You are on page 1of 20

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 102, NO.

D10, PAGES 11,145-11,164, MAY 27, 1997

Thermal band selection for the PRISM instrument

1. Analysis of emissivity-temperature separation algorithms


Vicente Caselies, Enric Valor, C•sar Coll, and Eva Rubio
Departamentde Termodinhmica,Facultatde Fisica,Universitatde Valencia,Burjassot,Spain

Abstract. One of the missionsbeingplannedby the EuropeanSpaceAgency(ESA)


within the frameworkof its Earth ObservationProgrammeis the ProcessesResearchby
ImagingSpaceMission(PRISM). The PRISM instrumentconsistsof a thermal sensor
whosemain objectiveis to retrieveaccurateland surfacetemperatures(LST) and whose
band positionsare 3.5-4.1 /•m, 8.1-9.5/•m, 10.3-11.3/•m, and 11.5-12.5/•m. We have
studiedthe optimal designof this instrumentto retrieve accurateLSTs. First, we have
analyzedseveralemissivity-temperature separationmethods(part 1) and atmosphericand
emissivitycorrectionalgorithms(part 2). Finally,we haveidentifiedthe optimalband
configuration(part 3). This paper is the first of a seriesof three and addresses
the
questionof the emissivity-temperature separability.Among all the existingalgorithms,we
have studied the "absolute methods," which are able to estimate the absolute value of
emissivityat satellitescaleand can yield better resultsin the emissivityestimate.These
methodsare the algorithmbasedon the temperature-independent thermal infrared
spectralindex (TISI), the alphacoefficients method,and the algorithmswhichusevisible,
near-infrared,and shortwaveinfrared data to estimatethe thermal emissivity(vegetation
covermethod(VCM)). The studyconsistedof an analysisof both random and systematic
errors of each method. The resultsindicate that emissivitycan be obtainedwith an error
of _+1.7-5%usingthe alpha coefficients,_+1.7-3%usingthe TISIs, and _+0.5-1.4%using
the VCM, dependingon the spectralregion.In all casesthe error decreaseswith
wavelength,and the lowesterrors are achievedin the 10-12/•m spectralregion, due to
smallvariabilityof emissivity.It is necessarythat the two first methodsuse radiosondes
simultaneouswith the satelliteoverpassto perform the atmosphericcorrectionson the
thermal data; in addition,they showimportant sourcesof systematicerrors,whichwill
increasethe uncertaintyin the emissivityestimate(evenin the bestpossiblecase).The
VCM doesnot use radiosondesand doesnot presentimportant sourcesof systematic
error. It appearsto be the procedurewith the most favorableerror propagation
characteristics.
Thus the VCM could be the most adequatemethod for retrievingthe land
surfaceemissivity(LSE), within the frameworkof thiswork.

1. Introduction 3.5-4.1 /•m, 8.1-9.5/•m, 10.3-11.3/•m, and 11.5-12.5/•m (this


initial band positionwill be optimizedin part 3).
The EuropeanSpaceAgency(ESA) is currentlyinvestigat- The main purposeof the thermal radiometer componentis
ing severalmissionsin its Earth ObservationProgramme.One to retrieve accurate LSTs from the radiance data. However,
of these missionshas as its main objective to increase the this radiance is affectedby several effectswhich must be re-
knowledgeof land/atmosphereinteractionsat the local scalein movedbefore obtainingthe temperature:(1) the Earth's at-
order to improve our understandingof processesat the re- mosphere,which actsas an emitting and absorbingbody dis-
gionaland globalscales.A key parameterin theseprocesses is turbingthe radiation emittedby the land surfaceto the sensor;
the land surfacetemperature(LST), which is involvedin en- (2) sincenatural surfacesare not perfectemitters,their emis-
ergy balance,energy-matter-momentum transfer, evaporation sivities must be known. In addition we must account for the
and transpirationof canopies,desertificationprocesses,etc., angulareffectspresentedboth by radiometrictemperatureand
and canbe usedasan indicatorof other processes suchasland by emissivity;theseeffectsare due to the directionalnature of
degradationand climaticchange.This missionis knownas the radiance measurementsand are important on heterogeneous
ProcessesResearchby Imaging SpaceMission (PRISM) see and roughsurfaces(suchasthe land). Consequently,the num-
EuropeanSpaceAgency(ESA), [1996]for detailson thisinstru- ber of bands,their positions,and bandwidthsneed to be opti-
ment). The PRISM is a high spatialresolutionmissioncom- mized in order to correct the above mentioned effects and to
posed of a visible, near-infrared, and shortwaveinfrared obtain the most accurate LST values from thermal data. In this
(VNIR/SWIR) imagingspectrometer, sensingcontiguously at sensewe have investigatedthe optimal designof the PRISM
0.012 /•m spectral resolution from 0.45 to 2.35 p•m, and a thermal sensorunder an ESA contract addressingall the in-
thermal infrared radiometer whose initial band positionsare
volvedquestions.The main objectiveof the studyhasbeen to
Copyright1997by the AmericanGeophysicalUnion. determinean optimizedband configurationfor the recoveryof
Paper number 97JD00344. LST from high-resolutionpassivethermal infrared data.
0148-0227/97/97JD-00344509.00 To perform the studywe have divided the work into two
11,145
11,146 CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

tropical
0.8 midlatitude
summer
midlatitude
winter ..,

ii.

,'

Wavelength (1•m)

Figure 1. Atmospherictransmittance
as a functionof wavelength
for threetypicalstandardatmospheres.

main tasks.First, we have reviewedthe suitabilityand limita- 2. Factors Which Affect Emissivity
tions of the varioustechniquesto separatethe emissivityand
The atmospheremodifiesthe radiationemittedby the land
temperatureinformation and to remove the atmosphericand
surfacethroughabsorptionand emissionprocesses. Therefore
emissivityperturbationsfrom the measuredradiances.Regard-
ing the first issue,we have studiedthe algorithmbasedon the
in orderto selectseveralbandsto retrievethe surfacetemper-
ature,we haveto searchfor the spectralregionswhichpresent
temperature-independent thermal infrared spectral index
(TISI) [Beckerand Li, 1990a],the alpha coefficientsmethod the minimum atmosphericdisturbance.Figure I showsthe
[Kealyand Hook, 1993],and the algorithmswhichuseVNIR/ atmospherictransmittancein termsof wavelengthin the 3-13
SWIR data into emissivityestimations.In relation to the at- /•m region for three standardatmospheretypes:midlatitude
mosphericand emissivitycorrectionalgorithms,they fall into winter(dry),midlatitudesummer,andtropical(humid).In this
three main categories:single-channel algorithms,split-window graphwe can clearlyidentify three regionsin whichthe atmo-
algorithms,and dual-anglealgorithms.We have identifiedthe sphereis transparentenoughto allow satelliteobservationsof
generalconditionsunderwhich eachmethodologycanbe uti- the Earth surface(the so calledatmosphericwindows)with
lized, in termsof the numberof bandsrequired,their approx- transmittances larger than 0.5: the 3.4-4.2/•m (whichwill be
imate position,and number of acquisitionangles.With this referredto aschannelA), the 8.1-9.4/•m (channelB), andthe
informationwe have performed a comparisonof the several 10.0-12.5 •m (channelsC andD) intervals.We havestudied
studiedmethods.The resultsof this analysishave led us to the feasibilityof land surfaceemissivity(LSE) and LST re-
establishthe most suitableemissivity-temperature separation trieval only in theseregions.
algorithm,and atmosphericand emissivitycorrectionmethod, Emissivityis the physicalpropertywhich definesthe capa-
and the most convenientspectralregion to obtain the most bility of a body to emit radiationat a giventemperaturein
accurate LSTs. relationto the blackbodythat is the perfect emitter (so its
In the secondpart of the work we have simulatedthe recov- emissivity is 1). Naturalsurfaces do not actasblackbodies: they
ery of LST usingthe proposedalgorithmsfor a setof potential haveemissivities whichare lessthan unity, changewith wave-
band configurations, locatedwithin the spectralregionresult- length, and are differentfor each sample.To illustratethis,
ing from the previousstudy.On the basisof the resultsof these Figure2 showstypicalemissivityspectralcurvesfor three sur-
simulationswe have derived a detailed set of optimal band facetypes:vegetation,soil,and rock.Greenvegetationusually
configurations.The set of solutionsincludesestimationsof the presentsa high emissivityvalue (•0.985) and a low spectral
magnitudeof errors associated with each configuration. variability.The typical spectralsignatureof vegetationis a
This paper is the first of a seriesof three and addressesthenearlyflat curve.Thesecharacteristics are dueto the scattering
analysisperformedon the emissivity-temperature of radiationin the leavesandon the soil [Salisbury
separation andDMria,
algorithms.In the secondpaper [Caselies 1992]. On the other hand, rocksand soilshave lower emissivi-
et al., 1997a]we will
exposethe studyperformedover the atmosphericand emis- ties and presentspectralfeaturesdependingon their compo-
sivitycorrectionalgorithms.Finally,in a thirdpaper[Caselies et
sition (the reststrahlenbands).These are due to absorption
al., 1997b]we will show the simulationprocedureand the phenomenawhich occur at molecularlevel and are related to
optimalband configurationobtainedfor the PRISM instrument. energytransitionsbetweenthe allowedvibrationalmodes;con-
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,147

sequently,these characteristicsdepend on the crystallattice


structurepresentedby the surface[Asrar,1989]. For example,
the soil features in Figure 2 are due to the existenceof car-
0.98
bonatesin its composition,and the emissivityminima of the
rock are related mainly to carbonatesand alsoto quartz.They
can be modified by roughness,water content, etc., but in any \

case,these surfacesshowhigh spectralvariability. If we con- ,_,_,_,_,•


0.96
._•
sideronlythe atmosphericwindowsdefinedabove,the highest .•_
variabilityin emissivityvaluesis observedin channelA (0.75- I,U 0.94
0.98), decreasesin channelB, and is minimumin channelD
(0.95-0.98). This fact will lead to a decreasein error in the .... sea
retrievedemissivitywith wavelength,channelsC and D being 0.92

the most suitableto obtain this physicalquantity. • vegetation


Besidesthe spectralvariation of emissivitywe also need to
determineits angularproperties.At present,most of the mod- 0.9

els assume the Lambertian behavior of the surfaces, in which 0 10 20 30 40 50 6½

surfaceshave equal emissionin all directions. However, an Angle (deg)


angular variation has been observedempirically in homoge-
Figure 3. Emissivityangular variation for three typical sur-
neoussurfaces.Figure 3 showsthe emissivityvariationwith the
faces:sea [Masudaet al., 1988], soil [Labedand Stoll, 1991],
viewingangle for three surfaces:soil, sea, and vegetation.Soil and vegetation[Norman et al., 1994].
and seapresenta decreasein emissivitywith increasingviewing
angles,whereasthe vegetationshowsan emissivitythat is al-
most constant. Nevertheless, some theoretical calculations correspondingto the radianceemitted by the surfaceat a given
made with radiative transfer models show that emissivitycan wavelength.Lack of this correctioncan introducelarge errors
present directional features in canopies,depending on the in temperature. This effect is partially compensatedby the
leaves orientation and on the cover percentage[Anton and reflection of the downwellingatmosphericradiance on the
Ross, 1990; Olioso, 1995]. This variation must be taken into surface,the compensationbeing larger in more humid atmo-
accountin the LST algorithms,the main problem being the spheres.Generally, both effectsdo not canceleach other out.
lack of emissivityangular measurements.The question be- Channel A shows a variation in the correction between 0.5 and
comes more difficult when we consider real land surfaces, 6 K. Channel B presentsthe maximum emissivitycorrection
which are usuallya mixture of differentelementsand may have andvariability(1-11 K), and channelsC (1-3 K) andD (1-2.5
a complexstructure.In this casewe need to establisheffective K) have the minimum ones.
parametersto characterizethe surface[Caseliesand Sobrino,
1989]. The way in which these facts influence the effective
emissivityis still undetermined, due in part to the lack of 3. Selection of Emissivity-Temperature
adequateemissivitymeasurements.More work is needed in Separation Algorithms
this direction. In relation to temperature retrieval we encountertwo main
To quantifythe effect of the surfaceemissivityon tempera- problems. First, assumingthat emissivityis known, we must
ture, Figures 4a, 4b, and 4c showthe emissivitycorrectionin definethe way to incorporateits correctionin any algorithmto
each atmosphericwindow, defined as the differencebetween retrieve the LST from space measurements.Several efforts
the real surfacetemperature and the blackbodytemperature havebeen made on this question[Beckerand Li, 1990b;Prata,
1993; Collet al., 1994a],which will be treated in detail in the
secondpaper of this series[Caselieset al., 1997a].The second
problem is that in the radiance measuredby a sensor(one
measurement)the emissivityinformationand temperaturein-
formation(whichare two independentvariables)are coupled.
Thus the systemof equationsis alwaysunderdetermined(the
so-called"missingequation"problem). Therefore someaddi-
0.9 tional hypothesesto separatethe emissivityand temperature
information in the measured radiances are needed.
Severalmethodshave been proposedin order to obtain the
LSE from space(seeTable 1). At the beginningsomeof them
0.8
aimed to enhanceemissivityinformation based on image pro-
cessingtechniques,suchasthe decorrelationstretch[Sohaand
• Rock
Schwartz,1978; Gillespieet al., 1986]. This techniquedisplays
Vegetation emissivitychangesand temperature variations by means of
....... Soil
different colorsand different intensities,respectively.The pro-
0.7
cessed
datastillc0h:tain
spectral
andthermal
information
and
cannot be compared with laboratory measurements.So this
Wavelength (!•m)
approachis not valid for quantitativestudies(the absolute
Figure 2. Emissivityspectralcurvesfor three typical natural value is not obtained)nor for multitemporalones(the images
surfaces:rock (sandstone),soil (aridisol), and vegetation are not comparable).This approachis onlyvalid in qualitative
(white pine) [Salisbury
and DMria, 1992, 1994]. studies(classification
and recognitionpurposes,for instance).
11,148 CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

