Mapping forest canopy height globally with spaceborne lidar[1-USB迷|专注于互联网分享

2024年4月12日发(作者：果郎)

JOURNALOFGEOPHYSICALRESEARCH,VOL.116,G04021,doi:10.1029/2011JG001708,2011

Mappingforestcanopyheightgloballywithspacebornelidar

MarcSimard,

NaiaraPinto,

andAlessandroBaccini

Received17March2011;revised19August2011;accepted27August2011;published19November2011.

[

]

Datafromspacebornelightdetectionandranging(lidar)opensthepossibilitytomap

entawall‐to‐wall,globalmapofcanopy

heightat1‐kmspatialresolution,using2005datafromtheGeoscienceLaserAltimeter

System(GLAS)aboardICESat(Ice,Cloud,andlandElevationSatellite).Achallengein

theuseofGLASdataforglobalvegetationstudiesisthesparsecoverageoflidarshots

(mean=121datapoints/degree

fortheL3Ccampaign).However,GLAS‐derived

canopyheight(RH100)valueswerehighlycorrelatedwithother,morespatiallydense,

ancillaryvariablesavailableglobally,whichallowedustomodelglobalRH100from

foresttype,treecover,elevation,ferencebetweenthemodel

predictedRH100andfootprintlevellidar‐derivedRH100valuesshowedthaterror

increasedinclosedbroadleavedforestssuchastheAmazon,underscoringthechallenges

inmappingtall(>40m)ultingmapwasvalidatedwithfield

eledRH100versusinsitucanopyheight

error(RMSE=6.1m,R

=0.5;or,RMSE=4.4m,R

=0.7without7outliers)is

conservativeasitalsoincludesmeasurementuncertaintyandsubpixelvariabilitywithin

the1‐ultswerecomparedagainstarecentlypublishedcanopyheight

dourvaluestobeingeneraltallerandmorestronglycorrelatedwith

revealsagloballatitudinalgradientincanopyheight,increasing

towardstheequator,aswellascoarseforestdisturbancepatterns.

Citation:Simard,M.,,,i(2011),Mappingforestcanopyheightgloballywithspaceborne

lidar,.,116,G04021,doi:10.1029/2011JG001708.

uction

[

]Forestverticalstructureremainspoorlycharacterized

despitebeingapredictorofabovegroundlivebiomass

[Lefskyetal.,2002;Drakeetal.,2002;Andersonetal.,

2006],primaryproductivity[Thomasetal.,2008],and

biodiversity[Goetzetal.,2007].Here,wemodelforest

verticalstructureusingdatafromtheGeoscienceLaser

AltimeterSystem(GLAS)aboardICESat(Ice,Cloud,and

landElevationSatellite).Thealtimetertransmitteda1024nm

lightpulseandrecordedthereflectedsignal(waveform)

[Zwallyetal.,2002].Weusethesedatatoconstructaglobal

wall‐to‐wallmapofforestcanopyheight.

[

]ICESat/GLASacquireddatagloballybetween2003

r,lidarshotsprovideanincompletecov-

otprintswereapproximately65m

indiameter,spacedby170malongtrackandseveraltensof

kilometersacrosstracks,adistancethatincreasedinthetro-

oundthisdatadearth,obstructionbyclouds

,producingawall‐to‐wallmap

JetPropulsionLaboratory,CaliforniaInstituteofTechnology,

Pasadena,California,USA.

DepartmentofGeography,UniversityofMaryland,CollegePark,

Maryland,USA.

WoodsHoleResearchCenter,Falmouth,Massachusetts,USA.

0148‐0227/11/2011JG001708

requiresexploitingtherelationshipbetweenfootprintlevel

lidar‐derivedcanopyheightestimatesandspatiallycontinu-

ousancillaryvariables,suchasdatafromtheModerate

ResolutionImagingSpectroradiometer(MODIS).

