance.RemoteSensingofEnvironment64:234Â…253. Asner,G.P. Inpress .Organismicremotesensingfortropical forestecology.AnnalsoftheMissouriBotanicalGarden. Asner,G.P.,D.E.Knapp,T.Kennedy-Bowdoin,M.O.Jones, R.E.Martin,J.Boardman,andC.B.Field.2007.Carnegie airborneobservatory:in-Â”ightfusionofhyperspectral imagingandwaveformlightdetectionandrangingfor three-dimensionalstudiesofecosystems.JournalofApplied RemoteSensing1(013536):1Â…21. Asner,G.P.,S.R.Levick,T.Kennedy-Bowdoin,D.E.Knapp, R.Emerson,J.Jacobson,M.S.Colgan,andR.E.Martin. 2009.Large-scaleimpactsofherbivoresonthestructural diversityofAfricansavannas.ProceedingsoftheNational AcademyofSciencesUSA106:4947Â…4952. Baldeck,C.A.,etal.2013.Soilresourcesandtopographyshape localtreecommunitystructureintropicalforests.ProceedingsoftheRoyalSocietyB280(1753):20122532. Blanchet,F.G.,P.Legendre,andD.Borcard.2008.Forward selectionofexplanatoryvariables.Ecology89:2623Â…2632. Borcard,D.,F.Gillet,andP.Legendre.2011.Numerical ecologywithR.Springer,NewYork,NewYork,USA. Camps-Valls,G.,L.Go Â´ mez-Chova,J.Calpe-Maravilla,J.D. MartÂš Â´ n-Guerrero,E.Soria-Olivas,L.Alonso-Chorda Â´ ,andJ. Moreno.2004.RobustsupportvectormethodforhyperspectraldataclassiÂ“cationandknowledgediscovery.IEEE TransactionsonGeoscienceandRemoteSensing42:1530Â… 1542. Cho,M.A.,P.Debba,R.Mathieu,L.Naidoo,J.vanAardt, andG.P.Asner.2010.Improvingdiscriminationofsavanna treespeciesthroughamultiple-endmemberspectralangle mapperapproach:canopylevelanalysis.IEEETransactions onGeoscienceandRemoteSensing48:4133Â…4142. Cho,M.A.,etal.2012.Mappingtreespeciescompositionin SouthAfricansavannasusinganintegratedairbornespectral andLiDARsystem.RemoteS ensingofEnvironment 125:214Â…226. Clark,M.L.,D.A.Roberts,andD.B.Clark.2005. Hyperspectraldiscriminationoftropicalrainforesttree speciesatleaftocrownscales.RemoteSensingofEnvironment96:375Â…398. Colgan,M.S.,G.P.Asner,S.R.Levick,R.E.Martin,and O.A.Chadwick.2012 a .Topo-edaphiccontrolsoverwoody plantbiomassinSouthAfricansavannas.Biogeosciences 9:1809Â…1821. Colgan,M.S.,C.A.Baldeck,J.-B.Fe Â´ ret,andG.P.Asner. 2012 b .Mappingsavannatreespeciesatecosystemscales usingsupportvectormachineclassiÂ“cationandBRDF correctiononairbornehyperspectralandLiDARdata. RemoteSensing4:3462Â…3480. Coughenour,M.B.,andJ.E.Ellis.1993.Landscapeand climaticcontrolofwoodyvegetationinadrytropical ecosystem:TurkanaDistrict,Kenya.JournalofBiogeography20:383Â…398. Dimitriadou,E.,K.Hornik,F.Leisch,D.Meyer,andA. Weingessel.2011.e1071:miscellaneousfunctionsofthe departmentofstatistics(e1071).Rpackageversion1.6. http://CRAN.R-project.org/package Â¼ e1071 TABLE3.CorrelationsbetweenspeciesabundancesandtheÂ“rstcanonicalaxis(C1)andtheÂ“rstÂ“veunconstrainedaxes(U1Â…U5). SpeciesC1U1U2U3U4U5 A.nigrescens / A.burkei 0.23 0.14 0.110.06 0.340.86 A.tortilis 0.08 0.060.000.01 0.04 0.08 C.apiculatum 0.550.840.010.01 0.020.01 C.collinum 0.190.270.000.09 0.020.00 C.hereoense 0.070.03 0.320.830.410.07 C.imberbe 0.19 0.06 0.060.03 0.050.01 D.mespiliformis 0.070.000.000.040.02 0.01 E.divinorum 0.21 0.060.910.200.150.10 P.violacea 0.13 0.02 0.030.020.01 0.05 S.africana 0.16 0.010.010.000.02 0.02 S.birrea / L.schweinfurthii 0.310.09 0.08 0.490.760.25 T.sericea 0.060.030.000.010.11 0.03 Note: PositivescoresontheordinationaxescorrespondtodarkgraysinthemapsofFig.4. C.A.BALDECKETAL. 92EcologicalApplications Vol.24,No.1
Dray,S.,P.Legendre,andG.Blanchet.2011.packfor:forward selectionwithpermutation(canocop.46).Rpackageversion 0.0-8/r100.http://R-Forge.R-project.org/projects/sedar/ Dray,S.,etal.2012.Communityecologyintheageof multivariatemultiscalespatialanalysis.EcologicalMonographs82:257Â…275. Dufrene,M.,andP.Legendre.1997.Speciesassemblagesand indicatorspecies:theneedforaÂ”exibleasymmetrical approach.EcologicalMonographs67:345Â…366. Feilhauer,H.,andS.Schmidtlein.2009.Mappingcontinuous Â“eldsofforestalphaandbetadiversity.AppliedVegetation Science12:429Â…439. Fe Â´ ret,J.B.,andG.P.Asner.2012.Treespeciesdiscrimination intropicalforestsusingairborneimagingspectroscopy.IEEE TransactionsonGeoscienceandRemoteSensing99:1Â…12. Ferrier,S.,G.Manion,J.Elith,andK.Richardson.2007. Usinggeneralizeddissimilaritymodellingtoanalyseand predictpatternsofbetadiversityinregionalbiodiversity assessment.DiversityandDistributions13:252Â…264. Foody,G.M.,andM.E.J.Cutler.2006.Mappingthespecies richnessandcompositionoftropicalforestsfromremotely senseddatawithneuralnetworks.EcologicalModelling 195:37Â…42. Fraser,S.W.,T.H.vanRooyen,andE.Verster.1987.SoilplantrelationshipsinthecentralKrugerNationalPark. Koedoe30:19Â…34. Gertenbach,W.P.D.1983.LandscapesoftheKrugerNational Park.Koedoe26:9Â…121. Gillespie,T.W.,G.M.Foody,D.Rocchini,A.P.Giorgi,and S.Saatchi.2008.Measuringandmodelingbiodiversityfrom space.ProgressinPhysicalGeography32:203Â…221. Higgins,S.I.,C.M.Shackleton,andE.R.Robinson.1999. Changesinwoodycommunitystructureandcomposition undercontrastinglandusesystemsinasemi-aridsavanna, SouthAfrica.JournalofBiogeography26:619Â…627. Hijmans,R.J.,andJ.vanEtten.2012.raster:geographic analysisandmodelingwithrasterdata.Rpackageversion 2.0-12.http://CRAN.R-project.org/package Â¼ raster Kerr,J.,andM.Ostrovsky.2003.Fromspacetospecies: ecologicalapplicationsforremotesensing.TrendsinEcology andEvolution18:299Â…305. Khomo,L.,A.S.Hartshorn,K.H.Rogers,andO.A. Chadwick.2011.Impactofrainfallandtopographyonthe distributionofclaysandmajorcationsingraniticcatenasof southernAfrica.Catena87:119Â…128. Legendre,P.,D.Borcard,andP.R.Peres-Neto.2005. Analyzingbetadiversity:partitioningthespatialvariation ofcommunitycompositiondata.EcologicalMonographs 75:435Â…450. Levick,S.R.,G.P.Asner,O.A.Chadwick,L.M.Khomo, K.H.Rogers,A.S.Hartshorn,T.Kennedy-Bowdoin,and D.E.Knapp.2010.Regionalinsightintosavannahydrogeomorphologyfromtermitemounds.NatureCommunications1:1Â…7. Levick,S.R.,G.P.Asner,T.Kennedy-Bowdoin,andD.E. Knapp.2009.TherelativeinÂ”uenceofÂ“reandherbivoryon savannathree-dimensionalvegetationstructure.Biological Conservation142:1693Â…1700. Milne,G.1936.NormalerosionasafactorinsoilproÂ“le development.Nature138:548Â…549. Milne,G.1947.Asoilreconnaissancejourneythroughpartsof TanganyikaTerritoryDecember1935toFebruary1936. JournalofEcology35:192Â…265. Morison,C.G.T.,A.C.Hoyle,andJ.F.Hope-Simpson.1948. Tropicalsoil-vegetationcatenasandmosaics:astudyinthe south-westernpartoftheAnglo-EgyptianSudan.Journalof Ecology36:1Â…84. Mountrakis,G.,J.Im,andC.Ogole.2011.Supportvector machinesinremotesensing:areview.ISPRSJournalof PhotogrammetryandRemoteSensing66:247Â…259. Nagendra,H.2001.Usingremotesensingtoassessbiodiversity. InternationalJournalofRemoteSensing22:2377Â…2400. Naidoo,L.,M.A.Cho,R.Mathieu,andG.Asner.2012. ClassiÂ“cationofsavannatreespecies,intheGreaterKruger NationalParkregion,byintegratinghyperspectraland LiDARdatainarandomforestdataminingenvironment. ISPRSJournalofPhotogrammetryandRemoteSensing 69:167Â…179. Ã˜kland,R.H.1999.Onthevariationexplainedbyordination andconstrainedordinationaxes.JournalofVegetation Science10:131Â…136. Oksanen,J.,F.G.Blanchet,R.Kindt,P.Legendre,P.R. Minchin,R.B.OÂHara,G.L.Simpson,P.Solymos, M.H.H.Stevens,andH.Wagner.2012.vegan:community ecologypackage.Rpackageversion2.0-4.http://CRAN. R-project.org/package Â¼ vegan RDevelopmentCoreTeam.2012.R:alanguageand environmentforstatisticalcomputing.RFoundationfor StatisticalComputing,Vienna,Austria.http://www. R-project.org/ Rao,C.R.1964.Theuseandinterpretationofprincipal componentanalysisinappliedresearch.Sankhyaa Â´ SeriesA 26:329Â…358. Rocchini,D.,H.Nagendra,R.Ghate,andB.S.Cade.2009. Spectraldistancedecay:assessingspeciesbetadiversityby quantileregression.PhotogrammetricEngineeringandRemoteSensing75:1225Â…1230. Saatchi,S.,W.Buermann,H.terSteeg,S.Mori,andT.B. Smith.2008.ModelingdistributionofAmazoniantree speciesanddiversityusingremotesensingmeasurements. RemoteSensingofEnvironment112:2000Â…2017. Sankaran,M.,etal.2005.Determinantsofwoodycoverin Africansavannas.Nature438:846Â…849. Scholes,R.J.,andB.H.Walker.1993.AnAfricansavanna: synthesisoftheNylsvleystudy.CambridgeUniversityPress, Cambridge,UK. Tuomisto,H.,A.D.Poulsen,K.Ruokolainen,R.C.Moran,C. Quintana,J.Celi,andG.Ca Â˜ n as.2003.LinkingÂ”oristic patternswithsoilheterogeneityandsatelliteimageryin EcuadorianAmazonia.EcologicalApplications13:352Â…371. Turner,W.,S.Spector,N.Gardiner,M.Fladeland,E.Sterling, andM.Steininger.2003.Remotesensingforbiodiversity scienceandconservation.TrendsinEcologyandEvolution 18:306Â…314. vanColler,A.L.,K.H.Rogers,andG.L.Heritage.2000. Riparianvegetation-environmentrelationships:complementarityofgradientsversusp atchhierarchyapproaches. JournalofVegetationScience11:337Â…350. Witkowski,E.T.F.,andT.G.OÂConnor.1996.Topo-edaphic, Â”oristicandphysiognomicgradientsofwoodyplantsina semi-aridAfricansavannawoodland.Vegetatio124:9Â…23.SUPPLEMENTALMATERIALAppendix SupplementarymethodsforestimationofclassiÂ“cationmodelaccuracy,communityanalysis,andÂ“elddatacomparison ( EcologicalArchives A024-005-A1 ). January2014 93 REMOTEANALYSISOFSAVANNACOMMUNITIES
CONTRIBUTION OF TROPICAL AGRICULTURAL TREES TO SPECIES DIVERSITY AND CARBON: A NEW LANDSCAPE PERSPECTIVE ENABLED BY HIGH RESOLUTION HYPERSPECTRAL AND LIDAR IMAGES By SARAH GRAVES A THESIS PRESENTED TO THE GRADUATE SCHOOL O F THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE UNIVERSITY OF FLORIDA 2014
Â© 2014 Sarah Graves
To my parents
4 ACKNOWLEDGMENTS I th ank my teac hers and mentors for inspiring me to continue a future in science and conservation ; the input and support of my committee members , Wendell Cropper and Jack Putz , and especially my advisor Stephanie Bohlman; my peers for their assistance and guidance ; and m y parents for their continued support and excitement . I also thank the individuals and organizations who helped complete this research including; the Tinker Travel Fund awarded through the Tropical Conservation and Development Program; Matt Colgan and Greg Asner of the Carnegie Institution for Science; Pablo Ramos and the staff the Smithsonian Tropical Research Institute; Jairo Batista Bernal and other staff at the Azuero Earth Project; and additional field assis tants , Lesly Candelaria, Luis Mancilla, and D iogenes Ibarra .