et al. [1980]. It assumesa constantand definedvalue of emis-


.... Soil sivityfor all the pixelsof a givenchannel.The spectralband to
• Rock be selected is determined by laboratory measurements:the
........ Vegetation channel in which several sampleshave the same emissivity,
beingthis closeto the assumedvalue,is the channelto be used.
Sinceits emissivityis known,we can retrieve the pixel temper-
ature and use it to obtain the emissivityin the remaining
channels.In thisway the obtainedquantitiesare channelemis-
sivitiesrelative to the reference channel.This method is very
simple in concept and application,but it has severallimita-
• 4 tions, as pointed out by Hook et al. [1992]. First, since we
assumea constantemissivityin a channel, all spectralinfor-
mation contained in it is lost. Second, the noise of the refer-
ence channel is introduced and amplified in the other ones.
4.2
Moreover, an incorrectassumptionin the emissivityvalue will
introduceerrorsin the retrievedtemperatures;so,the method
Wavelength (!•m)
is difficultto applyin thoseareaswherevegetationis distrib-
uted in varyingproportions,becausethe assumedhypothesis
b will not be suitablein thosecases.Consequently,the error in
emissivityand temperaturewill varydependingon the assumed
valuesin the referencechanneland the heterogeneityof the
10 area.As an example,Kealyand Hook [1993]reportederrorsof
about +_2K in temperature that have an equivalenterror on
emissivityof about4-5%, corresponding to the thermal infra-
red multispectral scanner(TIMS) andthe advancedspaceborne
thermal emissionreflectanceradiometer(ASTER) sensors.
2. A very similar method to the reference channel is the
emissivitynormalizationmethoddescribedby Gillespie[1985].
Soil
We assumea certain value of emissivityfor all the channels
and then calculate temperaturesfrom them. The highest of
Vegetation
thesetemperaturesis the closestto the actual temperatureof
each pixel and is assumedto be the correctvalue. Then, it is
8 8'.3 8'.6 8'.9 9'.2 9.5 usedto calculatethe emissivityin the remainingchannels.This
Wavelength (!• rn) method is also very simple to apply.Kealy and Hook [1993]
reported better resultsin this method than for the reference
channel,givingan accuracyof about2-3%. This methodshows
.... Soil
basicallythe sameadvantagesand disadvantages alreadyseen
12 • Rock
Vegetation
in the preceding method.
3. The spectralratio method [Watson,1992a] providesa
10
meansto obtain information about the spectralvariationson
._
emissivity.In this method the ratio of the radiancesmeasured
g 8
in two contiguouschannelsis related to the ratio of the emis-
sivitiescorresponding to the samebands.A roughestimateof
• 6
the real temperatureis needed to obtain the ratio, and its
.•_ accuracydepends on this estimate and on how narrow the
bands are. Watson [1992a] has reported that an estimate of
temperaturewith an accuracyof about 13 K permitsto obtain
the ratio with an accuracyof about 1% for the TIMS. Thuswe
can usethe maximumbrightnesstemperatureamongchannels
0

10 16.5 1'1 `1i.5 '12 '12.5 '13 as the temperature estimate,becausethe differencewith the
actual temperaturefalls within 13 K in many cases.The main
Wave length (!• rn)
limitation is that it is not possibleto retrieve the absolutevalue
Figure 4. Emissivitycorrection in terms of wavelengthfor of emissivity,only relative variations.Moreover, the method is
the sametypicalsurfacesasFigure2: (a) 3.4-4.2/xm, (b) 8-9.5 limited by the signal-to-noiseratio and the spectralbandwidth
/xm, and (c) 10-12.5/xm. of the used channels,due to the dependenceof the ratio
accuracyon these factors.
4. Another method which estimatesthe emissivitydiffer-
Other methodologiespermit to handle quantitatively the ences(As) betweentwo contiguous channelsis that presented
obtained results.Two main typesof algorithmscan be distin- by Collet al. [1994b].The methodis basedon the separation
guished:"relativemethods"and "absolutemethods."With the between the atmosphericand emissivityeffectsin the bright-
first type, we can obtain information about the spectralvaria- nesstemperaturedifferencemeasuredin these bands.To do
tions in emissivity: this, it is necessaryto use atmosphericprofiles simultaneous
1. The referencechannelmethodwas developedby Kahle with the satellite overpassand a radiative transfer model to
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,149

Table 1. Summaryof Emissivity-TemperatureSeparationAlgorithms


Method Advantages Disadvantages Error
Decorrelation stretch only qualitative
Reference channel simplein conceptand application lossof informationin referencechannel 4-5% with TIMS and ASTER
noise amplification Kealyand Hook [1993]
difficultapplicationin vegetatedareas
Emissivitynormalization sameas previous same as previous 2-3% with TIMS and ASTER
Kealyand Hook [1993]
Spectralratio only a rough estimateof no absolutevalue, only spectral 1% in spectralvariationwith TIMS
temperatureneeded variations Watson[1992a]
sensitivityto S/N ratio and bandwidth
Emissivitydifference use of radiosondes simultaneous with 0.5% in spectralvariation on
the satelliteoverpass AVHRR channels 4 and 5
accuracyof atmosphericcorrection Collet al. [1994b]
only spectralvariations
Multitemporal no previousspectralinformation hypothesisof invariabilityof emissivity 12% with TIMS
needed coregistration Watson[1992b]
sensitivityto S/N ratio
TISI no previousinformation needed coregistration 3% on AVHRR channels 4 and 5
about surfaceemissivity knowledgeof angular form factor and 5% on AVHRR band 3
temperatureindependent solar irradiance Li and Becker[1993]
Alpha residuals temperatureindependent knowledgeof mean emissivitiesof 4% with TIMS and ASTER
surfaces Kealyand Hook [1993]
Empirical relationships only applicableto areaswhere 0.7%
(e versusNDVI) measurements are made Van de Griend and Owe [1993]
difficultto extrapolateto satellitescales
Radiative transfer codes developedat microscopicscale
knowledgeof surfacecharacteristics
Geometric models developedat samesensorspatial knowledgeof surfacecharacteristics 0.5-2% in 10-12 tzm region
and spectralresolution Valor and Caselles[1996]

TIMS, thermalinfraredmultispectralscanner;ASTER, advancedspacebornethermalemissionreflectanceradiometer;TISI, temperature-


independentthermalinfraredspectralindex;AVHRR, advancedveryhighresolutionradiometer;NDVI, normalizeddifference
vegetationindex.
S/N, signal to noise.

apply the atmosphericcorrectionto the brightnesstempera- well studied in vegetation, and not valid when there are
tures.This methodhasbeen appliedto the advancedvery high changesin the surfaceconditions(moisturecontent,composi-
resolutionradiometer (AVHRR) channels4 and 5, and the tion,vegetationcoverage).Other problemsare more technical,
error analysisshowsan accuracybetter than 0.005 in the Ae suchas the coregistrationof the two imagesand the sensitivity
estimation[Call et al., 1994b].The main problemsare the need of the method to the signal-to-noiseratio. Related to this last
for radio soundingsand the accuracyrequired in atmospheric point, Watson[1992b]hasreported that with TIMS data, emis-
correction.This is inconvenientin moist atmospheres.How- sivitycanbe only retrievedwith an accuracyof about 12%, and
ever,the impactof atmosphericcorrectionis muchlowerin dry Stall [1991]hasobservedthat if the noiseincreases, the system
atmospheres,where the atmosphericprofiles can be avoided, of equationsbecomesdivergent,and no solutionis obtained.
as shownby Call et al. [1994a]. 6. Other methodsuse spectralindicesmore or less inde-
All thesemethodsproduceonly relative variationsof emis- pendent of temperature. An example is the temperature-
sivitybetweenbands.This solvespart of the problem. How- independentthermalinfraredspectralindex(TISI) developed
ever,whatwe needto performthe correctioncompletelyis the by Beckerand Li [1990a].TISIs are constructedby combining
absolutevalue of emissivity,not only its spectral shape. To by-productradiancesor temperaturescorrectedfor the atmo-
achieve this, we can consider some "absolute methods": spheric effect and with adequate exponentsfor each term.
5. The multitemporal algorithm [Watson, 1992b; Stall, They are related to certain combinationsof the emissivities
1991] allowsto determine the emissivityif the surfaceis ob- correspondingto the samechannelsas the involvedradiances
servedat two differenttemperatures.The methodhypothesis is (see section4.2). It is possibleto use them for classification
that emissivitydoesnot changeduringthe period in whichthe purposes,definingparticularTISIs sensitiveto particularfea-
measurements are made. These are taken at two different tures of the studied surfaces. Furthermore, Li and Becker
times,at which emissivitiesare the same,but temperaturesare [1993]defineda way to retrieveboth the temperatureand the
not. With this procedurewe obtain 2N measurements (in a absolutevalue of emissivityusingTISIs. Emissivitiesare ob-
sensorwith N channels),and we haveN + 2 unknowns(N tainedthrougha daytimeand a nighttimeimagecorresponding
emissivitiesand two temperatures);if N > 1, then the system to channels3, 4, and 5 of AVHRR and assumingthe hypoth-
is theoreticallysolvable.The measurementsshouldbe taken at esisthat TISI does not changebetween day and night. How-
midday and at midnight to obtain a maximum differencebe- ever, we have to considerseveralquestionswhich could limit
tween the two measured temperatures.As pointed out by the applicationof the method.First, the problemof the coreg-
Watson[1992b],the main advantageof this method lies in the istrationof the two imagesneeded(in fact, thismethodologyis
fact that previousinformationaboutthe spectralshapeof the a multitemporalmethod), becausea misregistrationcan pro-
emissivitycurve is not needed. However, it has severallimita- duce large errors dependingon the heterogeneityof the sur-
tions,one of them beingthe assumedtemporal invariabilityof face. Second,the method needsto considerthe angularform
emissivity.This hypothesisis correctin dry soilsand rocks,not factor of the surface,which relates the bidirectional reflectance
11,150 CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