[

]Therearenumerousapproachestoassociatethesparse

lidarfootprintswiththespatiallycontinuousancillaryvari-

ionistosegmentorclassifythestudysiteto

obtainpatchesthatshareameaningfulecologicalparameter

(,speciescomposition),suchthatlidarmeasurements

,Lefskyetal.,2005a;

Boudreauetal.,2008].Arecentglobalcanopyheightmap

[Lefsky,2010]wasproducedbysegmentingMODISreflec-

ethatthose

resultscanbedifficulttointerpretatcoarseresolutiongiven

that(i)MODISimagesweresegmentedbasedonspectral/

texturalheterogeneitythresholdsthatdonotreadilytranslate

intoforeststandproperties;(ii)modelerrorcanbeattributed

tobothsub‐pixel(500‐mMODIS)andsub‐patch(1–900

MODISpixels)ternative,weproposeto

associateselectedGLASshotswith1‐kmpixelsaswe

believethatintheabsenceofhigh‐resolutionforestdistur-

bance/agemapstoserveassegments,thepixelapproach

deservesevaluation.

[

]Weproducedaglobalwall‐to‐wallcanopyheightmap

bycombiningGLASRH100estimatesandglobalancillary

peralsoextendspreviouswork[Lefsky,

2010]byvalidatingresultsagainstfieldmeasurementsand

1of12

G04021

G04021SIMARDETAL.:FORESTHEIGHTWITHSPACEBORNELIDARG04021

consideringtheimpactofsub‐pixelvariabilityonmodel

ingcanopyheightglobally,weareinter-

estedincharacterizingfine‐scalevariability(attributableto

disturbance)againstabackdropofgenerallymorecoarsely

atgoalin

mind,weselectedgloballyavailableclimate,elevation,and

cludestheMOD44Bpercent

treecoverproduct[Hansenetal.,2003]fromMODIS,ele-

vationfromtheShuttleRadarTopographyMission,SRTM

[Farretal.,2007],aswellasclimatologymapsfromthe

TropicalRainfallMeasuringMission,TRMM[Kummerow

etal.,1998]andWorldclimdatabase[Hijmansetal.,2005].

Allancillaryvariableswereresampledtobringtheresolution

oftheoutputwall‐to‐wallmapto1‐km.

[

]Inthecontextofbuildingasystematicalgorithm,the

firststepwastodevelopanobjectiveproceduretoselect

waveformsandcorrectslope‐induceddistortionsinRH100

[Lefskyetal.,2005b]andcalibratecanopyheightestimates;

RH100estimateswerecalibratedwithfieldmeasurements.

Second,weusedaregressiontreeapproachtomodellidar

y,model

predictionswereindependentlyvalidatedagainstfieldesti-

matesfrom66FLUXNETsitesdistributedgloballyand

coveringabroadrangeofforesttypes.

[

]Ourerroranalysisaccountedforthedisparityinscales

betweenlidarfootprintsandtheresolutionoftheancillary

odel,lidarshotseffectivelyrepresented

samplesoftheunderlyingforeststructure,anddidnotalways

intersectthetallesttreewithina1‐sely,

ancillaryvariablesmightnotalwaysreflecttheenvironmental

conditionsatthelidarfootprint,butrathercomprisean

averageover1‐km

.Duetothescarcityoffine‐resolution

globalvegetationmaps,wehaveexaminedtheimpactofsub

pixelvariabilityonmodelerrorindirectlybylookingattwo

surrogatesofforestheterogeneity:degreeofdisturbanceand

,wehypothesizethatourestimatesatpro-

tectedsites(asdefinedbytheUNWorldDatabaseonPro-

tectedAreas,)havelesserrorduetoless

structuralvariabilityassociatedwithanthropogenicdistur-

,wehypothesizethatacrossforesttypes,

modelerrorincreaseswithvarianceincanopyheight.