5 TABLE OF CONTENTS page ACKNOWLED GMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 7 LIST OF FIGURES ................................ ................................ ................................ .......... 8 LIST OF ABBREVIATION S ................................ ................................ ............................. 9 ABSTRACT ................................ ................................ ................................ ................... 11 CHAPTER 1 INTRODUCTION: A NEW LANDSCAPE PERSPECTIVE ................................ ...... 13 Agricultural Production and Ecosystem Services ................................ .................... 13 Agricultural Tree Cover Dynamics ................................ ................................ .......... 15 Aerial Images of Changing Landscape s ................................ ................................ . 17 2 CLASSIFICATION OF INDIVIDUAL TREE SPECIES USING A HYPERSPECTRAL AERIAL IMAGE FOR A TROPICAL AGRICULTURAL LANDSCAPE ................................ ................................ ................................ .......... 19 Int roduction ................................ ................................ ................................ ............. 19 Applications of Species Maps ................................ ................................ ........... 20 Species Maps from Hyperspectral Aerial Data ................................ ................. 21 Methods ................................ ................................ ................................ .................. 25 Study Site ................................ ................................ ................................ ......... 25 Aerial and Field Data Collection ................................ ................................ ....... 26 Classification Algorithm Development ................................ .............................. 27 Classification Algorithm Testing ................................ ................................ ....... 28 Classification Algorithm Ac curacy ................................ ................................ .... 29 Application to Landscape ................................ ................................ ................. 30 Results ................................ ................................ ................................ .................... 31 Number of Speci es ................................ ................................ ........................... 31 Number of Crowns ................................ ................................ ........................... 32 Spectral Range ................................ ................................ ................................ . 33 Final Classification Model and Landscape Species Prediction ......................... 33 Discussion ................................ ................................ ................................ .............. 33 Dimensions of Hyperspectral Aerial Data and Their Importance for Specie s Classification ................................ ................................ ................................ . 34 Training Data Effects on Classification Accuracy ................................ ............. 36 Algorithm Selection ................................ ................................ .......................... 39 Conclusion ................................ ................................ ................................ .............. 40
6 3 COMPARISON OF INDIVIDUAL TREE AND PLOT BASED ABOVEGROUND BIOMASS (AGB) ESTIMATES FROM AERIAL LIDAR DATA FOR A TROPICAL AGRICULTURAL LANDSCAPE ................................ ................................ ............. 50 Introduction ................................ ................................ ................................ ............. 50 Methods ................................ ................................ ................................ .................. 52 Study Site ................................ ................................ ................................ ......... 52 Aerial Data ................................ ................................ ................................ ........ 53 Overview ................................ ................................ ................................ .......... 53 Development of Field Based Allometric AGB Model ................................ ........ 54 Development of lidar AGB model ................................ ................................ ..... 56 Landscape Map of Agricultural and Forest Tree Crowns ................................ . 56 AGB Estimation from Lidar Derived Tree Crowns ................................ ............ 57 Calculation of the Difference Between Tree and Plot Based AGB Estimates ... 58 Results ................................ ................................ ................................ .................... 58 Individual Tree AGB Lidar Estimates ................................ ................................ 58 AGB Estimates of Agricultural and Forest Trees ................................ .............. 59 Differences in Tree Based vs. Plot Based Estimates of Landscape AGB ........ 59 Discussion ................................ ................................ ................................ .............. 60 Contribution of Agricultural Trees to Landscape AGB ................................ ...... 60 AGB Lidar Model ................................ ................................ .............................. 61 Future Directions in Landscape AGB Estimation ................................ .............. 64 Conclusion ................................ ................................ ................................ .............. 65 4 CONCLUSION: ECOLOGICAL VALUE OF TROPICAL AGRICULTURAL LANDSCAPES ................................ ................................ ................................ ........ 71 APPENDIX A NDIV AND NIR SPECTRAL FILTERS ................................ ................................ .... 72 B COMPARISON BETWEEN FIELD AND LIDAR MEASURED TREE HEIGHT ...... 73 C SUMMARY OF TESTED LIDAR AGB MODELS ................................ .................... 74 LIST OF REFERENCES ................................ ................................ ............................... 75 BIOGRAPHICAL SKETCH ................................ ................................ ............................ 84
7 LIST OF TAB LES Table page 2 1 Summary of species classification studies with hyperspectral data. ................... 41 2 2 Sample size and mean NDVI of tree species included in this study. .................. 42 2 3 Overall accuracy of classifications with different spectral ranges. ...................... 43 2 4 Species classification accuracy of the final SVM classification model with 37 species ................................ ................................ ................................ ............... 43 3 1 Summary statistics of land cover, tree cover, and biomass across forested and agricultural areas. ................................ ................................ ........................ 66 3 2 Summary of differences between tree based and plot based AGB estimates . ... 66 C 1 AGB lidar models and their model fit statistics for 1, 100 field measured trees. .. 74
8 LIST OF FIGURES Figure page 2 1 Study site on the Azuero Peninsula.. ................................ ................................ .. 44 2 2 The overall accuracy of multiple SVM classifications with changes in t he number of classified species . ................................ ................................ .............. 45 2 3 Error and prediction bias for a classification with 44 species. ............................. 46 2 4 Overall accuracy and prediction bias for classifications with the full set of training da ta to only 5 crowns per species ................................ .......................... 47 2 5 Differences in classification accuracies for 44 species with the full spectral range to classifications with reduced spectral ranges ................................ ......... 48 2 6 Predicted abundance for 37 species across 22,58 7 ha study site ...................... 49 3 1 Study site on the Azuero Peninsula. ................................ ................................ ... 66 3 2 Outline of methods for landscape AGB analysis. ................................ ................ 67 3 3 Map of forest and agricultural tree classification for a subset of the study area . ................................ ................................ ................................ ................... 68 3 4 Comparison of final AGB lidar predictions and AGB ch ave predictions for all 1,100 field measured trees ................................ ................................ ........................... 69 3 5 Tree and plot based AGB density estimates ................................ ...................... 70 A 1 Overall accuracy for va riable NDVI and NIR thresholds. ................................ .... 72 B 1 Comparison of maximum tree crown height measurements ............................... 7 3
9 LIST OF ABBREVIATIONS AGB a boveground biomass ACD a boveground ca rbon density CA tree crown area CAO Carnegie Airborne Observatory cm centimeter CR tree crown radius D tree stem diameter measured at 1.4 m from the ground FAO Food and Agriculture Organization of the United Nations H maximum tree height kg kilogr am lidar Light Detection and Ranging m meter Mg m egagram , a unit used to measure aboveground biomass and carbon NDVI Normalized Difference Vegetation Index NEON National Ecological Observatory Network NIR n ear infrared range of the electromagnetic spectrum , 700 1 400 nm nm n anometer , a unit used to measure the wavelength size of the electromagnetic spectrum Pg p etagram , a unit used to measure global quantities of carbon REDD+ Reducing Emissions from Deforestation and Forest Degradation , a program of the United Nations SVM Support Vector Machine algorithm SWIR shortwave infrared range o f the electromagnetic spectrum, 14 00 2 500 nm
10 TCH t op of c anopy h eight , a metric used in plot based lidar measurements VIS visible range o f the electromagnetic s pectrum, 390 700 nm WD wood density or wood specific gravity
11 Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science CONTRIBUTIO N OF TROPICAL AGRICULTURAL TREES TO SPECIES DIVERSITY AND CARBON: A NEW LANDSCAPE PERSPECTIVE ENABLED BY HIGH RESOLUTION HYPERSPECTRAL AND LIDAR IMAGES By Sarah Graves August 2014 Chair: Stephanie A. Bohlman Major: Forest Resources and Conservation At 38% of global land surface area, a griculture exceeds forest as the dominant global biome . Despite losses of tropical forests, tropical agricultural landscapes contain substantial tree cover . While agricultural trees provide valuable services and products, tree species composition and aboveground biomass have not been quantified across agricultural landscapes due to inadequate methods to characterize dispersed and heterogeneous tree cover . The objective of this research is to characterize a tropical agricult ural landscape on the Azuero Peninsula of Panama in terms of tree species composition and above ground biomass using high resolution hyperspectral and lidar aerial data . First, w e use d hypers pectral reflectance (150 bands, 380 2510 nm) to develop a classifi cation model to identify 37 tropical tree species . The final model had an accuracy of 61. 4 % with species specific accuracy ranging from 8.3 96.7%. Second, we use d lidar data (1.12 m spatial resolution) to estimate above ground biomass of agricultural and fo rest trees . Approximately 54% of total landscape tree cover was found in agricultural areas with an average above ground biomass density of 12.7 Mg/ha, which accounted for more above ground biomass than forested areas. Furthermore, this
12 estimate is 60% great er than published landscape estimates from methods developed for forested areas, with the largest differences occurring in areas with dispersed tree cover . This research demonstrates the utility of novel aerial data to characterize species composition and biomass of a diverse and heterogeneous tropical landscape. As agriculture continues to expand to provide resources for growing global populations , hyperspectral and lidar aerial data can provide information about forest resources of tropical landscapes.
13 CHAPTER 1 INTRODUCTION: A NEW LANDSCAPE PERSPECTIVE Agricultural Production and Ecosystem Services The growing global population is causing unprecedented change in global land use and land cover. Using 2010 global assessments developed from national report s , the Food and Agriculture Organization of the United Nations (FAO) estimate that 38% of global land are a is used for agricultural production, whereas 31% is forested (FAOSTAT 2014). Similar conclusions about agricultural surpassing forest a s the dominant global biome have been reached using analysis of satellite data (Ramankutty et al. 2008 ). Global f orest loss , particularly of tropical forest s , driven by agricultur al expansion has global consequences for the carbon cycle and species diversity . Land use c hange in the tropics, which is dominated by the loss of forests and expansion of agriculture (Gibbs et al. 2010), has caused an average annual emission of 4 Pg of carbon to the atmosphere (Houghton 2003 ). Furthermore, while current and future estimates in global species extinction rates are widely variable depending on region, species taxa, and estimate method, ( Stork 2010) land use change is proposed as the primary driver of species extinction (Sala 2000) . Ecological consequences of these patterns in globa l land use change also cause numerous social consequence s, including the extinction of indigenous communities and loss of tropical forest products (Laurance 1999) . As global populations continue to increase, resulting in greater demand for food and fuel pr oduction, a major global question to address is how to meet these demand s without compromising the functions and diversity of natural ecosystems (Laurance et al. 2014). One proposed solution is to increase the production of current agricultural land to mee t demand and reduce the expansion of agricultural area (Foley et al. 2005) . In this
14 valuable ecosystems) are physically separate, allowing for intense agricultural p roduction, thereby sparing forests because agricultural demand has been met (DeFries and Rosenzweig 2010). In regions of the world that have low current production relative to the land potential, p ractices of intensified agricultural production, such as th e use of irrigation, application of fertilizer, could (Mueller et al. 2012). This strategy was implemented during the 1940s and 1950s with the green revolution, where it was estimated improved agricultural practices increas ed food production and prevented agricultural expansion onto 18 to 27 mi llion hectares of land area (Stevenson et al. 2013). of multi functional agriculture, where areas of agricultural producti on contain elements of natural ecosystems, such as tree cover (Foley et al. 2005). In this strategy , agricultural production may be lower than in an intensified system with high mechanization and high inputs , but the system maintains some diversity and ser vices of the natural ecosystem while still producing agricultural products . Progress in the development of these systems has been made in the field of agroforestry, where crop and trees are grown together ( Nair 2008) . There is not likely one scenario that can be applied throughout the globe and socio economic factors are major determinants of land management outcomes (Carberry et al. 2013). Furthermore, s olutions to th e food production problem will require investment in education, technology, and new ways o f thinking , such as the development of new crop varieties (Tilman et al. 2002). However, the land sparing and
15 land sharing debate is increasing ly informed by quantitative analysis of different scenarios. These studies highlight the ways in which land manag ement scenarios are evaluated , which often include a limited set of descriptors, such as animal or insect specie s richness, and forest biomass. This thesis research adds to the discussion by quantifying the two metrics by which landscapes are evaluated, tr ee species composition and abundance, and aboveground carbon . The unique element of this research is the use of aerial imagery to derive estimates of species diversity and biomass. This approach allows for examination of large areas across a complex landsc ape. Furthermore, it produces spatially explicit results, which can be used with other datasets, such as agricultural production. Agricultural Tree Cover Dynamics Field studies and analysis of satellite imagery indicates that agricultural landscapes are no t void of trees, suggesting that the ecological services trees provide may still be present , albeit at a lower level than the services provided by a tropical forest that likely preceded agriculture . Small scale studies using historic aerial imagery have re vealed that agricultural tree cover change s through time . While these studies provide information on some drivers of changes in tree co ver on agricultural landscapes, this understanding is limited to select sites and small scales , and therefore cannot be u sed to understand global trends (Gibbons et al. 2008, Fischer et al. 2010). At the global scale, a nalysis of satellite data has revealed that agricultural tree cover in the tropics is high . For example, i n Central A merica, 9 4 % of agricultural land has over 10% tree cover . While some of the patterns of tree cover can be explained by climate and population density, other factors of tenure, ma r ket, and national policies are likely
16 important factors in global trends of agricultural tree cover (Zomer et al. 2014 ) . To better understand tree resources in lands outside of forests, t he FAO, which compiles national and global reports of natural resources, particularly land use and forest cover , are calling for an incorporation of trees outside of forests into national and global inventories of resources (de Foresta et al. 2013). This highlights the global need for documentation of trees in agricultural landscapes. Agricultural landscapes, and the tree cover that goes along with it, will continue to change with growth i n human population and socio economic changes ( Munteanu et al. 2014 ) . Satellite estimates show that between 2000 and 2010 , agricultural areas with some form of tree cover have been increasing (Zomer et al. 2014) , though the driver of this trend is not clea r . Despite trends in agricultural tree cover that can be evaluated with historic aerial imagery, we do not know how changes in the rural population and demand for agriculture in the tropics will affect agricultural tree cover. Human populations are growing and urbanizing, with drastic change in the tropics. The demographic transition affects the composition of rural landscapes; however, there is little empirical evidence or mechanistic understanding of how these changes affect agricultural tree cover. Agric ultural intensification includes practices such as increased tillage and irrigation, which likely reduce tree cover (Maron and Fitzsimons 2007). At the same time, agricultural abandonment in occurring caused by population migration from rural to urban area s. Newly abandoned fields and pastures often lead to forest regeneration because tree cover provide conditions suitable for the establishment and growth of new trees (Chazdon 2003, Zahawi and Augspurger 2006), a process that has been well documented in tro pical regions such as Costa Rica (Schlawin and Zahawi
17 2008) and the eastern Amazon (Nepstad et al. 1996). To aid in land use decisions, new data and analysis of landscapes should capture the ecological values of agricultural landscapes to ensure it is bein g considered with these land use changes, and do it in a way that is repeatable, to allow for measurement of changes through time and better understand ing of the consequences of changes in these important landscapes . Aerial Images of Changing Landscapes Ae rial imagery ha s the potential to bridge the gap between what is known on the ground in terms of tree presence, and what is detected from satellites, and allow for repeated data collection through time. Traditionally, information about species diversity an d carbon stocks has been obtained through detailed field data collection. The primarily methodology for species diversity is sampling an area, recording species present and estimating of abundance, either by counting the number of individuals, their size, or their percent cover. While these methods provide detailed and useful information, the time consuming nature of this work in addition to accuracy problems in applying these measurements to larger areas, means uses of these data are limited in addressing landscape or global scale ecological questions. In this way, data and methodology from the field of environmental remote sensing is especially applicable. Aerial photography is a useful tool for assessing tree composition of landscapes because of its high spatial resolution and broad spatial scale. Digital images can be analyzed with computer programs, which remove the timely task of manual interpretation (Morgan et al. 2010). Finally, aerial sensor systems not have the capacity to collect high resolution d ata across a broad spectral range (hyperspectral), and can be equipped with lidar (light detection and ranging) systems to capture surface elevation (Asner et al. 2012). The growing importance of aerial data collection is illustrated in its
18 incorporation t o the National Ecological Observatory Network (NEON), a system of integrated ground, airborne, and satellite data collection for 89 core sites across North America. The motivation for aerial data collection for NEON is to connect ground measurements with t hose made from satellites, and have the ability for annual data collection (Kampe et al. 2010). The availability of aerial data, particularly high spectral resolution data, is growing, with many ecological applications. The objective of this research was t o develop methods to characterize tropical agricultural landscapes in terms of tree species abundance and aboveground biomass using novel hyperspectral and lidar aerial data. This work advances the field of aerial remote sensing and applications to ecology and conservation, and provides valuable data for discussions of what is the current state of tropical agricultural landscapes and how they can be manage for species conservation or carbon sequestration .