effectiveparametersof a surfacein terms of a vegetationpa-


rameter and which are derivedat the sensorspatialresolution,
as proposedby Valor and Caselles[1996] who estimated an
error range of 0.5-2% in the emissivityretrieval within the
10-12 •m spectralregion.
Consideringonly the algorithmswhich allow to estimate
Top absoluteemissivities,it becomesclear from the previousdis-
cussionthat only the spectral-indices-based
methodsand the
VNIR/SWIR methods can retrieve LSE with a reasonable de-
gree of error (the multitemporalalgorithmshowsan excessive
level of error). Regardingthe VNIR/SWIR algorithms,the
Side empirical approach(i.e., to obtain emissivity-vegetation pa-
rameterrelationships using"in situ"data) is somewhatlimited,
becausethese relationshipsdepend, to a great extent, on the
surfacecharacteristics (emissivityvalues,vegetationdistribu-
tion and structure,etc). Thus it is not possiblein principleto
Ground use empirical relationships,obtained in a particular area, in
other areas completely different. An additional restriction is
Figure 5. Scheme of direct radiances and first reflections that field measurementsare difficultto extrapolateto the sat-
which occurinsidethe cavitiesexistingbetweentwo contiguous ellite scale if a model of the surface is not used. On the other
vegetationelements[Caseliesand Sobrino,1989].
hand, radiativetransfermodelscanbe very usefulto studyhow
emissivityvariesdependingon severalfactors,but thesemeth-
ods are generallydevelopedat microscopicscaleand thus are
in channel3 to its emissivity,and also the solar irradianceon difficult to employ at the usual satellite spatial resolutions.
the surface.Li and Becker [1993] studiedthe main sourcesof Consequently,it seemsthat the most suitablemethod of this
error and their impact on emissivitycalculationand concluded categoryto be appliedis the one basedon the combineduseof
that emissivitycould be obtainedwith an accuracylower than geometricmodelsand VNIR/SWIR data. For the reasonsex-
3% in channels 4 and 5 and lower than 5% in channel 3. plained above,we have analyzedthree methodsin this work:
7. The alpha residualsmethod also usesthe definition of the TISI method, the alpha coefficientsalgorithms,and the
an index [Kealyand Hook, 1993;Hook et al., 1992].This index VNIR/SWIR geometricmodel method.
is derivedusingWien's approximation.It is calculatedfor each
channelby meansof a combinationof the radiancesmeasured
by the sensorin all bandsand is related to the emissivity.This 4. Description of Selected Algorithms
index is alsotemperatureindependent.The alpha spectrumof 4.1. Algorithms Incorporating VNIR/SWIR Data
the surfacesis usedfor recognitionand classificationpurposes. Into Emissivity Estimations
Kealy and Hook [1993] have proposeda way to retrieve the In order to do an appropriate analysisof such types of
absolute value of emissivity from the alpha indices. This methodologies,we need to relate theoreticallythe data taken
method needsto considerthe mean emissivityof the surfaces in the VNIR/SWIR and thermal infrared regions.We believe
in the spectralregion coveredby the sensor.Kealy and Hook that the onlyway to do this is throughsomevegetationparam-
[1993] proposeto use an empirical relationshipbetweenthe eters, such as the vegetationcover of the surface.There are
varianceof alpha residualsand the mean value of emissivityof severalmethodsto obtainthe vegetationcoverusingdata from
several surfacesin order to estimate this mean value. Using the VNIR/SWIR region (vegetationindices,for instance).
simulateddata, they havereported an accuracyof 4% in emis- Thus the questionlies in relating this parameter to emissivity.
sivityretrieved with TIMS and ASTER sensors. To this end we have chosenthe geometricmodel proposed
Apart from these two groupsof methodologieswhich use by Casellesand Sobrino [1989], following Valor and Caselles
radiancesmeasured in the thermal infrared region, we can [1996]. It considersthat the land surfaceis a heterogeneous
considera third type of methodwhichtriesto relate data taken (both in temperatureand in emissivity)and roughsystemcon-
on different parts of the spectrum:optical and microwaves stituted by a soil on which the vegetationis distributedin a
domains [Beckerand Choudhury,1988], optical and thermal more or lesscompactway. Each vegetationelement is repre-
bands [Li and Becker, 1990], or optical, thermal, and micro- sentedby means of a "box" made up by Lambertian surfaces
wave domains[Li and Becker,1990]. Some authorshave pro- and with length L, height H, and separationS. The model
posedto use relationshipsbetweenemissivityand somespec- does not take into account the shadow influence, which was
tral indicescorresponding to other partsof the spectrum,such consideredby Caselleset al. [1992], the systemhaving only
as the normalizeddifferencevegetationindex (NDVI). One three distinguishable elements:top, side,and ground(seeFig-
way to obtain them is empirically, i.e., taking simultaneous ure 5). It considersthe radianceemitted by the systemas a
measurementsof emissivityand NDVI and calculatingregres- weighted combinationof the radiancesemitted by each ele-
sion curvesbetween the measurements.An example of this is ment. If we neglect double-scatteringprocessesbetween the
givenby van de Griendand Owe [1993],who reportedan error differentpartsof the systemand simplifythe geometryaccord-
of estimate of 0.7% in emissivity.Another way is to use radi- ing to the infinitelylongLambertianboxesmodel suggested by
ative transfer codes which model the processessufferedby Sutherlandand Bartholic[1977],thenwe achievea definitionof
radiation in a given surfaceat microscopicscale [Anton and the effectiveemissivityof sucha systemgiven by
Ross,1990;Olioso,1995].Finally, we can alsocombineVNIR/
SWIR data with geometricmodels,which define the thermal s, = S,o+ ds, (1)
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,151

where eio is a term related to the radiancesdirectlyemitted by Table 2. Coefficients/• and n• Used in the Power
the elementswhich constitutethe surfaceand is given by Approximation Correspondingto PRISM Channelsand
Calculatedfor the Temperature Range [285, 325] K
I•io-- I•tvPvnt- I•t•7(1- P v) (2)
Band /3i, W/cm2sr cm-1 n, AT,, K
whereeiv andew are the vegetation
andgroundemissivities A 3.7306 X 10-39 12.597 0.2
measured in band i, respectively,and P• is the vegetation B 1.4440 X 10-•9 5.380 0.2
cover in the p•el. The factor d e• is related to the radiance C 1.0878 X 10--16 4.445 0.2
indirectlyemitted by the surfacethrough the reflections(to D 1.0630 X 10-•s 4.067 0.2
first order) that existinsidethe cavitiesdefinedin the rough
PRISM, ProcessesResearchby Imaging SpaceMission.
surface, because of the nonblacknessof the natural surfaces
(caviff effect).It is givenby [Caseliesand Sobrino,1989]

de,: (1 - e,a)e,•(1 - P0 N

----
•-' •'l aknk
TISI ]-[ (1/o (7)
k=l

wherePs is the sideproportionobse•ed by the sensor,and F,


whereT} is the ground-level
brightness
surfacetemperature
G, and F' are shape factors that represent the fraction of
measuredin channel k and obtained using a single-channel
radiation which, emitted by a side, reachesthe ground; the
algorithm.
If we useground-level
radiances
(Bk(T•)) instead
proportionwhich, emitted by the ground,reachesa side; and
of temperatures,then
the fraction which arrNes to a side comingfrom the other one,
respectNely(seeFigure5). Theseshapefactorsdependon the N

surfacestructureand are gNen by [Caselies and Sobhno,1989] TISI = M I-I (mk(Y•))


a•' (8)
k=l

F:(1+•)- •1+(•)2 (4, where M is a constantwhich dependson the channel combi-


nation and is given by
N
1+ - 1+
m--- H j•;ak (9)
k=l

F' = I + H (6) Additionally, the condition


N
Equations(1)-(3) relate the effectiveemissiviffto a vegeta-
tion parameter,whichis the vegetationcover(hereinafterthis Z aknk
-- 0 (10)
k=l
method will be referred to as the "vegetationcover method"
(VCM)). It canbe obtainedfrom VNIR/SWlR data usingone mustbe alwayssatisfied.In (10) it is possibleto chooseN - 1
of the severalexistingtechniques;for example,we can use a from the N a k coefficientswithout restrictions,and the Nth is
vegetationindex such as the NDVI [Rouseet al., 1974], the then determined. Accordingly, we can define an infinity of
SAVI [Huete, 1988], the ARVI [Kaufmanand Tanr•, 1992], TISIs, dependingon the used channelsand alsoon the partic-
etc., or more accuratetechniquessuchas the spectralm•ture ular coefficientschosenfor a given band combination.In this
analysis[Sabolet al., 1992].Therefore this model is able to link way, usingthe adequatecoefficientswe can tune the weight of
this ffpe of data to emissiviff.We have to remark that this each channel in a particular TISI definition. Each TISI is
method can only estimate the LSE and not the LST. The approximatelyequal to a combinationof emissivitiescalled
emissiviffinformation providedby the method must be intro- TISIE:
ducedinto a split-windowalgorithmin order to perform both
the emissivityand the atmosphericcorrectionsto obtain finally N

the LST [seeCaselleset al., 1997a,section4.2]. TISI = TISIE = ek (11)


k=l

4.2. TISI Algorithm where e• is the emissivitycorrespondingto band k.


The temperature-independent thermal infrared spectralin- Beckerand Li [1990a] designedthe TISI to determine the
dex(TISI) definedbyBeckerandLi [1990a]is derivedfrom the LST from AVHRR data. They derived a methodologyto ob-
ground-levelradiation equationby neglectingthe atmospheric tain the land surface emissivity using the TISIs for the
reflectionterm and by usingthe "power approximation."This AVHRR channels3, 4, and 5 (which correspondto channels
approximationconsistsof writing the band-integratedradiance A, C, andD of the PRISM sensor)[Li and Becker,1993].It is
as a power of temperature:Bi(ri) • [•ir• ', [•i and ni being based on the fact that the radiance emitted by the Earth's
coefficientsobtained by regressionanalysisbetween the tem- surfacein the region around 3.7/•m is of the same order of
peratures and their correspondingchannel radiances.These magnitudethat the Sun radiancereflectedon the surface.Thus
radiancesare obtained by integrating Planck's function over the signal measuredby day containsboth contributions.The
the spectralrange coveredby the band and weighted by its authors defined a three-channel TISI choosinga,q = 1 and
filter function. These coefficientsdepend on the channel and at> - - 1, then a c = - 1.924 (in what follows,ac represents
on the temperaturerange(seeTable 2). The generaldefinition the absolutevalue of a c, becauseits negative sign has been
of a TISI involvingN channelsis already taken into accountin the theoretical development).
11,152 CASELLES ET AL.' EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

This TISI is calculatedin two images,one takenby day and the atmosphericreflection term and by using Wien's approxima-
other by night, for the same area: tion. With these approximationsthey result in

BA(T•d) 1 N
TISId
= MBc(* ac
Tcd) ß
Bo(Tod) (12) X,In8i- .• 2 /•J
In%= c•,= X,InBi(r•
j=l
BA(T•.) eA
TISI,= MBc(T•,)aCBo(T•,)•
TISIE,= 8C• 8D (13) 1 N
whereBi(Ti*d/n)is the ground-level
radiancemeasured
in
N • XjInB•.(T•
j=l
+gi (18)
channeli by day(d) or bynight(n). The methodestimates the
emissivepart of the signalreceivedin channelA by day using where ei is the emissivityin channeli, ai is the alpha coeffi-
the emissionoccurredin this channelby night; the reflectance cientfor thisband,X• is itseffective
wavelength,
B•(T•) is the
is obtained from this part, and finally we can retrieve the ground-levelradiancemeasuredin channeli, and K• is a con-
emissivityapplying Kirchoff's law. The main hypothesisas- stantwhich dependson the characteristicsof the channel.
sumedto developthe model is that the TISIE (not the emis- From the left-handsideof (18) we seethat thesecoefficients
sivity)remainsinvariantalongtime, if the environmental con- are independentof the surfacetemperature.Kealyand Hook
ditionsdo not change;thus the TISIE by day and by night are [1993] developeda method to retrieve emissivityusingthem.
equal.Li and Becker[1993]proposethe followingrelationship Solving(18) for the emissivity
we obtain
to obtainthe emissivityin channelA, derivedusingthe defined
TISIs on the ground-levelradiation equationfor this channel
and in day conditions:
e•= exp X, (19)
w(TISId - TISI•) Bc(T•d)aCBD(TCd)
where• represents
thesecond
termon theleft-hand
sideof
(18). The alphacoefficients
are calculatedfrom the corrected
•:A(O)
= 1-- MR}A(Os)
cos
Of
A(Os,
O) (14) radiances.However, sinceit is not possibleto calculatethe X
whereR}A(0•) is theSunradianceat groundlevelin channel factorfrom satellitedata (we do not knowthe emissivityspec-
A, 0• is the Sun angle, and fA(O•, O) is the bidirectional tral shapeof anypixel),it is necessary to find a wayto estimate
angularform factor. The atmosphericeffectmustbe removed it. In fact, with this method the authors changethe problem of
from the measuredradiancesby means of a radiative transfer N + 1 unknowns(the N emissivities plus the temperature)
model, using a standard atmosphereor preferably a radio- into another problem in which temperatureis substitutedby
sonde simultaneouswith the satellite overpassin space and the• factor, whichisrelatedtotheN remaining variables(the
time. Emissivityin channelsC and D is obtained with the N emissivities). In thissense,the questionbecomesa problem
followingrelationshipsderived usingthe propertiesof TISIs: with N unknowns.However, now we have N - 1 measurements,
becausefrom the Na• coefficientsmeasured,only N - 1 are
independentvariables.This is due to the fact that the mean
ec= TISI•7•,•TISI• (15) value of all the coefficientsfor a givenpixelmustbe zero.Kealy
and Hook [1993] proposeda semiempiricalrelationshipthat
eo = e7f•TISIoc• (16) relates the X factor to the variance of the alpha coefficients
where noc = nD/nc. TISIoc,• is a two-channelindexdefined calculatedin all the channelsfor a given pixel:
as
•: ½(O'2a,)
(1/') (20)
Bo( To.)
*
where
rr2•,
isthevariance
ofthealpha
coefficients.
Theparam-
TISIoc,
= Moc
Bc(T•,)
""• (17) etersc andJ mustbe obtainedby regressionanalysisusingthe
where spectral curves of several surface samplesmeasuredin the
laboratory.We must only use the surfacesexistingin any stud-
• •I I)C
c ied area to obtaina goodperformanceof (20). Otherwisethe
MDC = error will increase.