[

]Theresultingwall‐to‐wallmapcanbedownloaded

fromtheweb()andreveals

regionalcanopyheightgradientsaswellascoarsedistur-

ultswerecomparedagainstanother

canopyheightproduct[Lefsky,2010],anddifferenceswere

examinedinlightofthechoiceofcalibrationalgorithm,

ancillaryvariables,andmodelingprocedures.

ationoftheGaussiandis-

tributionsisconstrainedtobebetweenthesignalbeginning

lly,thegroundcanbedeterminedasthelast

Gaussianpeak,whichworksbestinflatareasandopen

closedcanopies,locatingthegroundis

sometimesdifficult(elastpeakhaslowampli-

tuderelativetoanotherneighboringpeak)[Boudreauetal.,

2008].Ithasbeenshownthatusingaregressionthrough

waveformextentandaterrainindexderivedfromanancil-

laryDEMcanalleviategrounddetectionissuesandimprove

canopyheightestimates[Lefskyetal.,2005b;Rosetteetal.,

2008].However,theregressionsmaybesitespecificand

therhand,Rosette

etal.[2008]obtainedreasonableresultsusingthelocationof

thelastGaussianpeakasthegroundleveland,importantly,

foundithadthelowestmeanerror(0.39m).Sincethe

regressiontreemethodologyessentiallyrepresentsoverall

trends,itisimportantthatpotentialbiasintheGLASesti-

mateofcanopyheightbeminimized.

[

]Afterthesystematicselectionprocessdescribedin

section2.2,theRH100valueswereroundedtothenearest

meterandusedasinputintheregressiontreetoproducea

wall‐to‐wallcanopyheightmap(seesection2.3).Ifmore

thanoneGLASshotintersecteda1‐kmpixel,allpoints

doflocallycombiningmultipleGLAS

shots,modelaveragingisperformedasthelaststepofthe

regressiontreeapproach(seesection2.3).

rmSelection

[

]WeselectedthedataacquiredwithlaserL3C

between2005‐05‐20and2005‐06‐mpaignwas

chosenduetoitstemporaloverlapwiththe2005MODIS

PercentTreeCoverproduct(MOD44B).

[

]Theoverallgoalofthewaveformselectionprocedure

wastoisolatedatapointsfromforestedsiteswhilereducing

theimpactofslopesandcloudcontaminationoncanopy

ctedGLASshotsthatfellwithina

forestclassasdefinedbytheGlobcovermap[Hagolleetal.,

2005].Becausethe65mGLASfootprintsamplesonlya

fractionofthe1‐km

landcoverpixel,notallwaveformsare

aseswereproblematicgiven

ourobjectivetomodelthetallestcanopiesasopposedto

rethatshotswerereflected

fromaforestcanopy,weselectedwaveformscharacterized

bymorethanoneGaussianpeak,assumingthatwaveforms

withasinglepeakareduetogroundreflectiononly.

[

]GLA14waveformswerealsofilteredusingengi-

neeringandsignalparameterstoaccountforcloudcover

14productcontainsacloud

d,we

computedtheSignal‐to‐NoiseRatio(SNR)todetect

waveformhinderedbycloudsaswellaswaveformswith

therremovewaveformswithsignaldominatedbyclouds,

weselectedwaveformsthatwereco‐locatedclosetothe

groundasdefinedbySRTM(80mtoaccountfor

forestheightandSRTMelevationerrors).

[

]Terrainslopeisthemainfactorcontributingtocanopy

heightestimationerror[Lefskyetal.,2005b;Duncanson

etal.,2010].Theimpactofslopeistobroadenthelidar

waveform,therebyintroducingabiasincanopyheight

rgefootprintlidarsuchasGLAS,pulse

HeightEstimation

[

]OuranalyseswerebasedontheGLA14landproduct

A14waveformisafitoftheoriginal

GLASwaveform,modeledbyamaximumof6Gaussian

distributions[Brenneretal.,2003].TheGLAS‐derived

estimateofcanopyheightisthewaveformmetricRH100,

definedasthedistancebetweensignalbeginningandthe

locationofthelidargroundpeak[HardingandCarabajal,

2005;Sunetal.,2008;Boudreauetal.,2008].Inthe

GLA14productversion31,thesignalbeginningisdefined

asthelocationatwhichthesignalis3.5timesabovethe

2of12