19 CHAPTER 2 CLASSIFICATION OF INDIVIDUAL TREE SPECIE S USING A HYPERSPECTRAL AERIAL IMAGE FOR A TROPICAL AGRICULTURAL LANDSCAPE Introduction A t 3 8 % of global surface area, agric ulture exceeds fore st as the dominant global biome ( FAOSTAT 2014 ). Despite continual losses of tropical forests (Hansen et al. 2013 ) , caused largely by the expansion of agriculture (Gibbs et al. 2010), agricultural landscapes in the tropics contain substantial tree cover (Zomer et al. 2009). Based on c. 2010 assessments of global land cover data and satellite imagery, 40% of global ag ricultural area (9 billion km 2 ) is characterized by having over 10% tree cover. In the tropics, agricultural tree cover is particularly high . For example, 94% of agricultural land in Central America (253 thousand km 2 ) has over 10% tree cover, and 53% (270 thousand km 2 ) has over 30% tree cover (Zomer et al. 2009). Agricul tural trees, defined here as individual trees, scattered groups of trees, live fences, and windbreaks (Plieninger 2012) that exist in areas of agricultural production , are likely remnants of previous forest cover that have not been removed , have naturally established , or were deliberately planted in active agricultural areas (Plieninger 2012) . Consequently, most species found in agricultural areas are those that provide services or products r ecognized by landowners (Love and Spaner 2005, Griscom et al. 2011, Harvey et al. 2011). Agricultural trees also provide a diversity of ecological benefits, including wildlife habitat ( Harvey et al. 2006 , Medina et al. 2007 ) and carbon storage ( Kuyah et al . 2012 a ). Additionally, a gricultural landscapes in Cen tral America are documented to have high tree species richness , for example 32 species were documented in 4 ha of active pastures in Panama (Griscom et al. 2011), 17 20 species
20 in 1 ha of pasture in Nic aragua (Harvey et al. 2006), and 190 species across 237 ha in Costa Rica (Harvey and Habe r , 1999) . Applications of Species Maps While field inventories provide information on tree presenc e and species diversity at the local level to address local diversity and management at a particular time ( Chazdon et al. 2009 ) , studies of landscape level species composition can be used to address a variety of important ecological and social issues. One way to characterize landscape species diversity is by directly identi fying the species of individual trees on the landscape. These tree species maps can be used to identify areas of relative high or low diversity across broad areas , which allows for the evaluation of spatial patterns in tropical tree populations ( Condit et al. 2000) , including identification of biophysical drivers of diversity and coexistence (Colgan and Asner 2014) . For conservation activities , species maps indicate areas of high species diversity, which should be protected and maintained. Species maps may also indicate the diversity of the future landscape , because existing trees provide seeds, act at seed dispersal nuclei, and establish a suitable microclimate for the establishment of tree saplings (Chazdon 2003) . In applications not specific to diversity, s pecies maps may also allow for detection of individual species of importan ce , whether they are potential or current invasive species (Ustin et al. 2002) , or rare and endangered species with high intrinsic or ecological value ( e.g. Dipterix panamensis , Cl ark et al. 2005 ). Furthermore , species maps at various time periods can show landscape changes due to climactic change or socio economic development (Kerr and Ostrovsky 2003) . Finally, species maps can aid in landscape estimates of carbon, where species id entity can be used to estimate aboveground biomass ( Colgan et al. 2013 ).
21 Species Map s from Hyperspectral Aerial Data Aerial remote sensing has the potential to characterize the species composition of trees on agricultural landscapes because it cover s large areas and has high spatial resolution capable of resolving individual tree crowns . Furthermore, a irborne hyperspectral sensors collect reflectance of many spectral wavelengths in the visible, near infrared and sometimes shortwave infrared spectral range , which provides data for identifying tropical tree species based on spectral differences of leaf chemical compounds and crown characteristics (Clark 2011). At the leaf level, the biochemical properties controlling spectral reflectance include; chlorophyll a and chlorophyll b , influencing reflectance in the visible range (VIS; 380 700 nm), and water, protein nitrogen, cellulose, and lignin influencing the reflectance in the near infrared (NIR; 700 1327 nm) and short wave infrared (SWIR 1467 2435 nm ) range (A sner 2008). Differences in these leaf components among species allows for spectral discrimination of species. In an airborne hyperspectral sensor , where reflectance is measured over a scale larger than a single leaf , the unique reflectance patterns among s pecies is blurred because of spectral interference from other sources, including spectral absorption of compounds in the atmosphere and shadows ( Clark et al. 2005 ). Spectral contributions to airborne reflectance data include; branch and crown level feature s such as, leaf and branch density, leaf angle, distribution, and clumping , and crown shape (Clark et al. 2005). When branch and crown features vary among species (such as between conifers, broadleaf, or palm species), spectral separability may improve (Al onzo et al. 2014). For species with similar crown structure, which is typical in tropical forests species (Clark 2011), spectral differences among species are driven by leaf biochemical composition and structure ,
22 which is not as pronounced among species wi th airborne sensors compared to hand held sensors which collect spectra directly on the tissue . Biophysical features of tree crowns are known to affect spectral reflectance across the full spectral range ( VIS SWIR ), which is important for spectral discrim ination of tropical tree species . Because of the high cost and operational difficulties of including a SWIR sensor, many sensors include just the VIS to NIR range (Clark 2011). However, hyperspectral sensors with the capacity to detect reflectance in the S WIR range been increasingly used for species classifications . It is an open question however whether including the SWIR sensor improves species classification enough to warrant the use of VIS SWIR sensors over VIS NIR sensors. For example, i n an African sa vanna, crown level accuracies for 7 (Cho et al. 2012) and 13 (Colgan et al. 2012) species were quite high (80% and 76%, respectively) using only VIS NIR images. Using the same sensor, a study of tropical species in Hawaii , Feret and Anser (2012 a ) achieved 73.2% accuracy for 17 tropical tree species. In a wet tropical forest of Costa Rica , Clark and Roberts (2012) used spectral data that included the SWIR bands and classified 7 species with 87% accuracy . While it is known that species have unique spectral re flectance in the SWIR range , studies of the additional accuracy achieved with the SWIR range cannot be directly compared due to differences in the number of species, their spectral variance, the dat a source, and the classification method. Therefore, the ut ility of spectral reflectance in the SWIR range for discriminating tropical species has not been directly add ressed. Relative to the number of species present on the landscape , tropical tree species classifications have included few species . For example, c lassification of a wet tropical
23 forest in Costa Rica have only included seven tree species (Table 2 1), whereas approximately 400 tree species are known to exist in that forest (Clark et a l . 2005). As there is a move towards application of species classifi cations to develop landscape species maps , the potential of spectral classification models that include more species needs to be addressed. While the Support Vector Machine algorithm (SVM), the technique used for classification in this study, has been show n to be well suited to studies with limited field data (Melgani and Bruzzone 2004, Pal and Mather 2005, FÃ©ret and Asner 2012 a ), the appropriate sample size for classification of a large number of tropical species has not been examined. Species classificati ons of airborne images are performed by building a data set of tree crowns in the image that have been identified to species in the field, and using this dataset for model development and evaluation ( Baldeck and Asner 2014). The crown dataset used to develo p a classification model is usually referred to as the training data , and the number of crowns per species in the training data referred to as the sample size . Since classification models are often applied to each individual pixel in the image , the pixel i s the unit for which to measure the size of the training data. To determine the necessary species sample size for hyperspectral image classification, Baldeck and Anser (2014) caution that analys e s based on the number of pixels per species may inaccurately estimate sample size needed . Instead, evaluation of species sample size and classification accuracy should be done at the crown scale . Baldeck and Anser (2014) conclude that the appropriate number of crowns per species depends on the number of species and their spectral separability, and therefore the minimum number of crowns to achieve high classification accuracy is entirely dependent on the
24 species to be classified. The challenge of tropical forest species classifications is to obtain the sufficient samp le size across many species, a task difficult due to the apparent unknown number of samples needed, and the low occurrenc e of many species in the field. In ad dition to the number of species and the sample size of each species , the accuracy of a classificat ion is also determined by the distribution of samples across all species . Tropical tree species are not found in equal abundance , a pattern more pronounced in agricultural landscapes where there is a dominance of species that have high value for farmers ( L ove and Spaner 2005 ), and very few individuals of most other species. Field collection of training data will often reflect this abundance distribution, where the sample size will be highly imbalanced across all species. Th e imbalance of samples across spec ies (often referred to as class imbalance) is a problem atic because classification algorithms will favor the majority classes (those with a large sample size), which results low classification accuracies for the minority classes ( Japkowicz and Stephen 2002 , Lin and Chen 2013 ) . For tropical tree species classification studies, the class imbalance problem translates to unequal accuracy between common and rare species, which could affect conclusions about species diversity and distribution patterns made from l andscape species predictions (Shao and Wu 2008) , such as the abundance or scarcity of certain species. Despite known classification inaccuracies due to class imbalance with high dimensional data, species classifications with hyperspectral data have not exp lored patterns of accuracy due to imbalance d training data , though this is somewhat addressed with semi supervised classifications where the species identity of input training data is not provided (F Ã© ret and Asner 2012b) . Therefore, the challenge of
25 suffic ient sample size is complicated by the need to achieve similar sample sizes across all species, or adjust current classification algorithms to account for class imbalance. Here we perform a species classification of 44 tropical tree species using a single high spatial resolution (2 m) full spectrum (VIS SWIR) hyperspectral image . Our objective is to understand the species abundance patterns of tropical agricultural trees based on the application of a SVM classification algorithm. We a ddress the following pa rameters of the SVM classifier by varying the input data and measuring o verall and species level accuracy ; (i) number of species, (ii) imbalance of species sample size, and (iii) t he SWIR wavelength range . We then apply the classification algorithm to disc riminate 37 species of agricultural tree crowns in a 22,85 7 ha study site to quantify tree species abundance patterns of the agricultural landscape. Methods Study S ite The species classification for a tropical landscape was tested o n the Azuero P eninsula o f Panama. The 8,000 km 2 peninsula is located on the Pacific side of Panama at approximately 7.5Â° N, and 80.5Â° W (Fig ure 2 1 ). The Azuero has a long history of agricultural development, which has left a landscape dominated by crop fields, pastures, and litt le forest cover, which is a result of its long history of forest clearing for cattle and farming initiated by Spanish colonists that intensified during the 2nd half of the 20th Century . Though now dominated by agricultural land use, the historical ecoregio n coverage of the peninsula are tropical dry broadleaf forest to the south and east, and moist broadleaf forest in the west (Olson et al. 2001). In the most southern region of the peninsula, mean annual rainfall is 1,946 Â± 65 mm yr 1 with 4.1 drought month s characterized by less than 100 mm of rainfall per month. The study site for this
26 research is 22, 85 7 ha in the southernmost region of the peninsula and is dominated by cattle pastures on steep slopes , narrow riparian forests, and small secondary forest fr agments (Fig ure 2 1). Aerial and Field Data Collection In January 2012 the Carnegie Airborne Observatory (CAO) flew 22,85 7 ha of the Southern Azuero Peninsula using the AToMS sensor . The AToMS sensor collects hyperspectral data (380 2510 nm; 10 nm bands) a t a spatial resolution of 2 m . Further details about the AToMS system can be found in Asner et al. 2012. In May July of 2012 and 2013 individual tree crowns were mapped to be used in the development and evaluation of the species classification model. With the help of botanists and access permission from landowners, t rees on private lands were visited and marked in high resolution (1.12 m) georeferenced images using a tablet computer equipped with a GPS (Xplore Technologies; Austin, TX). Tree crown boundarie s were digitized in the lab on the images, aided by additional high resolution images and lidar data, using ENVI (Exelis Visual Information Solutions, Boulder, Colorado ) . A total of 1,140 crowns from 44 species were included in the classification (Table 2 1). All species with less than three individuals were excluded, which eliminated 32 species from the field dataset. The three crown threshold was selected because we performed a 3 fold cross validation of the classification , in which the field dataset was divided into three groups; two for training the classification algorithm, and o ne for determining its accuracy . Therefore, species with less than three individuals were excluded because they would not be represented in one of those groups and would result in inaccurate accuracy calculations .
27 Classification Algorithm Development The Support Vector Machine algorithm (SVM) is widely used for species classification of hyperspectral data (FÃ©ret and Asner 2012, Colgan et al. 2012, Baldeck and Asner 2014) because of its ability to handle high dimensional data despite a small dataset size (Melgani and Bruzzone 2004, Pal and Mather 2005). SVM distinguishes among classes by finding the plane that maximizes their separability , where separability is determined by the di stance between the closest points in the class and the separation plane. For problems of high dimensionality where linear separation is not possible, such as multiple species classification of hyperspectral data, the data can be mapped into higher dimensio nal space with a transformation kernel , in which the linear problem can be solved (Melgani and Bruzzone 2004). Parameterization of the SVM algorithm requires selecting values for two variables, the width of the transformation kernel , which transforms the d ata to be solved linearly , and the penalty parameter (C), which influences the penalty of misclassification for non separable classes. For this study, a radial basis function kernel was used and its value was optimized by iteratively testing range s of valu es and selecting those that gave the highest accuracy , until an optimum value was reached . The same optimization process was performed for the C parameter . The kernel width and C parameter values determined by the optimization process were used in the spec ies classification algorithm. in which the pixel s of each crown were input to the classification , with the output of a predicted species for each pixel. Thus, most crowns include pixels classified to more tha n one species . To produce a species prediction at the crown scale
28 method, in which each crown was assigned to the species that had the most pixels predicted as that species ( Clark et al. 2005) . Using the k f old design, classification accuracy was assessed on an independent group of tree crowns . Classifications were run in R statistical software (R Development Core Team package (Myer et al. 2013). Colgan et al. (2012) found that SVM cla ssification of African savanna species worked best with filters that removed pixels of high shade or low leaf density. W e selected pixels with the strongest live vegetation signal by removing those with low near infrared (NIR) reflectance (shaded) and a lo w value in the Normalized Difference Vegetation Index (NDVI ; low leaf density). Removal of pixels with NIR values (860 nm) less than 30 % and less than 0.5 in NDIV resulted in the optimum classification accuracy ( Appendix A ). These filters were applied to a ll pixels prior to running the SVM classification. Classification Algorithm Testing After the SVM parameters were optimized and the NDVI and NIR filters applied, we tested how varying the following three aspects of the SVM algorithm affected species classi fication: number of species ; number of crowns per species ; and , spectral range of the data. Number of species . We ran a series of classifications in which the number of species was progressively reduced from the full 44 to only six by s uccessively removing species with the fewest number of crowns per species. The full number of crowns per species, which ranged from 4 to 119 (Table 2 1 ), was used for each species in the classification.