In this work we have analyzedthis method,includingchan-


nel B, which was not consideredinitially by the authors. 5. Analysis of Methods
Now we addressthe analysisof the applicability of each
4.3. Alpha Coefficients Algorithm method.We have performedthe studyof channelsA, C, and
The alpha methodwasoriginallydevelopedto extractemis- D usingthe characteristicsof NOAA 11/AVHRR channels3
sivityinformationfrom TIMS data for geologicalstudiesand (3.7-4.1 /•m), 4 (10.3-11.3 /•m), and 5 (11.4-12.4 /•m). For
later tested for LST extraction using TIMS and simulated channel B we have used a Gaussian distribution centered at
ASTER data. This method was therefore basedon multispec- 1135cm-• (8.8/•m) andwithrr - 43 cm-•.
tral data taken from the 8 to 14/•m region and never intended We have carried out first a sensitivityanalysisof each algo-
for the 3.5/•m range.In this work we test the alpha algorithm rithm consideringonly the sourcesof random error. Certainly
on an instrumentwith fewer spectralbandsthan the method it is a difficult task to do an appropriate analysisof very dif-
was developedfor, and in additionwe have includedthe 3.5 ferent methods,if they are to be compared later. To obtain
/•m region into the analysis. "comparableresults"we have defined the same general con-
The alphacoefficients[Kealyand Hook, 1993]were derived ditionsin testingthe three methods;that is, we have intended
from the ground-levelradiation equation by neglectingthe to use the same general philosophyin analyzingthe three
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,153

Table 3. Rocks,Soils,and Vegetation EmissivityMean Values and StandardDeviations


CalculatedFrom Data of Salisburyand DMria [1992, 1994] for PRISM Channelsand Used
in Error Calculations

Emissivity

Channel A Channel B Channel C Channel D

Rocks 0.87 _ 0.07 0.88 _+ 0.08 0.93 _+ 0.03 0.959 _+ 0.014


Soils 0.80 _+ 0.07 0.92 _+ 0.05 0.967 _+ 0.007 0.975 + 0.005
Vegetation 0.985 _+0.010 0.985 _+0.010 0.985 +_0.010 0.985 _+0.010

methods.Particularly,we have considereda "mean case" for


(25)
each one (not extreme situations),and moreover,we have OPv= e,v-8tg-(l --Big
)etvF
assumedoptimal conditionsof application.For instance,we
have applied the VCM to an area with a 50% of vegetation
(26)
cover and the two other methods in midlatitude summer at- Op----•
= (1- eiv)(etaG
+
mosphereconditions.This analysisgivesus the best possible
H2
performanceof each algorithmin generalterms.
We havedefinedthe absoluteerror on emissivity(Be) from S3 H

the errors correspondingto the severalvariables (Sxi) that


take part in its calculationas

t=l
I•xi
lt•X
i (21) + - -- (27)
H
where we have consideredthe existenceof dependent vari-
ables.
To completethe study,we haveidentifiedand estimatedall H
the sourcesof systematicerrors, and additionally,the condi- S2
tionsof applicationof eachmethodhave been analyzed.This •--•: (1 -- Big)8tv(1
-- Pv) -- -
has allowed us to test the strengthsand weaknessesof each 1+
one. With this information we have performed a comparison
which is clearlylimited by the basicdifferencesbetweenthese 82
algorithms.Finally, we have identified the most favorable S
method in terms of error propagationto map the emissivity + (l -- eiv)Ps 2
from satellite. •q• (28)
5.1. Vegetation Cover Method
The way to applythis algorithmis (1) to estimatethe vege- We need to definea particularcaseto obtain an idea of the
tationcover(Pv) with datafrom the VNIR/SWIR regionusing errors we can find in general.We have chosena mean case:a
one of the existingalgorithms;(2) to calculatethe geometric mixedpixel with a vegetationcoverof 50%. For this situation
factors(F, G, F') usingthe mean dimensionsof the surface we have assumedthe best conditionsof applicationof the
(S, H); and (3) to introducethesefactorstogetherwith the methodology,consideringthat we have a sufficientknowledge
groundandvegetation (eiv, eia) in (1) obtaining of the area (typesof soil and vegetation,their structure,and
emissivities
finallythe effectiveemissivity.We have consideredthesevari- emissivitieswith the lowestfield experimentalerror). Let us
ablesin the sensitivityanalysis. seethe ordersof magnitudewe haveusedin the severalvariables.
5.1.1. Sensitivityanalysis. From (21) we will have an er- We have used the completedata set providedby Salisbury
ror in the effectiveemissivitygivenby and DMria [1992, 1994]to estimatethe mean emissivityvalues
for vegetationand groundexpectedin the PRISM channels
(seeTable 3). To achievea higherlevel of reliabilitywe have
• ig 8Pv
+ I8Ps checkedfrom our field measurements[Rubioet al., 1997] that
the mean value and standard deviation of the soil emissivities
in channels C and D coincide with the ones calculated from
(22)
laboratory data. Thus we believe that the samplesused by
where each coefficient is
Salisburyand D'Aria are studiedin conditionsquite similarto
those found in the field. On the contrary,the valuesfrom the
Salisburyand D'Aria workshavenot beenusedfor vegetation,
--=
Oe:•v
Pv + (1 - eia)F(] - Pv) - eiaGPs+ (1 - 2e,OF'Ps
becausethey are referred to leavesand not to whole plants;in
addition,we haveonlyconsideredgreenvegetation,becauseit
(23)
is the mostabundantand the bestobservablefrom satellite(it
: (1 -- •ivF)(1 -Pv) -[- (1 - •:iv)GPs (24) is more difficultto observethe bark, for example).For thistype
Ol•tg of vegetationthe emissivityis highand the spectralresponseis
11,154 CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

Table 4. Error on Effective EmissivityRetrieved With in this channeland also a lower variability(see Figure 2 and
Vegetation Cover Method in PRISM ChannelsWhen we Table 3).
Have "Rocks" as Backgroundand 50% Vegetation Cover 3. The error rangesfrom _+1-1.5%in bandA to _+0.5%in
Rocks
channelD, decreasingas wavelengthincreases.Therefore the
( SP•/P•), best bandsto use regardingthe emissivityretrieval are those
% A B C D C+D correspondingto 10.5-12.5/xm. This resultis explainedby the
fact that the larger the wavelengthsthe narrower the intervals
5 0.9 0.8 0.6 0.4 0.5
10 1.0 0.9 0.6 0.5 0.6
of emissivityvalues.
15 1.1 1.0 0.7 0.5 0.6 5.1.2. Sources of systematicerror. Let us now analyze
20 1.2 1.1 0.7 0.5 0.6 the sourcesof systematicerror.
25 1.3 1.2 0.8 0.5 0.7 Doublescattering: The definitionof the effectiveemissivity
Mean 1.1 1.0 0.7 0.5 0.6(equations(1) to (3)) hasbeenreachedby neglectingdouble-
scatteringprocesses betweenthe differentpartsof the radiative
The relative error in emissivityis shownin termsof severalrelative
system.The importanceof the doublescatteringwasevaluated
errors in the vegetationcoverestimate.The mean valuesare shownin
the last row. by Casellesand Sobrino[1989],who checkedthat this approx-
imation resulted in a systematicerror of about 0.1 K. If we
expressthis error in terms of emissivity,we will obtain the
nearlyflat, as shownby Salisburyand DMria [1992].They have systematicerror introducedinto the effectiveemissivityby the
observedthat the scatteringwithin a plant increasesemissivity doublescattering.In the 10.5-12.5/xm region and for a surface
and reducesthe spectraldifferences.This meansthat vegeta- temperature of 290 K, an emissivityerror of 1% leads to a
tion behavesas a gray body. For these reasonswe have con- temperature error of 0.6 K. Then, the 0.1 K evaluated by
sidered a value of 0.985 in all channels based on our field Casellesand Sobrino[1989] is equivalentto making an emis-
measurements[Rubioet al., 1997].With regardto the emissiv- sivity error of 0.17%. If we consider the double-scattering
ity errorswe have consideredthe bestcasewe can find in field processes,the effectiveemissivitywill increaseabout 0.17%,
measurements,whichis about _+0.005in all samplesand chan- then the difference between the "true" and the modeled emis-
nels [Nerryet al., 1990;Rubio et al., 1997]. sivitieswill be +0.17%, which is the systematicerror intro-
In relation to the dimensions we have considered that H and ducedby this approximation.We will use the samesigncrite-
S are of the same order of magnitude.Then we have got H rion in the evaluationsof systematicerrorswhichwill be shown
S • 1, becausewhat is importantis their relativevalue and not later on.
their absolute values. We have assumed that their errors are of Geometric model and shape factors: The VCM uses a
the order of a 10% of their values(SH • +_0.1 H and 8S combinationof two geometricmodels:(1) To evaluatethe
s). shape factors (F, G, F'), it uses the model proposedby
For the proportionswe have consideredmixedpixelswith a Sutherland and Bartholic[1977],and (2) to estimatethe vege-
50% of vegetationcover and 50% of soil. Therefore we have tationcover(P v), it usesthe geometrysuggested by Valorand
put P v • 0.5; we have alsoconsideredcertain angularobser- Caselles[1996]. The first one assumesthat vegetationis dis-
vation, thus Ps • 0.2 and Pt • 0.3. tributed following infinitely long Lambertian boxes;however,
Finally, sincethe methodfor the vegetationcoverretrievalis the secondone is a more realisticapproximation,considering
undetermined,we have used relative error values in this pa- the vegetation elements as square finite boxes. Taking into
rameter rangingfrom _+5%to _+25%.It is a reasonablerange, accountthe results of Colton [1996], who usesCasellesand
if we use one of the methodsmentioned above [Ichoku and Sobrino'smodel but varies the geometry [see Colton, 1996,
Karnieli, 1996]. Figure 11], we haveconsideredthat the systematicerror due to
Using thesevalueswe have obtainedthe errorspresentedin the use of two different geometricmodels is of the order of
Tables 4 and 5, where we have consideredseparatelythe case +0.1%.
in which we have a surfacebackgroundformed by rocksor by
soils. Besides channels A-D we have also considered the mean
channelsC + D, for whichthe methodwasinitially developed.
We can observethe following:
1. In both casesthe error in emissivityvaries slowlyalong
with the error in the vegetationcover,of the order of _+1%0 or Table 5. Error on Effective EmissivityRetrieved With
less each _+5% in P v. The two terms which determine the Vegetation Cover Method in PRISM ChannelsWhen we
weight of the error on P v in the total error on emissivityare Have "Soils" as Backgroundand 50% Vegetation Cover
given by (25) and (26). On the one hand, in (25) the first Soils
right-hand term refers to the difference between vegetation ( SP•/P•),
% A B C D C +D
and soil emissivities,which is a small value (both valuesare
closeto the unity). In the secondright-handterm we have a 5 1.2 0.6 0.4 0.4 0.4
factor (1 - eia) whichis alsosmallfor the samereason. 10 1.4 0.7 0.4 0.4 0.4
Therefore the difference between these two small terms will be 15 1.6 0.7 0.4 0.4 0.4
20 1.8 0.8 0.5 0.5 0.5
smaller.On the otherside,(26) hasalsoa verysmall(1 -
25 2 0.9 0.5 0.5 0.5
factor. These low values explain the slight sensitivityto the
Mean 1.6 0.7 0.4 0.4 0.4
error on the vegetation cover parameter.
2. The error in the caseof rocksis larger than the error in The relative error in emissivityis shownin terms of severalrelative
the caseof soilsin all channelsexceptin channelA. This is due errorsin the vegetationcoverestimate.The mean valuesare shownin
to the fact that rockspresenthigheremissivityvaluesthan soils the last row.
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,155