29 Maximum number of crowns . While the most common species had many crowns i n the field data set, there were many species only represented by a few crowns. We tested if reducing the number of crowns per species for the most abundant species in the field set affected the classification accuracy. We first classified all 44 species w ith all crowns in the dataset. This represented the most uneven field dataset because the number of crowns per species was 4 119. Next, we set a maximum number of crowns per species (N) at 5, 15, 25, 35, 45, and 55 crowns per species. This process was repe ated 5 times for each value of N, drawing a new random sample of crowns each time. For species with fewer than N crowns , all crowns of that species were included to ensure that no species were excluded across the range of N , and so we could compare results for the full set of 44 species . Spectral range . As a final test, we ran the classification with a progressively larger spectral range. The first run included only bands from the visible (VIS) range (437 700 nm), the second from the VIS through the near in frared (NIR) range (437 1327 nm), third from VIS through the first region of the shortwave infrared (SWIR1) range (437 1771 nm), and fourth from VIS through the second region of the shortwave infrared (SWIR2) range (437 2435 nm). The difference between acc uracies with the full range (VIS SWIR2) and with each smaller range (VIS only, VIS NIR, and VIS SWIR1), were compared for all species and for individual species. Classification Algorithm A ccuracy We report overall accuracy, errors of commission, errors of omission (Congalton 1991), and prediction bias. Overall accuracy is the percentage of crowns that have been correctly predicted. Reported accuracy of remote sensing classifications also usually includes user and producer accuracy. However, we report accura cy in terms of
30 commission and omission error because it is more intuitive and can be directly translated into a metric of prediction bias. Commission error (1 user accuracy) is the percentage of crowns that were classified as a species when it was not that species. Omission error (1 producer accuracy) is the percentage of crowns that were not classified as a species when it was that species (Congalton 1991). Prediction bias was calculated as the difference between commission and omission error. S pecies, for which commission error is higher than omission error, have positive prediction bias, meaning the classification resulted in more crowns of that species than exist in reality . The opposite is true for when omission error is higher than commission error. Us ing measurements of prediction bias is useful when the application of the classification is to determine relative abundances of species because the predicated number of crowns per species can be adjusted based on the prediction bias. Application to Landsca pe To produce a landscape species map, we applied a final classification model to individual agricultural tree crowns throughout the entire study site. An image segmentation function in the SAGA GIS program ( http: //www.saga gis.org/en/ ) was applied to 1.12 m lidar vegetation height was used to map the boundaries of all tree crowns for the full extent of the study site. The landscape consists of both agricultural trees and forest trees, both of which were segmented from the lidar data. Since we focused field sampling on agricultural trees, we could only apply the classification to these trees, which required distinguishing those crowns from crowns in the continuous forest. Tree crowns were classified as forest or ag ricultural trees based on the number of neighbors and the amount of edge shared with another tree crown. Crowns with less than two neighboring crowns or less than 50% of their edge touching another crown
31 were classified as agricultural trees because they w ere relatively isolated from neighboring crowns and therefore not part of a forest fragment with continuous canopy cover. Crowns with more than two neighbors and greater than 50% of their edge touching another crown were classified as forest trees. This me thod effectively removed all continuous forest tree crowns, including those in riparian forests. The SVM model was optimized based on the classification tests described above. Though 44 species were included in the training data set, only 37 species were i ncluded in the final model. The seven species removed had very low classification accuracy and in initial applications composed less than 1% of the trees in the landscape. Final model accuracy statistics were generated by running the optimized SVM algorith m 100 times and average the results . This final model was then applied to the landscape by predicting the species of every pixel and crown of agricultural trees. The output of this was a species map where all individual agricultural trees have a predicted species and a prediction error. The species distributions were calculated across the landscape and corrected based on relative over or under prediction of the classification. Results Number of S pecies Classifying the six most abundan t species in the field dataset had an accuracy of 80.7 % (Fig ure 2 2 a ). With 14 species, classification accuracy was 73.4%, and decreased at a similar rate until 31 species were included . Accuracy declined slowly in classifications of over 31 species. When all 44 species were inc luded, overall accuracy was 59.7%.
32 The pattern of overall accuracy by the number of species reflected the pattern of overall accuracy based on the number of crowns per species (Fig ure 2 2 b ). As the minimum number of crowns per species included decreased , t he number of species included in the classification increased. This suggests the accuracy trend was driven in part by the sample sizes of the species included in the classification, which is explored more in the next section . Number of C rowns For classific ations of 44 species, there was a non linear negative correlation between the number of crowns per species and the species error (Fig ure 2 3 a ). While similar trends were seen for omission and commission errors, there were slight differences between errors for each species. The prediction bias showed a non linear positive correlation with the number of crowns of each species (Fig ure 2 3 b ). For 31 out of 33 species with less than 40 crowns , the prediction bias was 2 to 77%, whereas only one species with ove r 40 crowns ( Spondias mombin ) had a slight negative prediction bias. When the maximum number of crowns per species was lowered, thus reducing the abundance of the species with many crowns in the field set, the overall accuracy declined while the prediction bias approached zero (Fig ure 2 4 ). Compared to the classification with the full set of crowns, the overall accuracy was lower at all values of N, with a drastic decline at N less than 35 . The average prediction bias for all species, which was 12% when al l crowns were included, approached zero as N decreased. When only 5 crowns per species were included, which represents nearly equal representation of all 44 species, the overall accuracy is at its lowest at 27% with a 2.5% prediction bias.
33 Spectral R ange Including the full spectral range gave the highest overall accuracy (Table 2 3 ). Species classification of 44 species with only the VIS range had an accuracy of 37.7%. Accuracy increased to 51.5% with the inclusion of the NIR range, an additional 13.8% in accuracy. The addition of the SWIR1 and SWIR2 ranges improved accuracy by 4.4% and 2.3%, respectively. For individual species, accuracy improvements with a larger spectral range followed similar patterns as overall accuracy, in which large increases were s een with the addition of NIR reflectance and smaller increases with the addition of SWIR reflectance (Fig ure 2 5 ) . However , for some species, accuracy decreased with a larger spectral range, primarily with the inclusion of the SWIR range. Final Classificat ion Model and Landscape Species Prediction Accuracy was 61.4% for 37 species, and ranged from 8.3 96.7% across individual species (Table 2 4) . The species composition from application of our classification revealed a landscape dominated by a few species, e ven after adjusting species abundance by the prediction error (Fig ure 2 5 ). The most common species, Guazuma ulifolia , represented 13.8% of the total agricultural tree crowns. Nearly 80% of the trees were found in the 10 most common species. Discussion Thi s study produced a classification of 37 species with 61% overall accuracy with high resolution (2 m) hyperspectral (380 2510 nm) aerial image (Table 2 4 ). Application of the classification model produced a landscape species abundance prediction, which show ed the dominance of species with high value to humans (Figure 2 5 ). As field studies of numerous individual farms suggest, the species composition of agricultural trees is dominated by species that have high value to farmers, though there are other
34 species as well. The results of this study highlight two important aspects to consider in application of species classification models to landscape studies: the resolution of the data, and the size of the input data. Dimensions of Hyperspectral Aerial Data and Th eir Importance for Species Classification S pecies classification s h ave been done at multiple scales in other studies (Table 2 1); we show the potential of using a single 2 m full spectral resolution image for tropical tree species classification. The colle ction and analysis of r emote sensing data involves trade offs along different dimensions of data resolution (Key et al. 2001) . D ata resolution is typically discussed along three dimensions ; spatial ( area represented by each measurement/pixel ), spectral (wi dth, number, and range of spectral bands), and temporal (frequency of data collection) . Spatial resolution . Species classification accuracy is largely dependent on the spatial scale at which the data is collected, in which classification accuracy tends to decrease with an incre ase in spatial scale due to increased spectral variability within a species (Clark et al. 2005) . C rown level classifications are the most relevant for applications to landscapes (Baldeck and Asner 2014). Early work in spectral discrim ination of species used hand help spectral systems, where data was collected on individual leaves, and interference from the atmosphere was not a factor (Price 1994, Cochrane 2000). These studies were cautiously optimistic in the ability to discriminate sp ecies based on spectra, but highlighted the potential, especially with hyperspectral data from the full spectral range . With the advent of hyperspectral sensors on aerial platforms, tree species classification shifted to pixel and crown level classificatio ns. At this level, the spectral signal of the biophysical components of tissues is diluted, in part
35 due to scatter reflectance from other components of branches (though this depends on the spatial resolution of the pixel) and from the atmosphere. Scaling f rom pixel to crown level can be done prior to classification, by averaging the spectral values within a crown, or post classification, by summarizing pixel level predictions for each crown. Methods like the majority rule (Clark et al. 2005, Feret and Asner 2012a), which was also used in this study, or performing a second classification for crowns using species prediction probabilities (Colgan et al. 2012), have higher accuracy than pre classification scaling methods (Clark et al. 2005) . Though it is not rep orted in the results, pixel level accuracy for this study was consistently higher than crown level accuracy. To improve crown level accuracy, using a secondary classification as in Colgan et al. (2012) could be explored. There has been a lot of work in sca ling from leaf to crown level species classification with the consensus that crown level species predictions are useful for application studies despite the reduced accuracy. Temporal resolution . Most tropical species classifications have been done with a s ingle image, which provides a snap shot of the spectral reflectance of species. Therefore, the spectral discrimination among species is determined by differences in reflectance at one point in time and does not incorporate temporal variation. T emporal vari ation in species reflectance is especially important for landscapes with seasonal variation in reflectance, such as tropical dry forest s, where a pronounced dr y season causes trees to lose their leaves ( Castro Esau and Kalacska 2008 ) . Thus, incorporating t emporal variation into species classifications may increase spectral separability, because species differences may become greater, or may decrease spectral separability, because many species will not have leaf reflectance, which has been the
36 basis for spec ies spectr al separability (but see Clark and Roberts 2012). The image used in this study was collected at the beginning of the dry season. Many of the deciduous species that lose their leaves in the dry season still had leaves during this time. This is sup ported by the high average NDVI values for the focus species ( Table 2 2 ). Given the distinct spectral reflectance of bark tissue (Bohlman 2008, Clark and Roberts 2012), potential exists for dry season image classification and determining the separability o f deciduous crowns with aerial hyperspectral data . Spectral resolution . F ull spectrum data (VIS SWIR) helped achieve high classification a ccuracies, but our results indicate that species classification with only VIS NIR achieves similar classification accu racy. While a substantial increase in overall accuracy of the full range compared to the visible range alone was expected (Rivard et al. 2008), the inclusion of the full wavelength range (380 2510 nm) compared to just visible through near infrared (380 132 7 nm) increased the overall classification accuracy by 7%. S ince this image was taken when most trees had leaves, the shortwave infrared range may not have been as important during this time of year . If the image was taken during the dry season, including the SWIR range may have improved discrimination between deciduous species because there would be more wood tissue reflectance , which has been shown to have high SWIR reflectance (Clark and Roberts 2012). W hile the SWIR range did contribute to high accuracy , sensors with VIS NIR still have high potential to discriminate among species, meaning the collection of this data is still valuabl e to landscape species studies. Training Data Effects on Classification Accuracy Classifications with a greater number of sp ecies results in lower classification accuracies, though the pattern is dependent upon the type of the classification algorithm
37 (Feret and Asner 2012a) . Additionally, classification accuracy depends on the number of samples for each class used to train the classification algorithm. Therefore, c lassification of tropical tree species involves a trade off between the classification accuracy, and the size of the training data, which can also be measured in terms of the investment in field data collection . Numbe r of species . Aerial hyperspectral data has the potential to produce high resolution thematic maps of tree species identity. It is likely that species classifications will always only include a subset of the species found in the ground, due in part to the ability to collect sufficient field data to accuracy classify rare species. The optimal number of species depends on the application of the classification. F or example, if only interested in dominance of G uazuma ulimfolia , one c ould only map this species, with the rest in a single group, or only include the other very common species. Rather than be guided by the number of species, which may involve ranking species based on perceived importance, the desired level of accuracy can also guide how many species a re classified. While the number of species or overall accuracy targets may not be easy to determine for many ap plications, classifications should at least balance omission and commission errors for each species. With this frame of mind, the way in which tr aining data should be collected, is to balance the sample size of the species . Training data sample size . SVM produced a classification with high accuracy, but is not immune to misclassifications due to data limitations. SVM has been widely used, due in pa rt to its ability to discriminate among classes with a small sample size. However, there is still some minimum sample size necessary to achieve accurate predictions. Colgan et al. (2012) found the biggest increase in accuracy was caused by
38 an increase the sample size. The results of our study show the same, and suggest that with more field data, primarily for species with less than 30 crowns, classification accuracy may increase. This is supported by our finding of a strong negative correlation between spec ies sample size and classification error, and a positive correlation between species sample size and prediction error. While various classification studies report producer and user accuracies, they are rarely discussed in terms of the relative species comp osition from the classification predictions. When considering classifications with multiple species, especially those with imbalanced training data, the sample size for each species has a large effect on the prediction accuracy, specifically the relative o ver and under prediction of species based on their sample size. In a landscape dominated by a few species, field data collection will tend to be skewed (i.e. more samples of the dominant species). This bias in the field data affects differences between omi ssion and commission error. Large differences between these types of errors mean that there is large over or underestimation of these species on the landscape, resulting artificially skewed species distributions. While there is a desire to use all training crowns, the unequal distribution of the sample size across species affects the overall accuracy, and species over or underestimation. Patterns in accuracy with species sample size are problematic in application of the classification because common species will tend to be over predicted, whereas uncommon species will tend to be under predicted. In application to a landscape with highly uneven species abundance distributions, the classification will exacerbate the range in species abundance and make the land scape species abundance distribution appear more uneven. In applications focused on mapping a
39 single species of importance, such as rare or invasive species, the effect of class imbalance is especially relevant because the abundance of these focal species can be very biased. In mapping relative species abundance of many species, with equal importance, the class imbalance problem is less important. However, these types of errors can influence landscape estimates if they are not accounted for in landscape ana lyses. In applications which require accurate classifications of the dominant species, without much concern for the rare species, class size imbalance is less of a problem. One example of this type of application is estimation of tree biomass, where specie s identify is used in biomass equations based on tree size. In this application, the species is used to factor in a species specific value of wood density. Errors in the rare species are very mi nimal compared to other errors. Algorithm S election While the SVM classification algorithm has high accuracy for tropical tree species classification, we question its utility in application to landscapes outside of where the field data were collected. The main reason is that the SVM classification is designed to pred ict a pixel as one of the species in the field dataset. If a species exists on the ground that does not exist in the field dataset, it will be forced into one of the species known to the classification algorithm. While we show that the SVM classifier can d iscriminate among 37 species with 61% accuracy, this is only a fraction of the species that actually exist on the landscape. For example, the original field dataset included 76 species that we know exist on the landscape. However, the species prediction ma p only has 37 species, a direct factor of the input species to the classification. Other classification methods that allow for pixels to remain unclassified if they are not spectrally similar to one of the input classes would be more useful for application to the
40 landscape. In this way, if a species exists on the ground and not in the training dataset (because it was not sampled in the field or the sample size was too small), there is than being forced into a species known by the classification algorithm. Conclusion Airborne hyperspectral remote sensing opens the door for a new perspective of landscapes . H ere we map species distribution for common tropical species . The species map prod uced from this study hel ps to understand the species composition of this landscape, which can be used to examine the drivers of species distribution patterns, and serve as a baseline for measuring change. A erial species mapping has high potential for ecolo gical research , application need to consider the type of data. Tree species classifications in species rich landscapes involve a tradeoff in the number of species to be classified and the accuracy of the classification. F or agricultural landscapes this per spective is important because we can capture data about functional landscapes, specifically, the tree species within them . This study support s field specific studies showing high diversity and value of trees in agricultural landscapes .