5.2. TISI Method In this case the variables we must take into account are the

The proceduredescribedby Li and Becker[1993] to obtain groundlevel brightnesstemperaturesmeasuredby day and by


with this algorithmis (1) to perform the atmo- night and the Sun radianceat surfacelevel. By applying(21),
the emissivities
sphericcorrectionson the measuredradiancesusingadequate the error on the effectiveemissivityfor each channelwill be
radiosondes;(2) to coregisterthe day-nightimages;(3) to o/A 7r
calculate
theTISIs;(4) to estimate
R.•A(Os)andfA(0s,0); (5) •A = s
to obtainthe emissivities from (14) to (16); and (6) to renor-
Rg.•(0,)cosO,f.•(O,,O)
malize the calculatedemissivitiesif it is necessary;that is, if the
emissivityis larger than 1 in some pixels. According to the ' nA --Ad OdTISI•
authorsthismay happenin the channelswith largestemissivities
(generallythe channelsat largestwavelengths), due to system-
atic errors arisingfrom the severalsourceswe can identify. +ncacTISI•Tcd
--Od
[•T Cd+ • C
5.2.1. Sensitivityanalysis. Let us first analyzethe sources
of random error of this method.
*•'"• * (33)
Sunradianceat surfacelevel(/•g.•(0s)): An estimationof + noTISI,,Tcd
TDd
k•T Dd+ 8TD
this quantityis neededto take into accountthe reflected com-
ponent of the signalmeasuredin channel.4. This estimation aec nc ae• aT• nc aT•
.... • nc , + (34)
can be made empiricallyin the field, or theoreticallyintroduc- ec ncac + no e4 Tc• ncac + no T•
ing the Sun radiance at the top of the atmosphere into a
radiative transfer code. The error on emissivitydue to this
=••+no , +no , (35)
sourceis [Li and Becker,1993] eo nc ec T c• Tm

6e.• 1 - e.• 6R•.•(0s) where 8T* is the error in the atmospheric


correctionof the
--
e.• e.• R•n0x) s
(29) brightnesstemperature measuredin channeli. Here we have
where6R}n(0x) is the errorin Sunradiance.We haveesti- only consideredthe random error in the atmosphericcorrec-
tion. The error due to the Sun radiance estimation has been
mated that an error of • 10% in this quantity,for instance,will
producean error of 21% in channelA emissivity. treatedbefore andwill be addedto ben (equation(33)). These
Atmospheric corrections: The atmosphericperturbation relationshipscorrespondto the combinationof bandsA, C,
must be removed from the radiances used to calculate the and D, initially defined by Li and Becker [1993]. We have
TISIs. Li and Becker[1993]proposeto usea radiativetransfer extendedthe studyby includingchannelB and by considering
code togetherwith the adequateatmosphericprofile calculat- all the possiblecombinationsof three channels,one of them
ing the atmosphericparameters (transmittance and atmo- being channelA, which is alwaysnecessa• due to the basisof
sphericradiance)and solvingthe radiativetransferequation. the method. The objectiveis to maintain the basicstructureof
This can only be accomplishedby using a radiosondesimulta- the used TISIs but changingthe involved channels.It is ve•
neouswith the satelliteoverpassin spaceand time (satellite easy to obtain the expressionsfor the error in each channel
soundingsare lessaccurate).Sincetheseconditionsare rarely combination,startingfrom (33) to (35): in theseformulaewe
satisfied,they studiedthe errorsarisingfrom the use of several only have to changechannelsC and D by two other channels,
standardatmospheresdifferentfrom the actual one. They an- calculating the correspondinga• coe•cient for the channel
alyzedthe systematicerrorsproducedby thesecalculationsby which substitutesthe C band, taking into accountthat the two
means of simulation and obtained a mean error in the emis- othercoe•cientsare 1 and -1, respectively, andthat (10) must
sivityretrieval of about 3% in channelsC and D and of 5% in be satisfied(note that in the derivedformulae this coe•cient
channelA. They claimedthat theseerrorsmightbe reducedto must be used without the negative sign, becauseit has been
lessthan 1% usingatmospherescloseto the actualone. In this alreadytaken into account).Table 6 showsthe a• coe•cients
used for each combination.
paper we analyzethe random errors associatedto the uncer-
We have considered a mean situation: the case of a midlati-
tainties in the measurement of humidity and temperature,
which arisewhen we use the adequateradiosonde,not atmo- tude atmospherein summer.We have also assumedoptimal
sphericmodels more or less closeto the actual one. In this conditionsof application:the useof an adequateradiosondeto
sensethese calculationsare complementa• to the work car- apply the atmosphericcorrection and to have a su•cient
ried out by Li and Becker[1993],who did not perform them. knowledgeon the studiedarea in order to use an appropriate
To simpli• the error calculations,we have rewritten previ- model to deal with the angular form factor. Let us see the
ously (14)-(16) in terms of temperatures,using the power ordersof magnitudewe have used.
approximation.They become The errors in the corrected temperaturesare due to the
random errors in the atmospherichumidiff and temperature
e.(0) = 1- measurementmade in a radiosonde.Maul and Sidran [1973]
R}n(Os)cosOsf•(Os, O) showedthat ffpical errorscouldbe • 10% in relative humidiff
and •5 K in temperature. These errors considerthe uncer-

.['T*.•r*"•
An T*
•d--•*ncacT*
nD
• ncacT*
Cd nD•
' Dd
]
Cn • Dn (30)
tainty in the measurementand also include the effect of the
spatialvariabiliff of the atmosphere.This effect resultsfrom
the extrapolationof a measurementtaken in a definedlocation
= • (31) to an extensivearea.Maul and Sidran[1973] estimatedthat for
thesevalues(consideringmidlatitudesummerconditions)we
: • (32) could expectan error of •0.5 K on the temperaturemeasured
Tc•J in AVHRR channel4 (PRISM channelC). We believe that
11,156 CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

Table 6. Coefficientsa i Used in Each StudiedChannel emissivityin channelA, we haveestimateda value of 0.9, using
Combination for TISI Algorithm the data from Salisburyand DMria [1992, 1994].
Channel
We haveconsidereda value of fA (0s, 0) • 1 (Lambertian
Combination an aB ac ao behavior),assumingthe knowledgeof the angularbehaviorof
the surface.
ACD 1 .... 1.919 -1
Finally, in order to simplify the calculations,we have as-
ADC 1 .... 1 - 2.004
ABD 1 -1.586 .... 1
sumedthat the three temperaturesmeasuredby day are equal,
ADB 1 -1 .... 1.775 and the sameappliesfor the temperaturesmeasuredby night
ABC 1 - 1.515 - 1 -.. (T•). This is a good approximationbecausethese tempera-
ACB 1 -1 -1.625 -.. tures only differ a few degrees.We have checkedthat it can
producevery smallchangesin the error estimate(•1%0). We
In the calculationsusingformulas(33) to (35), it is not necessary
to
put the negativesign,whichappearsin this table,becauseit hasbeen have performed the calculationsfor T d rangingfrom 280 K to
alreadyconsideredin the theoreticaldevelopment. 320 K, and T• being always15 K lower than Td.
Table 7 shows the results for the channel combination ACD.
The error increaseswith surfacetemperature;the higher the
temperature the larger the increase.On the other side, the
this is a reasonableand even an optimisticvalue, if we take into errorintroduced
byR}•(Os) affects
mainlye• (1% in •eA), its
accountthe comparisonbetweenradiosondestaken by differ- influencebeing smaller on the other channels(•0.3-0.5%).
ent meteorologicalservices(see,for example,EFEDA project Table 8 showsthe summary of the calculationsfor all the
results).In addition,we calculatedthe errorsin the corrected possiblechannelcombinations.In any case,there is alwaysa
temperaturesrelated to variationsof _+10% in humidity and channelwhich performswell, whereasthe other two present
+2 K in temperature, and values larger than _+0.5K were larger errors.The best result is obtainedin the band placedin
obtained [Caselieset al., 1997a]. To estimatethe error corre- the originalchannelC position.To obtainthe minimumerrors
spondingto the other channels,we haveusedthe resultsshown in each channel, we should use several combinations and ob-
in Figure3 (midlatitudesummercase)of Caselles etal. [1997a]. tain from them the value of the best channel in each case. This
There, we can find the relative sensitivityto the atmospheric would lead to the errors shown in the last row of Table 8.
correction of each channel if we use a single-channelalgo- Clearly, the error decreasesas wavelengthincreases.
rithm. Assumingan error of _+0.5K in channelC and using 5.2.2. Sourcesof systematicerror. This optimal casewill
these results,we estimated errors of +_0.2K in bandA, +_0.4K be affected inevitably by some sourcesof systematicerror,
in band B, and _+0.7 K in band D. which have been estimated only for the combinationACD,
For the atmospheretype considered,and a Sun angleof 30ø, becausein the previousanalysiswe have obtained similar re-
a typicalvalueof channelA transmittanceis 0.7 [seeCaselles
et sults in all combinations.
al., 1997a, Figure 1]. Image registration: The TISI method combinesthe TISIs
The Sun radiance at surfacelevel has been estimatedusing calculatedboth by day and by night in a pixel-by-pixelbasis.
the following approximaterelationshipproposedby Li and Thus to determinethe emissivity,the TISIs mustbe calculated
Becker[1993]' exactlyfor the samepixelsand in the same observationcon-
ditions. Otherwisewe can obtain large errors in emissivity.Li
R*}n(Os)
• 0.415536x 10-('•rrn(Os) (36)
and Becker[1993]did not treat this questionquantitativelyand
where rA(0s) is the atmospherictransmittancein the 0s direc- only suggestedthe applicationof averagingproceduresto re-
tion. Consideringa transmittancevalue of 0.7, we have esti- duce the error. Consequentlywe have made a rough estimate
mateda Sunradiancevalueof 10-•' W/(cm2 sr cm-•) in the of this source.
mentioned conditions.In addition, the error of eA due to this It is easyto see that if we calculatethe TISIs by day and by
source is of the order of _+1%, as we have shown. This error night in two different pixels (labeled 1 and 2), the error in
mustbe addedto the valuesobtainedfrom (33), becauseit only channel A will be
considersthe atmosphericinfluence.
Sincethe TISI is a combinationof emissivityvalues,closeto
unity, we have used a value of TISI • 1; and for the mean

Table 8. Relative Errors on Effective Emissivityfor Each


Channel, in Terms of the Used Channel Combination,
Table 7. Relative Errors on Effective Emissivityfor Obtained by Applying TISI Algorithm
CombinationACD Obtainedby ApplyingTISI Algorithm
Channel
r• Combination A B C D C + D
K A C D C+D
ACD 3 ... 1.7 3 2
280 1.8 1.4 3 2 ADC 3 ... 4 1.9 3
290 2 1.5 3 2 ABD 3 1.9 ... 3 ...
300 3 1.8 3 2 ADB 3 5 -.. 1.9 ...
310 4 2 3 3 ABC 3 1.8 3 ......
320 5 2 4 3 ACB 3 4 1.7 ......

Mean 3 1.7 3 2 Minimum 3 1.8 1.7 1.9 1.8

We haveperformedthe calculationsfor a surfacetemperaturerang- Last row showsthe minimum error we can achieveif we use the
ing from 280 to 320 K. Last row showsthe mean error. adequateband configuration.
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,157

1 - 8.• TISI,•(2) - TISI,•(1) Table 9. SystematicErrors for the ACD Combination Due
(37) to Power ApproximationUsed in Definition of TISIs
8A 8.a TISId(1 ) - TISI,•( 1)