41 Table 2 1. Summar y of species classification studies with hyperspectral data. Study, location, sensor Spatial resolution Spectral resolution Number of species Main findings Cochrane 2000 Para, Brazil FieldSpec ASD Leaf scale VIS NIR (450 950 nm) 11 species 1 5 trees and 3 36 samples per species Spectral signatures are not unique, but there is potential to separate species based on foliar reflectance Clark, Roberts, and Clark 2005 La Selva, Costa Rica HYDICE 1.6 m Pixel scale Crown scale VIS SWIR (400 2500 nm; 210 bands) 7 species 10 34 trees per species Decrease in accuracy from fine (leaf scale) to coarse (crown) scale 92% classification accuracy for individual tree crowns with 30 spectral bands Clark and Roberts 2012 La Selva, Costa Rica HYDICE ASD spectrometer 1.6 m Tissue scale (leaves and bark) Pixel scale Crown scale VIS SWIR (400 2500 nm; 210 bands) VIS SWIR (350 2500 nm) 7 species 10 34 trees and 300 pixels per species Leaves and bark have distinct spectral characteristics, which increases spectral variability at larger scales Crown level accuracy was 87% using the pixel majority method. Feret and Asner 2012a Hawaii, USA CAO Alpha system 0.56 m Pixel scale Crown scale VIS NIR (390 1044 nm; 24 bands) 17 species 1 168 trees and 117 53,710 pixels per species 73 % accuracy achieved with only spectral data Feret and Asner 2012b Hawaii, USA CAO Alpha system 0.56 m Pixel scale VIS NIR (390 1044 nm; 24 bands) 9 species 12 135 trees per species 61% classification accuracy achieved using semi supervised classificatio n Cho et al. 2012 Kruger National Park, South Africa CAO Alpha system 1.1 m Pixel scale VIS NIR (384 1054 nm; 72 bands) 7 species 88% classification accuracy achieved with 3x3 majority filter Colgan et al. 2012 Kruger National Park, South Africa CAO A lpha system 1.1 m Pixel scale Crown scale VIS NIR (384 1054 nm; 72 bands) 15 species 12 73 crowns and 114 3693 samples/pixels per species 76% classification accuracy achieved with only spectral data Accuracy improved with second SVM to scale from pixel to crown scale Notes: See Clark 2011 for overview of hyperspectral sensors. The main findings listed here are specific to species classifica tions with spectral data.
42 Table 2 2 . Sample size and mean NDVI of tree species included in this study. Species c ode Species Family Crowns Pixels Mean NDVI ALBIAD Albizia adinocephala Fabaceae 4 109 0.89 ALBIGU Albizia guachapele Fabaceae 9 591 0.91 ANACEX Anacardium excelsum Anacardiaceae 32 949 0.86 ANDIIN Andira inermis Fabaceae 24 835 0.81 ANNOPU Annona purp urea Annonaceae 15 240 0.85 AST2GR Astronium graveolens Anacardiaceae 14 588 0.86 BURSSI Bursera simaruba Burseraceae 18 238 0.79 BYRSCR Byrsonima crassifolia Malpighiaceae 29 563 0.85 CALYCA Calycophyllum candidissimum Rubiaceae 60 1540 0.77 CECRIN C ecropia insignis Cecropiaceae 5 39 0.84 CEDROD Cedrela odorata Meliaceae 84 2451 0.80 CHR2CA Chrysophyllum cainito Sapotaceae 5 139 0.83 COCCCA Coccoloba caracasana Polygonaceae 24 565 0.86 COCCLA Coccoloba lasseri Polygonaceae 16 505 0.84 CORDAL Cord ia alliodora Boraginaceae 32 544 0.81 CORDCO Cordia collococca Boraginaceae 9 158 0.85 DALBRE Dalbergia retusa Fabaceae 11 272 0.84 DIPHAM Diphysa americana Fabaceae 54 1057 0.88 ENTECY Enterolobium cyclocarpum Fabaceae 83 7366 0.86 COLUHE Colubrina h eteroneura Rhamnaceae 6 432 0.86 FICUSP Ficus spp. Moraceae 13 996 0.87 GENIAM Genipa americana Rubiaceae 25 398 0.84 GMELAR Gmelina arborea Lamiaceae 4 37 0.81 PSIDSP Psidium spp. Myrtaceae 11 208 0.87 GUAZUL Guazuma ulmifolia Malvaceae 119 4411 0.82 HURACR Hura crepitans Euphorbiaceae 62 3051 0.84 HYMECO Hymenaea courbaril Fabaceae 8 182 0.85 LUEHSE Luehea seemannii Malvaceae 21 650 0.86 MACLTI Maclura tinctoria Moraceae 20 924 0.83 MANGIN Mangifera indica Anacardiaceae 11 214 0.85 MANIZA Manil kara zapota Sapotaceae 4 120 0.92 OCHRPY Ochroma pyramidale Malvaceae 4 85 0.89 PLA1PI Platymiscium pinnatum Fabaceae 47 1841 0.85 POCHQU Pachira quinata Bombacaceae 30 1084 0.67 PSEUSE Pseudobombax septenatum Malvaceae 8 170 0.77 PSIDGU Psidium guaja ba Myrtaceae 11 137 0.83 SAPIGL Sapium glandulosum Euphorbiaceae 24 623 0.83 SCIAEX Sciadodendron excelsum Araliaceae 22 316 0.74 SPONMO Spondias mombin Anacardiaceae 78 2687 0.82
43 Table 2 2. Continued Species code Species Family Crowns Pixels Mean ND VI STERAP Sterculia apetala Malvaceae 21 971 0.80 TAB1GU Tabebuia guayacan Bignoniaceae 4 70 0.79 TAB1OC Tabebuia ochraceae Bignoniaceae 15 277 0.77 TAB1RO Tabebuia rosea Bignoniaceae 40 893 0.82 ZANTP1 Zanthoxylum panamense Rubiaceae 4 66 0.85 Notes : Mean NDVI (NDVI = NIR 860nm RED 650nm /NIR 860nm +RED 650nm ) was calculated for each pixel prior to NDVI and NIR filters. Table 2 3 . Overall accuracy of classifications with different spectral ranges. Spectral range Crown accuracy Percent increase VIS 37 .7% VIS NIR 51.5% 13.8% VIS SWIR1 55.9% 4.4% VIS SWIR2 58.2% 2.3% Total 20.4% Table 2 4 . Species c lassification accuracy of the final SVM model with 37 species Species Average accuracy Prediction bias Albizia guachapele 96.7% 51.1% Guazuma ulmi folia 87.1% 21.3% Calycophyllum candidissimum 85.2% 18.0% Diphysa americana 82.7% 4.6% Enterolobium cyclocarpum 80.7% 16.4% Cedrela odorata 80.1% 14.4% Platymiscium pinnatum 71.1% 2.6% Byrsonima crassifolia 68.7% 0.5% Annona purpurea 66.7% 8.1% Hura crepitans 62.5% 0.3% Anacardium excelsum 60.7% 15.9% Spondias mombin 60.3% 6.3% Sterculia apetala 56.7% 24.7% Luehea seemannii 54.3% 11.6% Chrysophyllum cainito 53.3% 2.6% Cordia alliodora 52.4% 14.1% Ochroma pyramidale 51.7% 25.6% Tabe buia rosea 45.4% 13.5% Andira inermis 43.8% 0.8% Tabebuia ochraceae 41.0% 11.6%
44 Table 2 4. Continued. Species Average accuracy Prediction bias Maclura tinctoria 40.1% 37.9% Genipa americana 39.6% 24.0% Psidium spp. 39.4% 17.0% Astronium graveolens 39.2% 14.1% Sapium glandulosum 38.8% 22.2% Cecropia insignis 38.3% 61.7% Coccoloba caracasana 28.8% 5.4% Ficus spp. 24.2% 9.9% Manilkara zapota 20.0% 18.1% Mangifera indica 19.7% 16.7% Tabebuia guayacan 18.3% 40.8% Psidium guajab a 17.2% 15.6% Dalbergia retusa 15.6% 40.8% Albizia adinocephala 13.3% 8.5% Coccoloba lasseri 11.3% 8.2% Sciadodendron excelsum 8.5% 24.4% Cordia collococca 8.3% 29.9% Figure 2 1 . Study site on the Azuero Peninsul a. Outline d region is the 22,587 ha study site of aerial data collection by the Carnegie Airborne Observatory.
45 Figure 2 2. The o verall accuracy of multiple SVM classifications with change s in the number of classified species. Overall accuracy as measure d by the number of species ( a ) and as measured by the minimum crown number across all species ( b ).
46 Figure 2 3 . Error and prediction bias for a classification with 44 species. (A) Omission (white circles) and commission (grey squares) error for each spec ies based on the number of crowns in the training data sample. (B) Prediction bias for each species based on the number of crowns in the training data sample. Prediction bias was determined as the difference between errors of commission and errors of omiss ion, where a positive value indicates the species was over predicted.
47 Figure 2 4. O verall accuracy and prediction bias for classifications with the full set of training data to only 5 crowns per species. The maximum number of crowns was reduced from ful l training data sample with up to 119 crowns per species (light grey squares, same data as white diamonds in Figure 4), to a range of 5 55 crowns.
48 Figure 2 5 . Differences in classification accuracies for 44 species with the full spectral range to cla ssifications with reduced spectral ranges Plotted differences are the species level accuracies achieved using VIS SWIR2 wavelengths, minus the accuracies achieved using only VIS (437 700 nm, light grey), VIS NIR (437 1327 nm , medium grey) and VIS SWIR1 (43 7 2435 nm , dark grey). A ccuracy across all 44 species is also plotted (overall) and specific values are found in Table 2 .
49 Figure 2 6. Predicted abundance for 37 species across 22,587 ha study site. Predicted values are from classification output. Correc ted values are adjusted based on the prediction bias for the final classification model (corrected = predicted * prediction bias).
50 CHAPTER 3 COMPARISON OF INDIVIDUAL TREE AND PLOT BASED ABOVEGROUND BIOMASS (AGB) ESTIMATES FROM AERIAL LIDAR DATA FOR A TRO PICAL AGRICULTURAL LANDSCAPE Introduction A t 3 8 agric ulture exceeds forest as the dominant global biome ( FAOSTAT 2014 ) . Across the globe, trees cover approximately 27% of agricultural area s and , if aggregated globally cover 17 mil lion km 2 (Zomer et al. 2009) . In the tropics, agricultural tree cover is particularly high, estimated at 30 90% in Central American and South East Asia. Agricultural trees are individual trees, scattered groups of trees, live fences, and windbreaks (Plieni nger 2012) that exist outside of forests in areas of agricultural production. Agricultural trees provide a diversity of benefits, from wildlife habitat and seed sources for forest recovery, to valuable products for landowners (Corbin and Holl 2012, Harvey and Haber 1999, Slocum and Horvitz 2000, Harvey et al. 2005, Harvey et al. 2006, LeÃ³n and Harvey 2006, Medina et al. 2007, Zahawi et al. 2013). Though agricultural trees can be part of agroforestry systems , in which woody and herbaceous species interact an d are managed for annual or perennial production (Somarriba 1992) , they do not necessarily occur as part of a planned management system. In addition to ecosystem services provided to farmers, wildlife , and cattle, agricultural trees may also be important s ources of global ecosystem benefits, namely carbon sequestration and storage . In an agricultural landscape in Western Kenya for example , the above ground biomass ( AGB ) of agri cultural trees in over 10,000 hectares was estimated at 36 Mg ha 1 , or 17.5 Mg C h a 1 (Kuyah et al. 2012 a ) . Alth ough compared to forested areas biomass per unit area is low , biomass contribution of
5 1 agricultural trees can be substantial because land in agricultural production is now a (Ramankutty et a l. 2008) . Accurate estimates of AGB are important for understanding global carbon and energy cycles, but also for operation of global carbon markets. Species diversity and carbon stocks of tropical forests grow in importance with development of internation al policies that use a system of carbon credits , designed to reduce emissions from land cover change. Panama , where this study was conducted, is one of the first countries to participate in the readiness phase of the REDD+ program ( William 2013 ) . In a surv ey of the readiness of Panama, Pelletier et al. (2011) found that the largest source of uncertainty in carbon emissions for the country, and therefore a limitation to the successful implementation of this policy, i s spatial variability of carbon stocks of mature forests , secondary forests, and fallow land. As a way to overcome the uncertainty in carbon stocks across the landscape, remote sensing data and analysis methods are being developed to estimate biomass and carbon stocks of tropical landscapes (Gibbs et al. 2007) . Airborne and space based lidar (light detection and ranging) measurements are increasingly used to estimate AGB. Frequently, these methods rely on canopy height and provide an estimate of AGB on an area basis. For example, Asner et al. (2013 ) used plot based methods to estimate country wide forest carbon stocks for Panama. However, plot based methods may have different level of error in different land cover types. For example, agricultural landscapes are not well suited for plot level studies because of spatial variability of trees over small areas (SÃ¡nchez Azofeifa et al. 2009) . In landscapes with dispersed tree cover, such as African savannas, AGB estimates are
52 frequently more accurate when individual trees are detected and quantified than w ith plot based method s (Colgan et al. 2013) . While Colgan et al. (2013) showed that individual tree as the unit of biomass estimation provided much more accurate estimates of biomass than plot based estimates; this has not been attempted in tropical agricu ltural landscape s . Here we develop a tree based method for estimating AGB and compare it with plot based AGB estimates for the same landscape. We a lso provide detailed estimates of agricultural trees in comparison to trees found in larger forest fragments. This study addresses the questions of (1) w hat is the AGB contribution of agricultural trees in a tropical landscape and what is the relative contribution of agricultural tree AGB compared to trees in forests , and (2) h ow do individual tree and plot bas ed approaches to AGB estimation vary, and what is the difference in the landscape AGB prediction? Methods Study Site The Azuero P eninsula of Panama has a long history of agricultural development, which has left a landscape dominated by crop fields and past ure. The 8,000 km 2 p eninsula is located on the Pacific side of the Panama located approximat ely at 7.5Â° N, and 80.5Â° W (Figure 3 1). Though the peninsula is now dominated by non forest land use, the historical ecoregion coverage of the area are tropical dr y broadleaf forest to the south and east, and moist broadleaf forest in the west, based on temperature, precipitation, and species distribution patterns (Olson et al. 2001) . In the most southern region of the peninsula, mean annual rainfall is 1,946 Â± 65 m m yr 1 with 4.1 drought months characterized by less than 100 mm of rainfall per month (Wishnie et al. 2007) . The current scarcity of f orest on the Azuero is a result of its long history of forest
53 clearing for cattle and farming that was initiated by Spani sh colonists and intensified during the latter half of the 20th Century. The study site for this research is 22 , 587 ha area in the southernmost region of the Azuero P eninsula (Fig ure 3 1 ) and is dominated by cattle pastures on hills with steep slopes . Aeri al D ata High resolution lidar and hyperspectral images of a 22,587 ha landscape of the southern Azuero Peninsula were used in this study to estimate landscape AGB. I n January 2012 the Carnegie Airborne Observatory (CAO) collected hyperspectral and lidar da ta using the AToMS system (more information found in Asner et al. 2012). The 2 m hyperspectral image was used to manually digitize a sample of tree crowns visited in the field. The data used for this study w ere discrete lidar return that provided a ground surface image (last return) and a maximum height image (first return), from which we were able to derive the height of surface features (difference between first and last return) . The 1.12 m lidar data aided in crown digitization and was used to collect tr ee dimens ions of height and crown area. Overview We used five steps to calculate landscape AGB based on individual trees with field and lidar data (Fig ure 3 2 ). First, we used standard allometric methods (Chave et al. 2005) to estimate AGB of 1,100 individ ual trees using field measured diameter and tree height, and species specific wood density from a pantropical dataset ( Chave et al. 2006). Second, for the field measured trees , we developed a relationship that predicted the individual tree AGB estimate ( fr om step one ) using lidar metrics of crown size and tree height. Third, we used lidar data to segment all individual crowns on the landscape, then classify tree crowns as either
54 areas, or those that occur as part of a forest with continuous canopy cover. Fourth, using the equation to predict field based individual tree AGB from lidar metrics ( from step two) and the segmented tree crowns (from step three) , we calculated the AGB of all trees in the landscape to produce a dataset of landscape AGB density (Mg ha 1 ). Fifth, we compared the landscape AGB density estimat e from individual trees (from step four) with an estimate f or the same study site from Asner et al. (2013) that was created with a plot based method. W e calculated the differences between the two estimates and trends across a gradient of tree cover. Development o f Field Based Allometric AGB Model In May July of 2012 and 2013, a sample of trees on private lands were located and marke d in high resolution (1.12 m) georeferenced images using a tablet computer equipped with a GPS (Xplore Technologies; Austin, TX). With a botanist who could identify tree species, and l andowner permission to access private lands, tree species, height and di ameter at breast height (at 1.3 m) were determined for each tree . Maximum tree height was measured with a laser altimeter (Nikon Forestry 550, Nikon Corporation; Tokyo, Japan) . Tree crown boundaries were digitized in the lab on the image with visualization aid from additional high resolution images and lidar data using ENVI (Exelis Visual Information Solutions ; Boulder, Colorado). Of the 1,396 trees sampled, 1,100 trees were used for the analysis. The 296 crowns excluded from analysis were either too small (less than 1 full pixel at 2 m resolution) or the crown boundaries were uncertain. Individual tree AGB was estimated with pantropical allometric relationships using tree diameter, height, and taxon specific wood density . Though we were not able to directly measure tree AGB by felling and weighing individual trees, the Chave et al.