and using the power approximation,it is


K K A C D

--Cd' Dd __
TncacTnøISA(2)
cos 8n(1)
Os8C(2)a•80(2) 1 (38)
8C(1)a•80(1) 300 285
295
+0.09
+0.03
+0.4
+0.5
+0.0007
+0.10
310 285 +0.19 +0.5 +0.03
The error depends on the differencesbetween the pixel 295 +0.09 +0.5 +0.12
emissivities.Therefore if we apply the method to a very ho- 320 285 +0.30 +0.5 +0.07
295 +0.16 +0.5 +0.14
mogeneousarea,the effectof thiserror will be small.However,
in heterogeneousareas,asthe land surfaceusuallyis, this error We haveperformedthe calculationsfor three day temperaturesand
couldbe large. For instance,if we merge a bare soil pixel with two night temperatureseach.
a vegetatedpixel, this error could rise to +3%.
Another problem is the change of the observationangle
between day and night. Again if we are observinga homoge- atures.To estimateits effect,we haveusedthe sameprocedure
neoussurface,it will not be problematic(althoughan emissiv- as above, introducingthis term into the simulatedradiances,
ity angularvariation hasbeen observedin somesurfaces,it will and consideringthree standardatmospheres:midlatitude win-
onlybe importantat large observationangles).The problems ter, midlatitude summer,and tropical. Figure 6 showsthe ob-
arisewhen dealingwith heterogeneousareas;in this case,the tained errors(definedas in the previouscase)for eachband
definition of the effective emissivitychangeswith the view consideringa midlatitude summer atmosphere.We have ob-
angle,becausethe "observedsystem"is different in each case tained similar relative distributions in the two other atmo-
[seeCaselleset al., 1997a,section4.3.2].An emissivityobtained spheretypesbut with different dispersions.Table 10 presents
using data taken at two different angles will be difficult to the summaryof the results,showingthe mean value, the stan-
interpret. In consequence,this question must be well con- dard deviation, the maximum, and the minimum values corre-
trolled in the applicationof this method. spondingto each atmosphere.From the mean valueswe can
Approximations of the theoretical development of TISIs: seethat this approximationproducesan overestimationof the
Two approximations are made:(1) the power approximation retrieved emissivities.The effect is larger in band C than in
and (2) the eliminationof the atmospheric reflectionterm on channelsA (due to a lessatmosphericeffect) and D (due to
the ground-levelradiation equation.Regardingthe first point, higheremissivity values).In winter conditionsthe effectwill be
Li and Becker[1993] did not estimatethe effectof this approx- small in all bands (of the order of the power approximation
imation on the emissivityretrieval, so we have performed this error), and it will be alwayssmallin channelA, which is less
estimate for the PRISM bands. We have simulated the radi- sensitiveto the atmosphereeffect. As for bands C and D in
ance at surfacelevel in each channelfor severalday and night summer and tropical conditions,the error can be significant
temperatures,using Planck's law and neglectingthe atmo- (-1 to -2%), beingvery large in somecases,as shownby the
sphericreflection term. The simulationhas been carried out minimum values.
usingthe emissivityvalues,which correspondto severalsam- The two sourcesof systematicerror, i.e., the power approx-
ples providedby Salisburyand D54ria [1992, 1994]. We have imation and the reflection term, have opposite effects and
calculated the TISIs from these radiances and applied the couldhave a partial compensation.Nevertheless,the ordersof
explainedmethodologyto obtainthe emissivities.Finally, com- magnitudeobtainedshowclearlythat the secondeffectwill be
paring the retrieved emissivitieswith the original ones, we dominant.
obtain an estimate of the error (defined as 8(sample)- There are other additional sources of error which we have
8(retrieved))relatedto the powerapproximation(becausewe not addressed in this article. We can mention the error due to
have obviatedthe reflectionterm in the simulation).Table 9 the lack of knowledgeof f,• (Os, 0 ), and the systematicerrors
showsthe systematicerrors obtained in the studied channels. arisingfrom the use of inaccurateatmosphericprofiles.These
The obtainedvaluesare small,beinglarger for channelC. The sourceswere studiedby Li and Becker [1993]. We have not
maximumvalue is 0.5%; thusthis sourcewill not be significant. consideredthem becausethey fall out of the "optimal condi-
Performingthesecalculationswe have observedthat the algo- tions of application"claimed for the three methodsto obtain
rithm isvery sensitiveto the valuesof the parametersused(the "comparable"results.Thus we have only taken into account
coetticients/'/i,ai, ]3i)' In fact, if we use approximatevalues those sourcesthat will inevitably affect the resultseven in the
insteadof the exact ones, systematicerrors arise quickly. In best of cases.
order to avoidtheseproblemsthe adequateparametersshould
be used. 5.3. Alpha Coefficients Algorithm
As for the secondissue, the removal of the atmospheric This methodis appliedfollowingthesesteps:(1) to perform
reflectionterm affectsthe resultin two ways:first, throughthe the atmospheric correction on the radiances using a single-
derivationand definitionof the TISI itself and its approximate channel algorithm along with the adequate atmosphericpro-
equalityto the TISIE; second,in the atmosphericcorrectionof file; (2) to calculatethe alphacoefficients
from theseradiances;
the measured radiances: since we do not know the emissivities, (3) to estimate
the• factorfromthealphacoefficients
vari-
we cannot take into accountthis term despitehaving a radio- ance using the appropriate empirical relationship, derived
sonde.Beckerand Li [1990a] made only a rough theoretical from laboratory spectracorrespondingto the surfacesin the
estimate of this source for TIMS channel 5 and a surface studiedarea; and (4) to derivethe emissivitywith (19).
temperature of 300 K. We have performed our own calcula- 5.3.1. Sensitivity analysis. Kealy and Hook [1993] only
tions for all the studiedchannelsand severalsurfacetemper- performed an analysisby using simulationprocedures,where
11,158 CASELIES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

they consideredonly the errors due to the approximationsof


the method. They did not study the errors related to the at-
Channel A
mosphericcorrection.Furthermore,their analysiswasmadein
terms of temperature, not of emissivity.In consequence,we
havecarriedout a completelynew analysis,extendingthe study
made by the authors.Let us first see the random errors.
Atmospheric correction and estimation of X: Here we an-
alyze the same situation as in the TISI method: the error
(random)due to the uncertaintiesin the adequateradiosonde.
Together with this sourcewe have consideredthe error intro-
ducedbytheempirical
estimation
of the,• factor.
To simplifythe error calculations,
we haverewritten(19) in
terms of temperatures,usingWien's approximation:

-2 i i i i
XilnAi T,* • j=A
XjlnAj- Tj] +Ki+
0.5 0.6 0.7 0.8 0.9 e,= exp •,
(39)
Channel C where C 2 is the secondPlanck constantand
C1
A,= X•rr' (40)
where C• is the first radiation constant.
The main variables which have an influence in the total error
are ground-level brightness temperatures measured in all
channelsand the X factor. Applying (21), the error in any
channel will be

-3

-5
e•-:4
X,3•-•-•+ • Tj'•+•,&•'(41)
ß where ST,*.is the error in the atmospheric
correctionof the
i i i
brightnesstemperaturemeasuredin channeli, and 8X is the
0.8 0.85 0.9 0.95 error in the estimate of the • factor.
These relationshipscorrespondto the combinationof the
four channels A, B, C, and D. We have also studied all
Channel D possiblecombinationsof three and two channels.In the first

Table 10. SystematicErrors Due to Removal of


AtmosphericReflection Term in TISI Algorithm
iSsls,%

Mean s.d. Maximum Minimum

Midlatitude winter
A -0.1 0.1 +0.1 -0.5
C 0.0 0.4 +0.7 -1.3
D 0.0 0.3 + 1.0 -0.6
-2
Midlatitude summer
A -0.3 0.5 +0.8 - 1.8
-3 C -0.8 1.5 + 1.1 -6
i i

D -0.6 1.1 + 1.8 -3


0.9 0.92 0.94 0.96 0.98 1
Tropical
A -0.1 0.6 +1.4 -1.7
C -1.7 1.8 +0.4 -8
D -1.2 1.1 +0.7 -4
Figure 6. Systematicerrors for the channelsA CD combina-
tion due to the atmosphericreflection term approximation
To perform the calculations,we have consideredthree standard
usedin the definitionof the temperature-independentthermal atmospheretypes: midlatitude winter (T s -- 290 K), midlatitude
infrared spectralindices,consideringthe midlatitudesummer summer(Ts = 300 K), and tropical(T• = 310 K). In addition,we
atmosphere. have usedthe emissivitylaboratoryspectraprovidedby Salisburyand
DMria [1992, 1994]for severalsoil and rock typesand consideredthe
vegetationas a gray body with s -- 0.985.
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,159

case,(41) is basicallythe same,changingthe factor 1/4 into 1/3 Table 12. Relative Errors on EffectiveEmissivityfor
and adding the three consideredchannels instead of four CombinationABCD Obtained Using Alpha Coefficients
bands. Similarly, in the case of two channels,the factor be- Algorithm
comes1/2, and two bandsare added.The error expressions for
ae/e, %
these cases are
T,
K A B C D C+D
(3 channels)
280 7 3 2 2 2
290 7 3 2 2 2
(42)
3x, 2E'W+ Z*r,'] 300
310
6
6
3
3
2
2
2
2
2
2
320 6 3 2 2 2
(2 channels)
Mean 6 3 2 2 2

We haveperformedthe calculations
for a surfacetemperaturerang-

•,=2,•, •, T•I +•8•


j=l (43) ing from 280 K to 320 K. Last row showsthe mean error.

In this method the atmosphericperturbation must be also


removedfrom the temperatures.We have usedthe samecon- Table 12 showsthe resultsobtainedin the caseof usinga
ditionsasin the TISI algorithmanalysis:the caseof a midlati- four-channelcombination.There is a slight decreasein the
tude summeratmosphereand the best conditionsof applica- error when temperature increases.Table 13 showsthe sum-
tion, i.e., usingan adequateradiosondeand havinga sufficient mary of all calculations,beingsimilarfor all combinations.The
knowledgeof the studied area to have the minimum error in greatestdifferencesarise in channelA. We have observedthat
the • estimate.
when usingfewer channels,the error in the atmosphericpart
The ordersof magnitudewe haveusedare the following: slightlydecreases,
but the errorin the• termincreases,
the
With regardto temperatures, we haveconsideredagainthat final result being that the total error remainsunchangedor
their errorsare related to the uncertaintiesin the appropriate increases(exceptfor the AD combinationwhere a slightde-
radiosondeswhen we use them to perform the atmospheric creaseis observed).As with the TISIs, if we wish to obtain the
correction. Thus we have used the same values as in the TISI
best results in each channel, we should use several combina-
algorithmanalysis. tions simultaneouslyand get the emissivityfor each channel
As effectivewavelengths,
we haveused•A - 3.7 •m, •B - from the adequatecombination.With thisprocedurewe would
8.8 /•m, Xc = 10.8 /•m, andXD = 11.9 /•m. obtain the minimum error valueswhich are presentedin the
To estimate the error on the X term, we have followed the last row of Table 13. The error alsodecreases with increasing
procedureproposedbyKealyandHook [1993].Takingthe data wavelength.
providedby Salisburyand DMria [1992, 1994],we havecalcu- We haveanalyzedthe methodas definedby Kealyand Hook
latedfor eachchannelcombinationtheX factorand the alpha [1993],who only usedthe thermal infrared channels.Sincethe
coefficients
variancecorrespondingto each sample.Then, we PRISM instrumentalsoprovidesmeasurementsin the VNIR/
havemade a regressionanalysisusingrelationshipsaccording SWIR region, the alpha method could benefit from an esti-
to (20), and we haveusedthe error of estimategivenby the mate of P v. With this estimatewe could selecta defined subset
regression
asthe erroron the]( factor.Table11 showsthe of points(thosecorresponding to a givenP•, value) to obtaina
values for each band combination.
particularX-variance curveoptimizedfor this case.This would
We have also assumedthat temperaturesmeasuredin the allow us to obtain better estimates of this curve. We have
different channelsare approximatelyequal and checkedthe
validity of such approximation.The calculationshave been
made for LST rangingfrom 280 K to 320 K.
Table 13. Relative Errors on EffectiveEmissivityfor each
Channel, in Terms of the Used Channel Combination,
Table 11. Values ObtainedFrom RegressionAnalysis ObtainedApplyingthe Alpha CoefficientsAlgorithm
PerformedUsingEquation (20) With Data Set Providedby
8e/e, %
Salisburyand DMria [1992, 1994] Channel
Combination A B C D C + D
Channel
Combination c J r2 8•I•, /•m ABCD 6 3 2 2 2
BCD ... 3 2 2 2
ABCD - 1.2025 4 0.79 0.13
ACD 6 ... 2 2 2
BCD -1.1357 5 0.78 0.14
ABC 8 3 3 ......
ACD -1.1904 4 0.67 0.13
ABD 6 3 ... 2 -. ß
ABC - 1.3892 4 0.67 0.2
AB 12 6 .........
ABD -1.4369 3 0.83 0.14
AC 6 ... 2 ......
AB -1.5788 5 0.36 0.4
AD 5 ...... 1.7 .- ß
AC -1.1700 5 0.53 0.18
BC '" 4 3 ......
AD -0.9041 7 0.50 0.13
BD '" 4 ... 2 ...
BC -1.4241 5 0.69 0.2
CD ...... 2 2 2
BD - 1.3298 4 0.82 0.17
CD - 1.8043 4 0.83 0.14 Minimum 5 3 2 1.7 1.9

J andc are the coefficients


of the regression,
r 2 is the correlation Last row shows the minimum error we can achieve in each case if we
coefficient,
and8• isthe• errorof estimate
givenbytheanalysis. use the adequateband combination.
11,160 CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

Table 14. SystematicErrors for ABCD CombinationDue the reflectionterm is basicallythe productbetweenthe down-
to Wien's Approximation Used in Definition of Alpha ward atmosphericradiance and the surface reflectance; in
Coefficients bandA the former is small and increasesto channel D, and the
latter hasthe oppositebehavior.Consequently,in the extremes
ae/e, %
T, (bandsA and D) this term is lower than in the center(bands
K A B C D B and C). This justifieslower errors in the extremeregions
(.4, D) and highervaluesin the centeredones(B). Consid-
280 + 1.8 +0.5 -0.2 -0.8
300 +2 +0.6 -0.3 -1.0
ering the differentatmospheres,in the midlatitudewinter case,
320 +3 +0.8 -0.4 -1.3 the effectis small exceptin bandsA andB. In the summerand
tropicalcasesthe errorsare significantin all channels(1-3%),
We have made the calculationsfor three different surfacetemper- therebeingsampleswith verylargevalues(4-10%), aswe can
atures.
see from the maximum and minimum columns.

5.4. Discussion

estimatedthat this procedurecould reducethe errors in emis- We considerfirst the mean situation analyzedin each case,
sivityby 20% approximately. assumingthe best conditionsof application,where we find the
5.3.2. Sourcesof systematicerror. This case will be af- minimum errors we can achievein each algorithm in general
fected by some sourcesof systematicerror, even if we use the terms. Figure 8 showsall the resultsobtained in these condi-
adequate
radiosonde
andthe bestpossible
•-variancerela- tions. We can see that in the three cases the error decreases
tionship.We have estimatedthem consideringonly the com- with increasingwavelength,due to the increaseof emissivity
bination ABCD, becausewe have obtained similar results for valuesand the diminution of their variability in thesespectral
all possibilities. regions.We can comparetheseresultswithin the limits defined
Approximations of the theoretical developmentof alpha co- by the assumptionsof this work. We must not forget that the
efficients: Two approximationsare used:firstwe haveWien's conditionsof applicationof the VCM are very different from
approximation,which works better at short wavelengths.To the other methods;thuswe mustplacethe comparisoninto the
estimate its effect we have used the same procedure as for proper context.From this comparisonwe could concludethat
TISIs; that is, we have simulatedthe ground-surfaceradiances the VCM seemsto presenta better performancein the error
by usingthe Salisburyand D'Aria's databaseand by neglecting propagation.Thus this method could be the most favorablein
the atmosphericreflectionterm. Then we haverecoveredfrom the LSE retrieval within the philosophicframework used in
them the emissivityby usingthe alpha coefficientsand, finally, this work. At this point we must mention that we have not
comparedthe initial and the retrieved emissivities.Table 14 consideredatmosphericerrors in the VCM analysis,because
showsthe obtainedvaluesfor three surfacetemperatures.The this method does not use thermal radiances to estimate LSE,
errorscan be significant,especiallyin channelsA (where they as the other ones do. However, this same reason leads us to use
can rise to +3%) and D (-1.3%). It is worth notingthat the the VCM results together with a split-windowalgorithm in
largesterrors appear in channelA, where Wien's approxima- order to retrieve LST. This issue will be treated in the second
tion functionsbetter in terms of radiance.It is easyto seefrom part of this series[Caselieset al., 1997a].
(19) that the ratio between the calculatedand the original These resultscorrespondto the optimal conditionswe can
emissivitiesis givenby a factor exp { v,Aa,), where v• is the find. To achievethis, a sufficientknowledgeof the studiedarea
wavenumberand Aa i is the changein the alphacoefficientdue in the three casesis needed:in the alpha coefficients,to have
to the approximation.Table 15 showsboth the variation in anappropriate •-variancerelationship, whichisobtained tak-
radiance and in the coefficients. Variation in radiance is much ing into accountonly the surfacetypesin the area; in the TISI
lower in channel .4 than in the other ones. However, in terms method, to know the angular form factor; and in the VCM
of alpha coefficientswe obtain similar orders of magnitude. case,to estimatethe surfacetypes(emissivity,geometricstruc-
This is due to the fact that the alpha coefficientin one channel ture) andtheir distributionon the area.Apart from this,in the
is calculated with the radiances measured in all bands; thus all two first methods,adequateradiosondesare needed in order
the radianceerrors are propagatedinto this coefficient.Since to perform the atmosphericcorrections(one for the alpha
the emissivityerror dependson the product(v, Aai), we will coefficients, two for the TISIs) on the thermalradiances.Cer-
obtain larger valuesfor higher wavenumbers. tainly this condition is not usual, except for well-controlled
The other sourceis the removal of the atmosphericreflec- experiments.More realistic is the use of satellite soundings
tion term in the ground-levelequation,as in the caseof TISIs, (whichare lessprecise)or standardatmospheres closeto the
and will affect the result in the sameway: on the alpha defini- particular conditionsof the area (which do not coincide,in
tion and on the atmosphericcorrection carried out in the general,with the actualprofile). Theseprocedurescan intro-
measured radiances. To estimate this source, we have used the duce significantsystematicerrors in the retrieved emissivity,
same simulationprocedureas previously.Figure 7 showsthe
errorsobtainedin the midlatitudesummeratmospherecase.In
the two others we have obtained similar distributions but with
Table 15. Error in Radiancesand Alpha Coefficientsfor
different dispersions.Table 16 showsthe summaryof the re-
sults. From the mean values we observe an underestimation of Each Channel ConsideringCombinationABCD Due to
Wien's Approximation
emissivityin band A and an overestimationin channel B;
channelsC andD presenta nearlyunbiaseddistribution.From A B C D
the deviations it is clear that the error increases from A to B
and decreases in C and D. This is due to the increase of both AB,/B,, % 3 X 10-4 0.4 1.2 1.8
Aa,, cm-• -9 X 10-6 -6 X 10-6 3 X 10-6 1.2x 10-s
the atmosphericeffect and the emissivityvaluesfrom A to D:
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,161

Channel A Channel B

10

2
o
-2

-4

-6

-8
-12 -10
& 0.5 0.6

Channel C Channel D

AA

Figure 7. Systematicerrors for the channelsABCD combinationdue to the atmosphericreflection term


approximationused in the definition of the alpha coefficients,consideringthe midlatitude summer atmo-
sphere.

thusa gooddescriptionof the atmospherebecomesnecessary. side,we have seenin the analysisthat both the alpha and the
Furthermore,the TISI method needs,in addition, a good es- TISI algorithmspresent important sourcesof systematicer-
timate of the Sun radiance at surface level and the VCM an rors,whichwill further increaseand propagatethe considered
estimateof the vegetationcover in each pixel. On the other errors even in the best situation. In the first case we have the
two approximationsmade in the theoretical developmentof
the alpha coefficients:Wien's approximationand the removal
Table 16. Systematic ErrorsDue to Removalof Atmospheric of the atmosphericreflectionterm. In the secondone, we have
ReflectionTerm in Alpha Coefficients Algorithm similar approximations,and moreover, the problem of the
coregistrationof the day-nightimages.We have checkedthat

Mean s.d. Maximum Minimum

Midlatitude winter
A +0.6 0.7 +3 -0.7 .ALPHA
B -0.6 0.8 +0.5 -3 ElTISI
C +0.1 0.3 + 1.0 -0.7 .QVCM(realisticcase)
D +0.1 0.3 +0.9 -0.2
E]VCM
Midlatitude summer
A +3 2 +10 -1.4
B -2 3 +3 -10
C +0.0 1.0 +3 -4
D + 0.2 0.8 +3 - 1.3
Tropical
A +4 3 +13 -0.2
B -2 3 +3 -12
C -0.2 1.3 +3 -5
D +0.0 1.0 +3 -1.7 ,

A
To perform the calculations,we have consideredthree standard
atmospheretypes: midlatitudewinter (Ts = 290 K), midlatitude Band
summer(Ts = 300 K), and tropical(T• = 310 K). In addition,we
have usedthe emissivitylaboratoryspectraprovidedby Salisburyand Figure 8. Comparisonof the resultsobtainedin the analysis
DMria [1992, 1994] for severalsoil and rock typesand consideredthe of the three studied algorithms. For the vegetation cover
vegetationas a gray body with e = 0.985. methodwe showboth the optimal and the realisticcases.
11,162 CASELLES ET AL.' EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS

Table 17. Conditionsof ComparisonTestsWith Their AssociatedAssumptions


Random Error, Auxiliary Main Sources of
Method % Information SystematicError

Alpha coefficients ___


1.7-5 one radiosonde Wien's approximation
removal reflection term
TISI _+1.7-3 two radiosondes(day and night) removal reflection term
Sun radiance at surface level coregistration
VCM _+0.5-1.4 vegetation cover vegetationcover

VCM, vegetation cover method.

these sourcescan have an important influence in the final The main ideasto be highlightedare as follows:
result.Furthermore, it is very difficultto assess
their combined 1. Similar results are obtained in both cases in bandsA and
effectin a particularcase,where they can compensate partially D. ChannelsB and C showa significantimprovementin the
or add their contributions,being in any caseuncontrolledfac- case of soils.
tors. The P v estimate in the VCM may be also a potential 2. If we compare these values with the results given in
sourceof systematicerror. Table 17 summarizesthis analysis. Tables4 and 5, we observethat the error increases by approx-
imately a factor of 2-3 in bandsA and B, a factor of 2 in
5.5. VCM in a Realistic Case
channelC, and finally a factor of 1.5-2 in band D. This in-
Once seen that the VCM could be the most suitable method creases the differences between these bands.
to estimatethe LSE, we can have a look at its possibilitiesin 3. Emissivitycan be mappedwith an accuracyof _+3%in
casesnot so favorableas thoseusedin the comparison.In the band A, _+2-3% in band B, _+0.7-1.4% in channel C, and
previousstudywe have consideredthe best conditionsof ap- +0.6-0.9% in channel D. We have shown these results in
plication, i.e., to know preciselythe soil and vegetationemis- Figure8. Even in thisworsecase,we obtaincomparableresults
sivitiesmeasuredin the field with the minimum experimental with those obtained in the other methods. In bands C and D it
error (_+0.005in emissivity).Actually,in manycases,emissivity is still possibleto retrieve LSE with a reasonabledegreeof
will present larger errors. This will be the case,for instance, error around_+1% (takinginto accountthe conditionsof ap-
whenwe are analyzingan area with differenttypesof soilsand plication).
vegetation(not just one type)' the error in assigningone emis-
sivityvaluewill increaseaccordingto the variabilityin the pure
elements. Another case will be that in which we have no in- 6. Summary and Conclusions
formation about emissivityat all. In this casewe can only use In this paper we have addressedthe problem of the emis-
"general mean values" obtained from laboratory and field sivity-temperatureseparationin the radiancesmeasuredby a
measurements,carried out by different investigatorsaround thermalradiometer,findingout whichmethodcanbe the most
the world, and assumethe "variability" presentedby these suitableone to map the LSE, in order to useit later in the LST
measurements as errors. retrieval.A first analysishasled us to selectthree candidatesto
To obtain an idea of how the VCM works in suchconditions, be studied:the TISI method,the alphacoefficients algorithm,
we have recalculatedthe errors on emissivityusingthe same and the VCM. We haveperformeda detailedanalysisof them
mean valuesbut usingthe dispersionsshownby the samples by means of an error analysiscarried out under the same
givenby SalisburyandD'Aria's measurements (seeTable 3) as generalphilosophyand by identifyingthe conditionsof appli-
errorson •a and• •,.We mustremarkthatwearenotestimat- cability of each one.
ing here systematicerrors.We are only testingthe VCM in a We have discussedthe applicabilityof thesealgorithms.It
situation which is not optimal but can be more usual. The seems that the VCM shows the most favorable characteristics
resultscanbe seenin Tables18 (rockscase)and19 (soilscase). regardingthe error propagation.In addition,it doesnot need

Table 18. Error on Effective EmissivityRetrieved With Table 19. Error on EffectiveEmissivityRetrieved With
Vegetation Cover Method in "Realistic" Case of Vegetation Cover Method in "Realistic" Case of
Application When we Have "Rocks" as Application When we Have "Soils" as
Backgroundand 50% Vegetation Cover Backgroundand 50% Vegetation Cover
Rocks Soils
(8P•,/P•,), (8P•,/P•,),
% A B C D C+D % A B C D C+D

5 3 3 1.3 0.8 1.1 5 3 1.8 0.7 0.6 0.7


10 3 3 1.3 0.9 1.1 10 3 1.9 0.7 0.6 0.7
15 3 3 1.4 0.9 1.2 15 3 2 0.7 0.6 0.7
20 3 3 1.5 0.9 1.2 20 4 2 0.7 0.7 0.7
25 3 3 1.5 0.9 1.2 25 4 2 0.7 0.7 0.7

Mean 3 3 1.4 0.9 1.2 Mean 3 1.9 0.7 0.6 0.7

The relative error in emissivityis shownin terms of severalrelative The relative error in emissivityis shownin terms of severalrelative
errorsin the vegetationcoverestimate.The meanvaluesare shownin errors in the vegetationcoverestimate.The mean valuesare shownin
the last row. the last row.
CASELLES ET AL.: EMISSIVITY-TEMPERATURE SEPARATION ALGORITHMS 11,163

the use of radiosondes to correct the thermal radiances be- P,, vegetationcover percentage.
causeit doesnot use them. The necessityof radiosondesis an R•A Sunradianceat groundlevelin channelA.
important condition in order to apply the TISI and alpha S separationbetweenplants.
algorithms, and in fact, it is one of their main sourcesof T or Ts surfacetemperature.
random error. However, this same reason implies that the Tg groundtemperature.
resultsobtainedby the VCM must be introducedinto a split- Ti brightnesstemperaturein channeli.
window method to perform the atmosphericand emissivity T,* ground-level
temperature
in channeli.
corrections. T d day temperature.
On the other side, the VCM has no significantsource of T,, night temperature.
systematicerror. However, we have seen that the TISI and
alphacoefficientmethodsshowsomeimportant systematicer- ai alpha coefiScientfor channel i.
rors,whichinevitablywill affectthe resulteven in the optimal /3, slopeof the power approximationrelationship.
conditions.These errorsare linked to the approximationscar- de, reflectionterm of the effectiveemissivityfor
ried out in the theoreticaldevelopment.Furthermore,they are channel i.