55 (2005) allometric models have been widely used for estimating AGB across the tropics , and is the basis for quantification of current forest carbon pools in the Global Forest Carbo n Initiative of the Center for Tropical Forest Science (CTFS) of the Smithsonian Tropical Research Institute . In addition , the Chave et al. (2005) allometric models have shown to be a good approximation in several studies directly related to estimating AGB of dispersed tree cover. In a study of dispersed trees in an African savanna , which was the model for the method in this study , including a sa vanna landscape in South Africa (Colgan et al. 2013) , and an agricultural landscape in Western Kenya, ( Kuyah et a l. 2012 b ) . The Chave allometric model used in this study (Chave et al. 2014) contains updated allometric relationships from the Chave et al. (2005) allometric model and additional data specifically from dry forest and woodland ecosystems, which are found i n this Azuero Peninsula study site . We calculated AGB with the Chave et al. (2014) pantropical equation: AGB chave =0.0673 *(WD*D 2 *H) 0.976 ( 3 1 ) where; AGB chave is the estimated biomass ( k g ), D = tree diameter at breast height (cm), H = tree height (m) and WD = taxon specific wood density (g c m 3 ). Diameter was measured for all stems in the field. Though tree height was measured in the field, we use lidar tree height for the AGB calculation. Field and lidar heights were very similar, but lidar height was slightly larger than field height for all trees (Appendix B ). WD was obtained using a Neotropical wood density database (Chave et al. 2006) . Of the 50 species included in this study, 38 had WD values in the Neotropical database. For the remaining 12 speci es, genus or family level wood density average were used. Trees with multiple stems were common , with 30% of trees having more than a single stem.
56 For trees with multiple stems, we applied the Chave model to each stem, and summed the stem level AGB to gene rate a whole tree estimate of AGB (Chave et al. 2008) . Development of lidar AGB model Using AGB chave as our observed AGB for the set of 1,100 trees, we developed a predictive AGB allometric model based on lidar derived tree size metrics of crown size and m aximum crown height, as well as taxon specific wood density (AGB lidar ) . Crown area w as determined from the hand digitized tree crown s for all field measured trees . We explored several functional forms to relate lidar size metrics to AGB, with a simple line ar relationship between lidar derived crown area (CA) and tree height (H) found to be the best predictor of AGB chave (Appendix C). For the Azuero Peninsula, a species map was produced for 37 species using a hyperspectral data of the same spatial resolution and geographic pixel location as the lidar data (Asner et al. 2012) , providing the opportunity to include species information for landscape AGB estimates . Therefore, w e also included a term for taxon specific WD in the AGB lidar model . The final AGB lidar m odel chosen was : log(AGB lidar )=0.26 +0.61*log(CA)+1.7*log(H) ( 3 2 ) where the intercept term includes a correction factor from the residual standard error (RSE) of the final model calculated as; CF= exp(RSE 2 /2). This correction factor is necessary when u sing log linear models (Baskerville 1972, Sprugel 1983) . Landscape Map of Agricultural and Forest Tree Crowns An image segmentation f unction of the SAGA GIS program ( http://www.saga gis.org/en/ ) was applied to th e 1.12 m lidar height image of the study site . Tree crowns were classified as forest or agricultural trees based on the number of neighbors and the amount of edge shared with neighboring tree crown s (Fig ure 3 3) . Crowns with less
57 than two neighboring crown s or less than 50% of their edge touching another crown were classified as agricultural trees because they were relatively isolated from other trees and therefore not part of a forest with continuous canopy cover. Crowns with more than two neighbors and gr eater than 50% of their edge touching another crown were classified as forest trees. Crown classification was performed in ArcMap 10.1 (Environmental Systems Research Institute; Redlands, CA). The classified crowns were used to generate a map of forest cov er and agriculture area of the study site at 1 m resolution F orest area was defined as all areas with forest tree crowns. Agricultural area was defined as all non forest area, and include d the area of agricultural tree cover . These data were used to calcul ate the density of trees and AGB per forest or agricultural area. The classified tree crowns were also used to create a 1 ha resolution map of percent tree cover, which included tree cover from both forest and agricultural trees . This map was used to analy ze patterns of AGB across a tree cover gradient. AGB Estimation from Lidar Derived Tree Crowns Using C A and H for each tree crown, the AGB lidar equation was applied to each segmented tree crown polygon to produce an individual tree based estimate of AGB ( kg). Since each tree crown was classified as an agricultural or forest tree, we calculated the contribution of each of these categories to landscape AGB. After converting each crown polygon to a point located in the center of the polygon, the AGB estimate was converted to Mg AGB at a scale of 1 h ectare . This dataset represented our individual tree based estimate of AGB (Mg ha 1 ) for the entire landscape , which could be directly compared to a published plot based estimate of AGB for the same study site (Asne r et al. 2013).
58 Calculation of the Difference Between Tree and Plot Based AGB Estimates Asner et al. (2013) estimated above ground carbon density (ACD) for all of Panama , which can be converted to AGB (ACD=0.48*AGB). The national AGB estimate was developed using field and lidar data from 6 study sites, one of which was the 22,857 region of the Azuero Peninsula. As with our approach , Asner et al. (2013) used the pantropical allometric relationship (Chave et al. 2005) to calculate field based AGB. In 33 second ary forest plots ( 0. 1 ha), they measured the diameter of every tree stem larger than 5 cm, and applied the Chave allometry to derive AGB for the measured plot . For each plot, they found the top of canopy height (TCH , the mean lidar height over the plot ), a nd fit a model to the estimated AGB from the Chave equation. This produced a model that predicted AGB for a plot using the TCH variable. AGB from this plot based method (AGB plot ) was calculated for the entire study site by extracting TCH for every hectare and applying the equation; AGB plot = 0.48(0.359* TCH) 1.7676 ( 3 3 ) The product of applying equation 3 3 to the lidar data was a plot based AGB density in Mg ha 1 for the entire study site that was compared pixel by pixel to our tree based AGB density es timate . Results Individual Tree AGB Lidar Estimates The pantropical Chave model (AGB chave ) for the field measured trees (N=1,100) resulted in a total AGB of 1,060 Mg, with a mean AGB of 478.3 kg per tree and range of 5.3 to 3,920.0 kg . The lidar model (AGB lidar ) for field measured trees resulted in a total AGB of 1,124 Mg, with a mean of 423.5 kg per tree and a range of 20.0 to 2,569.6 kg for all trees in the field data . The AGB lidar prediction was larger than AGB chave values by
59 11 % , an overestimate driven by the tendency of the model to over predict AGB of small trees ( Figure 3 4 ). AGB Estimates of Agricultural a nd Forest Trees Although tree density was much lower in agricultural areas versus forested areas (14.2 and 109.4 t rees per hectare, respectively) , t ree cover of the 22, 857 ha Azuero study site is composed of roughly equal number of forest and agricultural trees (Table 3 1). The equal distribution of trees between forest and agriculture is because the study site has much greater coverage of agricultu ral area versus forest fragments ( 20,175 ha vs. 2,681 , respectively). Similarly , although biomass density in agricultural areas was nearly 8 times lower than biomass density in forested areas ( 12.1 Mg ha 1 and 74.88 Mg ha 1 , respectively ), agricultural tre es contain ed 56,000 Mg more AGB than forest trees across the landscape . Differences i n Tree Based vs. Plot Based Estimates o f Landscape AGB Across the entire study site, the tree based estimates were an average of 28 Mg ha 1 higher than plot based estimate s, with a total difference of 183,791 Mg between the tree and plot based methods (Table 3 2) . Though differences in AGB per 1 ha plot (tree based minus plot based) ranged from 108 to 122 Mg ha 1 , most differences occurred at low AGB density (Figure 3 5a) and were small and positive, between 0 and 10 Mg ha (Figure 2 5b). The d ifferences in AGB estimates were lowest at the ends of tree cover spectrum, indicating both the tree and plot based methods are predicting nearly the same AGB in areas where there are very few trees (Figure 3 5c).Differences in percent AGB reveal that despite small differences in Mg AGB at the low end the of tree cover gradient, the tree based method estimates nearly 150 200% more AGB, and gradually declines with higher tree cover Figu re 3 5d). In terms of Mg AGB, t he greatest
60 differences between the two methods occurred at intermediated levels of tree cover, with the tree based estimate of AGB frequently h igher than the plot based estimate. These intermediate forest cover areas represe nt areas of the landscape where the tree cover is heterogeneous, characterized by individual trees among agricultural fields or pastures . Discussion Contribution of Agricultural Trees t o Landscape AGB We found that while there are roughly equal numbers of trees in forest and in agricultural areas, agricultural trees are a larger proportion of landscape AGB than forest trees. By segmenting lidar data into individual tree crowns and classifying them as either forest or agricultural trees, we found that the n umber and canopy cover of forest and agricultural trees was approximately equal (Table 3 2). However, at an average AGB density of 12.71 Mg ha 1 , agricultural trees contained 256,406 Mg , which is 64,655 Mg more AGB than forest trees in the study site. This mean s that AGB estimates of this landscape which include only data from forest ed areas miss more than 50% of landscape AGB . The high contribution of agricultural trees to landscape biomass highlights the need to accurately estimate, rather than ignore, tr ee cover in agricultural areas. It also shows the value of maintaining tree cover in highly deforested areas. In the extreme case that all agricultural trees were removed either through agricultural intensification, loss of live fences or long term failure to regenerate pasture trees, the AGB of this landscape would be cut in half. Based on the average AGB density of the forests (75 Mg ha 1 ) , and the total amoun t of AGB in agricultural trees (256 thousand Mg) the loss
61 of agricultural trees is equivalent to the removal of 3,413 ha of forest from this landscape , which is 732 more hectares of forest than exists in the study site. Tree based estimates of AGB density were higher than plot based estimates, particularly for areas of intermediate forest cover. This supports findings in an South African savanna which found that dispersed tree cover is not accurately quantified with plot based estimates and is likely due to how variability in tree height is incorporated into the AGB prediction model. Colgan et al. (201 3) found a similar trend in an African savanna were plot based methods underestimated field measured AGB at the low end of the AGB range by approximately 25%, and concluded that the tree based AGB estimates had lower error than plot based estimates. The fu ndamental reason for the discrepancies between the methods is that the tree based method, AGB is predicted for each individual tree, whereas the plot based method predicts AGB using an average tree height in a 1 h ectare area. Our comparison between tree an d plot based methods from an agricultural landscape dominated by dispersed tree cover results support those by Colgan et al. (2013) , where differences in AGB predictions results large difference s in the landscape AGB estimate. AGB Lidar Model The estimate of agricultural tree cover depends of course on the definition of what is agricultural and forest land cover, which is difficult in landscapes such as the Azuero where there is both fine spatial scale heterogeneity as well as temporal variation. The debate about a forest definition is not new, but is particularly relevant and challenging as mapping of forest and non forest is now accomplished primarily through remote sensing data, where a quantitative definition of forest is required (Putz and Redford 2010) . In this study, we distinguished between forest and agricultural tree cover
62 based on the relative isolation of tree crowns, driven by the dominance of scattered trees in pastures on this landscape . The threshold in number of neighboring crowns that deline ates agricultural versus forest trees is necessarily subjective, just as with all other classifications of forest cover, in the field and from the air . We suggest that including metrics of crown isolation when available in conjunction with tree cover may p rovide an improved method for classifying forest cover . One of the primary uncertainties of this method is estimating AGB of trees with small crowns (< 20 m 2 ). We found that there was a systematic overestimation of AGB . One explanation for overestimation o f small crowns is that the lidar pixel size was too large to accurately map the crown area of small crowns . Our data has a spatial resolution of 1.12 m so each pixel has a size of 1.25 m 2 . For a 5 m 2 (4 pixel) crown, the addition of a single pixel is a 25 % change in the measured crown area. For a 100 m 2 crown (80 pixel) , one added pixel increases crown area by only a 1. 25 % . There may be a tendency in digitizing individual crowns to include a pixel even if less than 50% of the crown is perceived to be in the pixel. For large crowns, this lead s to only small overestimation. For small crowns, this would lead to a large overestimation. A second explanation for low predictability of AGB for small crowns could be that cattle and humans cause disproportionally larg e damage to the crowns of small trees compared to the crowns of large trees because of the greater access ibility of crowns to the ground . Kuyah et al. (2012b ) found that in agricultural trees in Western Kenya, browsing or other damage on smaller, accessibl e trees affected the relationship between height, diameter, and crown area, thereby weakening the allometric
63 relationship in small trees. A gricultural trees have known uses for humans , thus increasing the chance that accessible crowns are altered by human use. Including taxon specific wood density increased the accuracy of lidar AGB prediction. However, applying this model to the landscape is contingent upon knowing the species of all tree crowns in the study area. For this study, this was possible because of the work in species identification based on hyperspectral data for the same area. Given the abundance of lidar data relative to hyperspectral data, many areas will be restricted to predicting AGB without species specific metrics. This study shows that t he lidar AGB model without taxon specific wood density is comparable with the models using crown area and height alone. As integrated hyperspectral and lidar data become more common (Ghosh et al. 2014) , utilizing the l idar model with wood density will impr ove AGB estimates, but for regions with only lidar data, the AGB estimates based on liar data are still a good estimate. Although we do not present an accuracy assessment of the lidar segmentation of crown areas here, we believe that l idar metrics of crown area and maximum height are well suited to estimate AGB for trees in a tropical agricultural landscape compared to continuous forest areas . The majority of the trees on this landsc ape have few neighboring crowns, so edges of crowns can be easily detected among the background of grass. Greater number of neighboring crowns , such as conditions in a closed canopy forest, can cause errors in segmentation algorithms because crown edges are hard to detect. T ree crowns in closed canopy forests tend to be irregular in shape and overlap, making the automatic tree crown segmentation difficult based on lidar data alone. Furthermore, secondary forests, which are common i n this landscape, have trees with
64 small crowns making spectral species identification and crown delin eation difficult because there are fewer pixels for each crown. This reduces the sample size of spectral data on which to predict species. In addition, the crown segmentation algorithm relies on detecting the curvature of the crown by identifying the talle st and lowest pixels. Therefore, with small crowns comprised of few pixels, the variation in height is more subtle and the algorithm may not detect the crown shape. Future Directions i n Landscape AGB Estimation Global climate change is a widely recognized problem being driven by a myriad of causes, one of which is believed to be the trend of tropical forest loss for the expansion of agriculture (Achard et al. 20 0 4 ) . To combat th e loss of tropical forest cover , international programs, such as REDD+, are bein g developed to integrate tropical forests, and the goods and services they provide, into the global economy . However , success of these programs requires accurate , reliable , and repeatable measurements of landscape carbon stocks and sequestration rates (Gib bs et al. 2007) . While methods to estimate forest carbon stocks from remote sensing continue to be developed (Asner and Mascaro 2014) , we stress the need to consider methods that can account for high variability in tree cover. This study demonstrate s that high resolution airborne data , in which individual trees can be resolved , may be a key to accurately assessing carbon stoc ks in agricultural landscapes. In addition, the AGB estimates based on these high resolutions images, which are often only available f or limited areas; aerial data may be used to make more accurate estimates from coarser resolution images that cover larger areas.
65 Conclusion We have shown that agricultural tree cover has a non negligible amount of biomass. For our study site in a dry fore st region of Panama, the biomass of agricultural tree cover was equivalent to the biomass of forest fragments. While methods are being developed to map biomass across entire nations as for Panama (Asner et al. 2013) , we advocate that it is necessary to ens ure methods capture the biomass of non forest as well as forest areas. While plot based methods do provide wall to wall AGB estimates, they will likely underestimate AGB in a large part of the global landscape. Without an accurate estimate of AGB in non fo rest areas, there is no way to quantify the necessary carbon stocks for these payment schemes. We believe that individual based methods provide a more accurate estimate of AGB, at least in landscapes that have a mixture of forest and non forest land cover.
66 Table 3 1 . Summary statistics of land cover, tree cover, and biomass across forested and agricultural areas. Measurement Forest area Agricultural area Area (ha) 2,681 2 0 , 175 % of study site 12% 88% % of tree cover 46% 54% # of trees 293,214 285,977 trees/ha 109.4 14.2 Total AGB (Mg) 200,751 256,406 AGB density (Mg/ha) 74.88 12.71 Table 3 2. Summary of differences between tree based (this study) and plot based AGB estimates in a 22, 587 ha landscape. AGB estimate Tree based Plot based Difference Average Mg/ha 35.8 7.4 28.44 Total Mg/ha 457,151 273 , 360 18,3791 Notes: Plot based data from Asner et al. 2013 Figure 3 1 . Study site on the Azuero Peninsula. Outlined region is the 22,587 ha study site of aerial data collection by the Carnegie Ai rborne Observatory.
67 Figure 3 2 . Outline of methods for landscape AGB analysis .
68 Figure 3 3 . Map of forest and agricultural tree classification for a subset of the study area . Tree crowns were segmented from 1.12 m lidar data and classified as an agric ultural tree (red) or forest tree (black) based on the shared edge of neighboring crowns The background image is a high resolution visible color image from the CAO AToMS system (Asner et al. 2012) .
69 Figure 3 4 . Comparison of final AGB lidar predictions and AGB chave predictions for all 1,100 field measured trees Final model statistics were generated from fitting the AGBlidar model 1000 times on 10% of the field data. Distributions of model parameters are shown in Appendix C.
70 Figure 3 5 . Tree and plot base d AGB density estimates (Mg ha 1 ). ( a ) C omparison between predicted AGB using the tree based method (this study) with the plot based estimate (Asner et al. 2013) for each 1 ha pixel in 22,587 ha study site . ( b ) Histogram of differences in AGB estimates. ( c ) Difference in AGB estimates between tree and plot based methods along a gradient of tree cover , with a local polynomial regression (loess fit , red line ) . (d) Difference in percent AGB estimates along a gradient of tree cover, with loess fit.
71 CHAPTER 4 CONCLUSION: ECOLOGICAL VALUE OF TROPICAL AGRICULTURAL LANDSCAPES While global forest coverage is declining, human population s and demand for food and fuel are increasing with the recognition that agricultural production must also increase (Godfray et al. 2 010) . The debate of how to increase agricultral production without causing additional carbon emissions or species extinctions is lively , with a focus on how and where to divide the landscape for agriculture and natural ecosystems (van Asselen and Verburg 2 013, Gilroy et al. 2014) . Fundamental to the evaluation of land management scenarios is the way in which production and ecological value is quantified. As suggested in this study, some agricultural areas that maintain high tree cover that are composed of m ultiple species with high valued to humans, and can contribute as much carbon as the remaining forest fragments. Furthermore, accounting for species presence and biomass can be done from novel aerial data, which reduces the need for expensive and time cons uming field data collection, and increases the ability to collect data at regular time intervals.
72 APPENDIX A NDIV AND NIR SPECTRAL FILTERS We varied both NDVI ( NDVI = NIR 860nm RED 650nm / NIR 860nm + RED 650nm ) and NIR (reflectance at 860 nm) and measure d overall accuracy for classification of 44 species . NIR has a bigger effect on overall accuracy relative to NDVI. The optimal NIR threshold of 30 % reflectance was chosen because above this level , accuracy dramatically declined. The optimal NDVI threshold selected at 0.50. Figure A 1. Overall accuracy for variable NDVI and NIR thresholds.
73 APPENDIX B COMPARISON BETWEEN FIELD AND LIDAR MEASURED TREE HEIGHT Figure B 1. Comparison of maximum tree crown height measurements . Field measurements were mad e in the field using a laser altimeter, and lidar measured height from a 1.12 m resolution lidar image.
74 APPENDIX C SUMMARY OF TESTED LIDAR AGB MODELS Table C 1. AGB lidar models and their model fit statistics for 1,100 field measured trees. Model Parame ters df AIC RSE A dj . R 2 fit1 CA+H 4 2162.14 57.33 0.65 0.64 fit2 CA 3 2597.34 492.53 0.79 0.47 fit2r CR 3 2597.34 492.53 0.79 0.47 fit3 H 3 2566.92 462.11 0.78 0.42 fit4 CA+CA 2 +H 5 2161.98 57.17 0.64 0.64 fit5 CA 2 +H 4 2176.53 71.72 0.65 0.64 f it6 CA+H+WD 4 2104.81 0.00 0.63 0.66 Notes: Model parameters are crown area (CA), height (H), and wood density (WD). Model (CR). df is the degrees of freedom for the model, AIC is Akai ke Informati on lowest AIC, RSE is the Residual Standard Error, Adj. R 2 is the coefficient of determination , adjusted for the number of model parameters.
75 LIST OF REFERENCES Achard, F., H. D. Eva, P. Mayaux, H. J. Stibig, and A. Belward. 2004. Improved estimates of net carbon emissions from land cover change in the tropics for the 1990s. Global Biogeochemical Cycles 18: GB2008 . Alonzo, M . , B. Bookhagen, and D. Roberts. 2014. Urban tree spe cies mapping using hyperspectral and lidar data fusion. Remote Sensing of Environment 148: 70 83. Asner, G. P. 2008. Hyperspectral remote sensing of canopy chemistry, physiology, and biodiversity in tropical rainforests. Pages 261 296 in M. Kalacska and G. A. Sanzhez Azofeifa, editors. Hyperspectral Remote Sensing of Tropical and Sub Tropical Forests. First edition. CRC Press, Boca Raton, FL. Asner, G. P., and J. Mascaro. 2014. Mapping tropical forest carbon: Calibrating plot estimates to a simple LiDAR met ric. Remote Sensing of Environment 140:614 624. Asner, G. P., D. E. Knapp, J. Boardman, R. O. Green, T. Kennedy Bowdoin, M. Eastwood, R. E. Martin, C. Anderson, and C. B. Field. 2012. Carnegie Airborne Observatory 2: Increasing science data dimensionality via high fidelity multi sensor fusion. Remote Sensing of Environment 124:454 465. Asner, G. P., J. Mascaro, C. Anderson, D. E. Knapp, R. E. Martin, T. Kennedy Bowdoin, M. van Breugel, S. Davies, J. S. Hall, H. C. Muller Landau, C. Potvin, W. Sousa, J. Wrig ht, and E. Bermingham. 2013. High fidelity national carbon mapping for resource management and REDD+. Carbon B alance and M anagement 8: 1 14 . v an Asselen, S., and P. H. Verburg. 2013. Land cover change or land use intensification: simulating land system chan ge with a global scale land change model. Global C hange B iology 19: 3648 3667 . Baldeck, C. A., and G. P. Asner. 2014. Improving remote species identification through efficient training data collection. Remote Sensing 6 :2682 2698. Baskerville, G. L. 1972. U se of l ogarithmic regression in the estimation of plant biomass . Canadian Journal of Forest Research 2:49 53. Bohlman, S. 2008. Hyperspectral remote sensing of exposed wood and deciduous trees in seasonal tropical forests . Pages 177 192 in Kalacksa, M. and G. SÃ¡nchez Azofeifa, editors. Hyperspectral Remote Sensing of Tropical and Sub Tropical Forests. CRC Press, Boca Raton, FL, USA. Carberry, P. S., W. Liang, S. Twomlow, D. P. Holzworth, J. P. Dimes, T. McClelland, N. I. Huth, F. Chen, Z. Hochman, and B. A. Keating. 2013. Scope for improved eco efficiency varies among diverse cropping systems. Proceedings of the National Academy of Sciences of the United States of America 110:8381 838 6.
76 Castro Esau, K. L, and M. Kalacska. 2008. Tropical dry forest phenology and discrimination of tropical tree species using hyperspectral data. Pages 1 26 in Kalacksa, M. and G. SÃ¡nchez Azofeifa, editors. Hyperspectral Remote Sensing of Tropical and Sub Tropical Forests . CRC Press, Boca Raton, FL, USA. Chave, J., C. Andalo, S. B rown, M. A. Cairns, J. Q. Chambers, D. Eamus, H. FÃ¶lster, F. Fromard, N. Higuchi, T. Kira, J. P. Lescure, B. W. Nelson, H. Ogawa, H. Puig, B. RiÃ©ra, and T. Yamakura. 2005. Tree allometry and improved estimation of carbon stocks and balance in tropical fore sts. Oecologia 145:87 99. Chave, J., R. Condit, H. C. Muller Landau, S. C. Thomas, P. S. Ashton, S. Bunyavejchewin, L. L. Co, H. S. Dattaraja, S. J. Davies, S. Esufali, C. E. N. Ewango, K. J. Feeley, R. B. Foster, N. Gunatilleke, S. Gunatilleke, P. Hall, T . B. Hart, C. HernÃ¡ndez, S. P. Hubbell, A. Itoh, S. Kiratiprayoon, J. V Lafrankie, S. Loo de Lao, J. R. Makana, M. N. S. Noor, A. R. Kassim, C. Samper, R. Sukumar, H. S. Suresh, S. Tan, J. Thompson, M. D. C. Tongco, R. Valencia, M. Vallejo, G. Villa, T. Ya makura, J. K. Zimmerman, and E. C. Losos. 2008. Assessing evidence for a pervasive alteration in tropical tree communities. PLoS biology 6:e45. Chave, J., H. C. Muller Landau, T. R. Baker, T. A. Easdale, H. ter Steege, and C. O. Webb. 2006. Regional and ph ylogenetic variation of wood density across 2456 Neotropical tree species. Ecological Applications 16:2356 23 67. Chave, J., M. RÃ©jou MÃ©chain, A. BÃºrquez, E. Chidumayo, M. S. Colgan, W. B. Delitti, A. Duque, T. Eid, P. M. Fearnside, R. C. Goodman, M. Henry, A. MartÃnez YrÃzar, W. a Mugasha, H. C. Muller Landau, M. Mencuccini, B. W. Nelson, A. Ngomanda, E. M. Nogueira, E. Ortiz Malavassi, R. PÃ©lissier, P. Ploton, C. M. Ryan, J. G. Saldarriaga, and G. Vieilledent. 2014. Improved allometric models to estimate t he aboveground biomass of tropic al trees. Global Change Biology DOI: 10.1111/gcb.12629 Chazdon, R.L. 2003. Tropical forest recovery: legacies of human impact and natural disturbances. Perspectives in Plant Ecology, Evolution and Systematics 6: 51 71. Chazdo n, R. L., C. A. Harvey, O. Komar, D. M. Griffith, B. G. Ferguson, M. MartÃnez Ramos, H. Morales, R. Nigh, L. Soto Pinto, M. van Breugel, and S. M. Philpott. 2009. Beyond reserves: a research agenda for conserving biodiversity in human modified tropical lan dscapes. Biotropica 41:142 153. Cho, M. A., R. Mathieu, G. P. Asner, L. Naidoo, J. Van Aardt, A. Ramoelo, P. Debba, K. Wessels, R. Main, I. P. J. Smit, and B. Erasmus. 2012. Mapping tree species composition in South African savannas using an integrated air borne spectral and LiDAR system. Remote Sensing of Environment 125:214 226. Clark, M. L. 2011. Identification of canopy species in tropical forests using hyperspectral data. Pages 423 445 in P. S. Thenkabail and J. G. Lyon, editors. Hyperspectral Remote Sen sing of Vegetation . CRC Press, Boca Raton, FL , USA .
77 Clark, M. L., and D. A. Roberts. 2012. Species level differences in hyperspectral metrics among tropical rainforest trees as determined by a tree based classifier . Remote Sensing 4:1820 1855. Clark, M. L. , D. A. Roberts, and D. B. Clark. 2005. Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sensing of Environment 96:375 398. Colgan, M. S., and G. P. Asner. 2014. Coexistence and environmental filtering of sp ecies specific biomass in an African savanna. Ecology 95: 1579 1590. Colgan, M. S., G. P. Asner, and T. Swemmer. 2013. Harvesting tree biomass at the stand level to assess the accuracy of field and airborne biomass estimation in savannas. Ecological Applic ations 23:1170 1184. C olgan, M. S., C. A. Baldeck, J. B. FÃ©ret, and G. P. Asner. 2012. Mapping savanna tree species at ecosystem scales using support vector machine classification and BRDF correction on airborne hyperspectral and LiDAR data. Remote Sensing 4:3462 3480. Condit, R., and P. S. Ashton, P. Baker, S. Bunyavejchewin, S. Gunatilleke, N. Gunatilleke, S.P.Hubbell, R.B. Foster, A. Itoh, J.V. LaFrankie, H. Seng Lee, E. Losos, N. Manokaran, R. Sukumar, and T. Yamakura. 2000. Spatial patterns in the dist ribution of tro pical tree species. Science 26:1414 1418. Congalton, R. G. 1991. A review of assessing the accuracy of classifications of remotely sensed data. Re mote Sensing of Environment 37: 35 46. Corbin, J. D., and K. D. Holl. 2012. Applied nucleation a s a forest restoration strategy. Forest Ecology and Management 265:37 46. DeFrie s, R., and C. Rosenzweig. 2010. Toward a whole landscape approach for sustainable land use in the tropics. Proceedings of the National Academy of Sciences of the United States of America 107:19627 196 32. Food and Agriculture Organization of the United Nations , FAOSTAT database . 2014. http://faostat3.fao.org/faostat gateway/go/to/download/R/RL/E (accessed, June 2014) FÃ©ret, J. B., and G. P. Asner. 2012 a . Tree species discriminati on in tropical forests using airborne imaging spectroscopy. IEEE Transactions on Geoscience and Remote Sensing 99 :1 12. FÃ©ret, J. B., and G. P. Asner. 2012 b . Semi supervised methods to identify individual crowns of lowland tropical canopy species using ima ging spectroscopy and lidar . Remote Sensing 4: 2457 2476.