not negligible,especiallyin humid conditions.A recalculation 8x or 8(x) error on the x variable.


of the VCM errors has been made in a more realistic(not A ai variation in the alpha coefiScients
due to Wien's
optimal) case,where lessinformationcan be available.Even in approximation.
this worse situation, the VCM is capable of obtaining the A e emissivityspectraldifferencebetweentwo bands.
emissivityestimatewith at least the same error as the other •, effectiveemissivityin channeli.
methodsin their optimal conditions.In bands C and D it is •o directterm of the effectiveemissMtyin channeli.
possibleto retrieve emissivitywith an error of about _+1%, e,g groundemissivity
in channeli.
whichis still acceptable.This demonstratesthat in anycasethe •,,, vegetationemissivityin channeli.
spectralregion in which it could work better is the 10-12 X• channeli wavelength.
spectralinterval (coveredby channelsC and D), where the • channel i wavenumber.
lowest errors are obtained. 0 observationzenith angle.
The VCM could work better with radiometers with few 0x Sun zenith angle.
bands, such as the PRISM and obtain accurate LSTs. However 2
O-oz
! varianceof alphacoefiScients.
the resultsmight be very differentwith multispectralinstru- ß•(0) atmospherictransmittancein channeli for an
ments such as ASTER or TIMS. The VCM is not able to observationangle 0.
extractinformationfrom multispectraldata becauseit doesnot
usethe thermal radiances.However,the TISI and alpha algo- Acknowledgments. This study was carried out under the ESA/
rithmscan be usedfor this purposeand can be very usefulfor ESTEC contract11402/95/NL/CN.We wishto expressour gratitudeto
geologicalstudies. the EuropeanSpaceAgency,especiallyto U. del Bello and P. Kealy.
In consequence,we will assumethe use of the VCM in the J. W. Salisbury(JohnsHopkinsUniversity,USA) is alsoacknowledged
LSE retrieval and apply the obtainedresults(optimal case) for providingemissivitylaboratoryspectra.We alsothank the anony-
mousreviewersfor their usefulsuggestions,which have contributedto
both to the theoreticalanalysisof the LST retrieval algorithm improve the paper. The financial support suppliedby the Spanish
and to the simulationscarriedout to obtain the optimal band Ministeriode EducaciOn y Ciencia(grantsof E. Valor and E. Rubio)
configurationfor the PRISM sensor.This work will be pre- and CICyT (ProjectAMB94-1208) is alsoacknowledged.
sented in detail in the two following papers of this series
References
[Caselieset al., 1997a,b].
Anton, Y. A., and Y. K. Ross,Emissivityof a soil-vegetationsystem,
Soy.J. RemoteSens.,7(5), 859-869, 1990.
Notation Asrar, G. (Ed.), Theoryand Applicationsof Optical RemoteSensing,
John Wiley, New York, 1989.
a t TISIs definition coefiScientfor channel i. Becker, F., and B. J. Choudhury,Relative sensitivityof normalized
B i Planck functionweighted in channeli. differencevegetationindex(NDVI) andmicrowavepolarizationdif-
Bi(T,*.) ground-level
radiancein channeli. ferenceindex(MPDI) for vegetationand desertificationmonitoring,
Remote Sens.Environ., 24, 297-311, 1988.
c slopeof the.•-variance
relationship
(alpha Becker,F., and Z.-L. Li, Temperature-independent
spectralindicesin
coefiScients). thermal infrared bands, Remote Sens.Environ., 32, 17-33, 1990a.
C1 first Planck radiation constant. Becker, F., and Z.-L. Li, Towardsa local split-windowmethod over
C2 second Planck radiation constant. land surfaces,Int. J. Remote Sens.,11,369-394, 1990b.
F side-groundshapefactor. Caselies,V., and J. A. Sobrino, Determination of frosts in orange
groves from NOAA-9 AVHRR data, Remote Sens. Environ., 29,
F' side-sideshapefactor. 135-146, 1989.
f•4(0•, 0) bidirectionalangularform factor. Caselies,V., J. A. Sobrino,and C. Coll, A physicalmodel for inter-
G ground-sideshapefactor. preting the land surfacetemperature obtained by remote sensors
H height of plants. over incompletecanopies,RemoteSens.Environ.,39, 203-211, 1992.
Caselies,V., C. Coll, E. Valor, and E. Rubio, Thermal band selection
J exponent
of the•-variancerelationship
(alpha for the PRISM instrument,II, Analysisand comparisonof the ex-
coefiScients). isting atmosphericand emissivitycorrectionmethods,J. Geophys.
Li'l'at
m downwelling
atmospheric
radiance
in channel
i. Res., in press, 1997a.
M TISIs constantdefined as a combinationof •i Caselies,V., E. Rubio, C. Coll, and E. Valor, Thermal band selection
coefiScients. for the PRISM instrument,III, Optimal band configuration,J. Geo-
phys.Res.,in press,1997b.
n• exponentof the power approximation Coll, C., V. Caselies,J. A. Sobrino,and E. Valor, On the atmospheric
relationshipcorrespondingto channel i. dependenceof the split-windowequationfor land surfacetemper-
Px side proportion. ature, Int. J. Remote Sens., 15, 105-122, 1994a.
11,164 CASELLESET AL.: EMISSIVITY-TEMPERATURESEPARATIONALGORITHMS

Coll, C., V. Caselies,and T. J. Schmugge,


Estimationof land surface els, in Proceedings
of the 6th InternationalSymposium on Physical
emissivity
differences
in the split-window
channels
of AVHRR, Re- Measurements and Signaturesin RemoteSensing, ESA, 749-758,Eur.
mote Sens.Environ., 48, 127-134, 1994b. SpaceAgency,Val d'Isere,France,1994.
Colton,A. L., Effectivethermalparametersfor a heterogeneous land Olioso,A., Simulating
the relationship betweenthermalemissivityand
surface,RemoteSens.Environ.,57, 143-160, 1996. the NormalizedDifferenceVegetationIndex,Int. J. RemoteSens.,
EuropeanSpaceAgency(ESA), TheNine Candidate EarthExplorer 16(16), 3211-3216,1995.
Missions,coordinatedby C. J. Readingsand M. L. Reynolds, Prata, A. J., Land surfacetemperaturesderivedfrom the Advanced
ESA-SP1196,ISBN 92-9092-383-0,ESA Publ.Div., ESTEC, Noord- Very High ResolutionRadiometerandthe AlongTrackScanning
wijk, Netherlands,1996. Radiometer,I, Theory,J. Geophys. Res.,98, 16,689-16,702,1993.
Gillespie,
A. R., Lithologic mapping of silicaterocksusingTIMS, in Rouse, J. W., R. H. Haas, J. A. Schell, D. W. Deering, and J. C.
The TIMS Data User'sWorkshop, pp. 29-44, JPL Publ. 86-38, Jet Harlan, Monitoringthe vernal advancement of retrogadationof
Propulsion Lab., Pasadena,Calif., 1985. naturalvegetation,Type III, final report, 371 pp., NASA/GSFC,
Gillespie,
A. R., A. B. Kahle,andR. E. Walker,Colorenhancement of Greenbelt, Md., 1974.
highly correlated images,I, Decorrelation and HSI contrast Rubio, E., V. Caselies,and C. Badenas,Emissivitymeasurementsof
stretches,RemoteSens.Environ.,20, 209-235, 1986. severalsoilsandvegetationtypesin the 8-14 tzmwaveband, Anal-
Hook, S. J., A. R. Gabell,A. A. Green,andP.S. Kealy,A comparison ysisof twofieldmethods,RemoteSens.Environ.,in press,1997.
of techniques informationfrom thermal Sabol,D. E., J. B. Adams, and M. O. Smith, Quantitativesubpixel
for extractingemissivity
infrareddatafor geologicstudies,RemoteSens.Environ.,42, 123- spectraldetectionof targetsin multispectralimages,J. Geophys.
135, 1992. Res., 97, 2659-2672, 1992.
Huete, A. R., A soil-adjusted vegetationindex(SAVI), RemoteSens. Salisbury, J. W., andD. M. D'Aria, Emissivity of terrestrialmaterials
Environ., 25, 295-309, 1988. in the 8-14 tzm atmosphericwindow,RemoteSens.Environ.,42,
Ichoku,C., and A. Karnieli,A reviewof mixturemodelingtechniques 83-106, 1992.
for subpixel landcoverestimation, RemoteSens.Rev.,13, 161-186, Salisbury, J. W., andD. M. D'Aria, Emissivity of terrestrialmaterials
1996. in the 3-5 tzm atmospheric window,RemoteSens.Environ.,47,
Kahle, A. B., D. P. Madura, and J. M. Soha,Middle infraredmulti- 345-361, 1994.
spectralaircraftscannerdata:Analysisfor geologicapplications, Soha,J. M., and A. A. Schwartz,Multispectralhistogramnormaliza-
Appl. Opt., 19, 2279-2290, 1980. tioncontrastenhancement, in paperpresentedat the Fifth Canadian
Kaufman,Y. J., and D. Tanrd, Atmospherically resistantvegetation Symposium on RemoteSensing, Victoria,BritishColumbia,Can-
index(ARVI) for EOS-MODIS,IEEE Trans.Geosci. RemoteSens., ada, 1978.
30, 261-270, 1992. Stoll,M., Land surfacetemperatureand emissivity retrievalfrom pas-
Kealy,P.S., andS.J. Hook,Separating temperature andemissivityin siveremotesensing measurements, inProceedings of the3rdThermal
thermalinfraredmultispectralscannerdata:Implicationsfor recov- InfraredMultispectral Scanner(TIMS), pp. 10-19, Jet Propul.Lab.,
eringlandsurfacetemperatures, IEEE Trans.Geosci. RemoteSens., Pasadena, Calif., 1991.
3•(6), 1155-1164,1993. Sutherland,R. A., and J. F. Bartholic, Significanceof interpreting
Labed,J., and M.P. Stoll,Angularvariationof landsurfacespectral thermal radiation from a terrestrial surface,J. Appl. Meteorol.,16,
emissivity in thethermalinfrared:Laboratory investigations
onbare 759-763, 1977.
soil, Int. J. Remote Sens.,12, 2299-2310, 1991. Valor, E., and V. Caselies,Mapping land surfaceemissivityfrom
Li, Z.-L., and F. Becker,Propertiesand comparisonof temperature- NDVI, Applicationto European,AfricanandSouth-American ar-
independentthermal infrared spectralindiceswith NDVI for eas,RemoteSens.Environ.,57, 167-184, 1996.
HAPEX data, Remote Sens.Environ., 43, 67-85, 1990. van de Griend, A. A., and M. Owe, On the relationshipbetween
Li, Z.-L., and F. Becker,Feasibilityof land surfacetemperatureand thermalemissivity
and the normalizeddifferencevegetationindex
emissivity determinationfromAVHRR data,RemoteSens.Environ., for natural surfaces,Int. J. RemoteSens.,14, 1119-1131, 1993.
43, 67-85, 1993. Watson,K., Spectralratio methodfor measuringemissivity,
Remote
Masuda,K., T. Takashima,and Y. Takayama,Emissivityof pure and Sens.Environ., 42, 113-116, 1992a.
seawatersfor the model seasurfacein the infrared windowregions, Watson,K., Two-temperature
methodfor measuringemissivity,
Re-
Remote Sens.Environ., 24, 313-329, 1988. mote Sens.Environ., 42, 117-121, 1992b.
Maul, G. A., and M. Sidran,Atmosphericeffectson oceansurface
temperature
sensing
fromtheNOAA satellitescanning
radiometer, V. Caselies,C. Coll, E. Rubio, and E. Valor, Departamento de
J. Geophys.
Res.,78, 1909-1916,1973. Termodinhmica,Facultat de Ffsica,Universitatde Valencia, 50 Doctor
of landsur- Moliner,E-46100Burjassot,
Nerry,F., J. Labed,and M.P. Stoll,Spectralproperties Spain.(e-mail:enric.valor@uv.es)
facesin the thermalinfrared,2, Field methodfor spectrallyaveraged
emissivity
measurements,
J. Geophys.
Res.,95, 7027-7044,1990.
Norman, J. M., S. Castello,and L. K. Balick, Directional infrared (ReceivedOctober18, 1996;revisedJanuary10, 1997;
temperature
andemissivity
of vegetation:
Measurements
andmod- acceptedJanuary11, 1997.)

You might also like