78 Fischer, J., A. Zerger, P. Gibbons, J. Stott, and B. S. Law. 2010. Tree decline and the future of Australian farmland biodiversity. Proceedings of the National Academy of Sciences of the United Stat es of America 107:19597 19602. Foley, J. A., R. Defries, G. P. Asner, C. Barford, G. Bonan, S. R. Carpenter, F. S. Chapin, M. T. Coe, G. C. Daily, H. K. Gibbs, J. H. Helkowski, T. Holloway, E. a Howard, C. J. Kucharik, C. Monfreda, J. a Patz, I. C. Prentic e, N. Ramankutty, and P. K. Snyder. 2005. Global consequences of land use. Science 309:570 57 4 d e Foresta, H., E. Somarriba, A. Temu, D. Boulanger, H. Feuilly, and M. Gauthier. 2013. Towards the assessment of trees outside forests. Forest Resources Assessm ent Working Paper 183. Rome, 2013 Ghosh, A., F. E. Fassnacht, P. K. Joshi, and B. Koch. 2014. A framework for mapping tree species combining hyperspectral and LiDAR data: r ole of selected classifiers and sensor across three spatial scales. International Jo urnal of Applied Earth Observation and Geoinformation 26:49 63. Gibbons, P., D. B. Lindenmayer, J. Fischer, A. D. Manning, A. Weinberg, J. Seddon, P. Ryan, and G. Barrett. 2008. The future of scattered trees in agricultural landscapes. Conservation Biology 22:1309 1319. Gibbs, H. K., S. Brown, J. O. Niles, and J. A Foley. 2007. Monitoring and estimating tropical forest carbon stocks: making REDD a reality. Environmental Research Letters 2:045023. Gibbs, H. K., A. S. Ruesch, F. Achard, M. K. Clayton, P. Holm gren, N. Ramankutty, and J. A. Foley. 2010. Tropical forests were the primary sources of new agricultural land in the 1980s and 1990s. Proceedings of the National Academy of Sciences of the United States of America 107: 16732 16737 . Gilroy, J. J., P. Woodco ck, F. A. Edwards, C. Wheeler, C. A. Medina Uribe, T. Haugaasen, and D. P. Edwards. 2014. Optimizing carbon storage and biodiversity protection in tropical agricultural landscapes. Global Change Biology 20: 2162 2172 . Godfray, H. C. J., J. R. Beddington, I . R. Crute, L. Haddad, D. Lawrence, J. F. Muir, J. Pretty, S. Robinson, S. M. Thomas, and C. Toulmin. 2010. Food security: the challenge of feeding 9 billion people. Science 327:812 81 8. Griscom, H. P., A. B. Connelly, M. S. Ashton, M. H. Wishnie, and J. D eago . 2011. The structure and composition of a tropical dry forest landscape after land clearance; Azuero P eninsula, Panama. Journal of Sustainable Forestry:37 41. Hansen, M. C., P. V . Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend. 2013. High resolution global maps of 21st century forest cover change. Science 342:850 853.
79 Harvey, C. A., and W. A. Haber. 1999. Remnant trees and the conservation of biodiversity in Costa Rican pastures. Agroforestry Systems 44:37 68. Harvey, C. A., A. Medina, D. M. SÃ¡nchez, S. Vilchez, B. HernÃ¡ndez, J. C. Saenz, J. M. Maes, F. Casanoves, and F. L. Sinclair. 2006. Patterns of animal diver sity in different forms of tree cover in agricultural landscapes. Ecological Applications 16:1986 1999. Harvey, C. A., C. Villanueva, H. Esquivel, R. GÃ³mez, M. Ibrahim, M. Lopez, J. Martinez, D. MuÃ±oz, C. Restrepo, J. C. SaÃ©nz, J. VillacÃs, and F. L. Sincl air. 2011. Conservation value of dispersed tree cover threatened by pasture management. Forest Ecology and Management 261:1664 1674. Harvey, C. A., C. Villanueva, J. VillacÃs, M. ChacÃ³n, D. MuÃ±oz, M. LÃ³pez, M. Ibrahim, R. GÃ³mez, R. Taylor, J. Martinez, A. Navas, J. Saenz, D. SÃ¡nchez, A. Medina, S. Vilchez, B. HernÃ¡ndez, A. Perez, F. Ruiz, F. LÃ³pez, I. Lang, and F. L. Sinclair. 2005. Contribution of live fences to the ecological integrity of agricultural landscapes. Agriculture, Ecosystems & Environment 111: 200 230. Houghton, R. A. 2003. Revised estimates of the annual net flux of carbon to the atmosphere from changes in land use and land management 1850 2000. Tellus B 55:378 390. Japkowicz, N . , and S. Stephen. 2002. The class imbalance problem: a systematic study. Intelligent Data Analysis 6: 429 449. Kampe, T. U. , B. R. Johnson, M. Kuester, and M. Keller. 2010. NEON: the first continental scale ecological observatory with airborne remote sensing of vegetation canopy biochemistry and structure. Journal of App lied Remote Sensing 4:043510. Kerr, J. T., and M. Ostrovsky. 2003. From space to species: ecological applications for remote sensing. Trends in Ecology and Evolution 18: 299 305. Key, T., T.A. Warner, J.B. McGraw, and M. Fajvan. 2001. A comparison of multi spectral and multitemporal information in high spatial resolution imagery for classification of individual tree species in a temperate hardwood forest . Remote Sensing of Environment 75: 100 112. Kuyah, S., J. Dietz, C. Muthuri, R. Jamnadass, P. Mwangi, R. C oe, and H. Neufeldt. 2012 a . Allometric equations for estimating biomass in agricultural landscapes: I. Aboveground biomass. Agriculture, Ecosystems & Environment 158:216 224. Kuyah, S., C. Muthuri, R. Jamnadass, P. Mwangi, H. Neufeldt, and J. Dietz. 2012 b . Crown area allometries for estimation of aboveground tree biomass in agricultural landscapes of western Kenya. Agroforestry Systems 86:267 277.
80 Laurance, W. F. 1999. Reflections on the tropical deforestation crisis. Biological Conservation 91:109 117. La urance, W. F., J. Sayer, and K. G. Cassman. 2014. Agricultural expansion and its impacts on tropical nature. Trends in E cology & E volution 29:107 16 . LeÃ³n, M. C., and C. A. Harvey. 2006. Live fences and landscape connectivity in a N eotropical agricultural landscape. Agroforestry Systems 68:15 26. Lin, W . , and J. J. Chen. 2013. Class imbalanced classifiers for high dimensional data. Briefin gs in Bioinformatics 14: 13 26. Love, B., and D. Spaner. 2005. A survey of small scale farmers using trees in pastures in H errera P rovince, P anama . Journal of Sustainable Forestry 20:37 65. Manning, A. D., J. Fischer, and D. B. Lindenmayer. 2006. Scattered trees are keystone structures Implications for conservation. Biological Conservation 132:311 321. Maron, M. , and J.A. Fitzsimons. 2007. Agricultural intensification and loss of matrix habitat over 23 years in the West Wimmera, south eastern Australia . Biological Conservation 153: 587 593. Medina, A., C. A. Harvey, D. SÃ¡nchez Merlo, S. Vilchez, and B. HernÃ¡ndez. 2007. Bat Diversity and Movement in an Agricultural Landscape in MatiguÃ¡s, Nicaragua. Biotropica 39:120 128. Melgani, F., and L. Bruzzone. 2004. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and R emote Sensing 42:1778 1790. Morgan, J. L., S. E. Gergel, and N. C. Coops. 2010. Aerial p hotography: a r apidly e volving t ool for e cological m anagement. BioScience 60:47 59. Mueller, N. D., J. S. Gerber, M. Johnston, D. K. Ray, N. Ramankutty, and J. A. Foley . 2012. Closing yield gaps through nutrient and water management. Nature 490:254 25 7. Munteanu, C., T. Kuemmerle, M. Boltiziar, V. Butsic, U. Gimmi, D. Kaim, G. KirÃ¡ly, Ã‰. Konkoly GyurÃ³, J. Kozak, J. LieskovskÃ½, M. Mojses, D. MÃ¼ller, K. Ostafin, K. Ostapow and agricultural land change in the Carpathian region A meta analysis of long term patterns and drivers of change. Land Use Policy 38:685 697. Myer, D., E. Dimitriadou, K. Hornik , A. Weinges sel, F. Leisch, C. C. Chang, and C. C. Lin. 2013. e1071: Misc . Functions of the Department of Statistics (e1071), TU Wein.
81 it is time for a paradigm shift . Journal of Tropical Agriculture 46:1 12. Nepstad, D. C., C. Uhl, C. A. Pereira, and J. M. C. da Silva. 1996. A comparative study of tree establishment in abandoned pasture and mature forest of eastern Amazonia. OIKOS 76: 25 39. Olson, D. M., E. Dinerstein, E. D. Wikramanayake, N. D. Burgess, G. V. N. Powell, E. C. F. Allnutt, T. H. Ricketts, Y. Kura, J. F. Lamoreux, W. W. Wettengel, P. Hedao, and K. R. Kassem. 2001. Terrestrial ecoregions of the world: a new map of life on earth . BioScience 51:933 938 . Pal, M., and P. M. Mather. 2005. Support vector machines for classification in remote sensing. International Journal of Remote Sensing 26:1007 1011. Pelletier, J., N. Ramankutty, and C. Potvin. 2011. Diagnosing the uncertainty and detectability of emission reductions for REDD + under current capabilities: an example for Panama. E nvironmental Research Letters 6:024005. Plieninger, T. 2012. Monitoring directions and rates of change in trees outside forests through multi temporal analysis of map sequences. Applied Geography 32:566 576. Price, J. C. 1994 . How unique are spectral signa tures? Remote Sensing of Environment 3:181 186. Putz, F. E., and K. H. Redford. 2010. degradation, deforestation, long term phase shifts, and further transitions . Biotropica 42:10 20. R Development Core Team. 201 3 . R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. http://www.R project.org/ Ramankutty, N., A. T. Evan, C. Monfreda, and J. A. Foley. 2008. Farming the planet: 1. Geographic distri bution of global agricultural lands in the year 2000. Global Biogeochemical Cycles 22:GB1003. Rivard, B., G. A. Sanchez Azofeifa, S. Foley, and J. C. Calvo Alvarado. 2008. Species classification of tropical tree leaf reflectance and dependence on selection of spectral bands. Pages 141 160 in M. R. Kalascka and G. A. Sanchez Azofeifa, editors. Hyperspectral Remote Sensing of Tropical and Sub Tropical Forests. First edition. CRC Press, Boca Raton, FL. Sala, O. E. 2000. Global Biodiversity Scenarios for the Ye ar 2100. Science 287:1770 1774
82 SÃ¡nchez Azofeifa, G. A., K. L. Castro Esau, W. A. Kurz, and A. Joyce. 2009. Monitoring carbon stocks in the tropics and the remote sensing operational limitations: from local to regional projects. Ecological Applications 19:4 80 494. Schlawin, J. R., and R. neotropical wet forests: the legacy of remnant trees. Journal of Vegetati on Science 19: 485 492. Shao, G . , and J. Wu. 2008. On the accuracy of landscape pattern analysis using remote sensing data . Landscape Ecology 23: 505 511. Slocum, M. G., and C. C. Horvitz. 2000. Seed arrival under different genera of trees in a neotropical pasture. Plant Ecology 149:51 62. Somarriba, E. 1992. Revisiting the past: an essay on agrofore stry definition. Agroforestry Systems 19:233 240. Sprugel, D. G. 1983. Correcting for bias in log transformed allometric equations . Ecology 64:209 210 . Stevenson, J. R., N. Villoria, D. Byerlee, T. Kelley, and M. Maredia. 2013. Green Revolution research sa ved an estimated 18 to 27 million hectares from being brought into agricultural production. Proceedings of the National Academy of Sciences of the United States of America 110:8363 836 8. Stork, N. E. 2010. Re assessing current extinction rates. Biodiversit y and Conservation 19:357 371. Tilman, D., K. G. Cassman, P. A. Matson, R. Naylor, and S. Polasky. 2002. Agricultural sustainability and intensive production practices. Nature 418:671 67 7. Ustin, S. L, D. PiPietro, K. Olmstead, E. Underwood, and G. J. Sche er. 2002. Hyperspectral remote sensing for invasive species detection and mapping . IEEE International Geoscience and Remote Sensing Symposium 3:1658 1660. Williams, L. G. 2013. Putting the pieces together for good governance of REDD+: an analysis of 32 RED D+ country readiness proposals. Pages 1 20. World Resources Institute, Washington D.C. Wishnie, M.H., D. H. Dent, E. Mariscal, J. Deago, N. CedeÃ±o, D. Ibarra, R. Condit, and P. M. S. Ashton. 2007. Initial performance and reforestation potential of 24 tropi cal tree species planted across a precipitation gradient in the Republic of Panama. Forest Ecology and Management 243:39 49. Zahawi, R. A ., K. D. Holl, R. J. Cole, and J. L. Reid. 2013. Testing applied nucleation as a strategy to facilitate tropical forest recovery. Journal of Applied Ecology 50:88 96.
83 Zomer, R. J., A. Trabucco, R. Coe, and F. Place. 2009. Trees on f arm: a nalysis of global extent and geographical patterns of agroforestry. ICRAF Working Paper no. 89. World Agroforestry Center, Nairobi, Kenya . Zomer, R . J. , A. Trabucco, R. Coe, F. P lace, M. van Noordwijk, and J. Xu. 2014. Trees ecological characteristics. ICRAF Working Paper no. 179. World Agroforestry Center, Nairobi , Kenya.
84 BIOGRAPHICAL SKETCH Sarah Graves received her Bachelor of Science degree in e nvironmental s cience , p olicy , and m anagement from the University of Minnesota , where she was involved in numerous research projects which ranged from examining the carb on fluxes of arid ecosystems to quantifying the soil quality and forest species diversity of multi functional dairy farms . After graduation in 2009, s he worked in Costa Rica teaching u ndergraduate students from universities and colleges in the United State s about sustainability issues in the tropics , a topic she developed interest in while studying abroad in Costa Rica in 2008 . Sarah returned to school at the University of Wisconsin Madison in 2011 and completed a professional certificate degree in Geographic Information Science . It was after courses in remote sensing and job experience in creating land cover maps that Sarah decided to pursue a graduate degree in tropical forest conservation , which could combine her interest in remote sensing with ec ology , conservation , and sustainability . In addition to her primary thesis research in Panama, Sarah has contributed to research in Florida by authoring a paper on fire defense strategies of native oak species. After completion of her Master of Science deg ree, Sarah will continue her studies in the PhD program in the School of Forest Resources and Conservation at UF as a Graduate School Fellow. Sarah is also an advocate and supporter of women and girls involved in science and engineering fields. She leads a co authored a paper on the gender composition of editorial boards for scientific journals in the ecological sciences.