2. Consistent scene identification and anisotropic factors A. Coakley

advertisement
JOURNALOF GEOPHYSICALRESEARCH,VOL. 101,NO. D16,PAGES21,253-21,263,
SEPTEMBER27, 1996
Biasesin Earth radiation budget observations
2. Consistentsceneidentificationand anisotropicfactors
QianYe• andJamesA. CoakleyJr.
Collegeof OceanicandAtmospheric
Sciences,
OregonStateUniversity,Corvallis
Abstract.Simplethreshold
sceneidentification
methodsaredeveloped
to reducetheeffects
of errorsin sceneidentificationon the anisotropyof reflectedandemittedradiances
inferredfrom EarthRadiationBudgetExperiment(ERBE) scannerobservations.The
ERBE maximum likelihoodestimate(MLE) sceneidentificationis assumedto be accurate
for nadirfieldsof view. Variouscombinations
of neighboring
ERBE scannerfieldsof
view at nadirareusedto determinethepopulationof cloudscenetypesasa functionof
field of view size. Longwaveand shortwavethresholdsarethendeterminedfor eachof the
ERBE solarzenith,satelliteview zenith,andrelativeazimuthangularbinssothatthe
populationof cloudscenetypesat a particularsatelliteview zenithangleis consistent
with
the field of view sizeat theparticularsatelliteview zenithangle. Differencesbetweenthe
anisotropy
of reflectedsunlightandemittedlongwaveradiationobtained
usingthenew
sceneidentificationmethodandthatobtainedusingthe ERBE MLE methodshowthatthe
ERBE radiativefluxeshavesatelliteview zenithangledependent
biases.Thresholds
are
alsodevelopedfor cloudsceneidentificationwith fieldsof view thatareconstructed
to
havea constantsizewith satelliteview zenithangle. The angulardependence
of reflected
sunlightandemittedlongwaveradiationfor scenesidentifiedwith thesethresholdsshow
littledependence
onfield of view size. Thislackof dependence
is a necessary
condition
for usingscanningradiometerdatato obtainradiativefluxes.
1. Introduction
In part 1 [Ye and Coakley, this issue] a potentialview
zenith angle dependentbias in Earth radiation budget
observations,
like thosemadeby the Earth RadiationBudget
Experiment(ERBE), was described. In ERBE the radiative
fluxes were derived by first identifyingwithin each scanner
field of view thecloudscenetype: clear,partlycloudy,mostly
cloudy,or overcast. The identificationwas made on the basis
of the shortwaveand longwaveradiancesobservedfor the
field of view [Wielicki and Green, 1989]. An anisotropic
factorassociatedwith that scenetypewas thenusedto convert
the radiancesto radiative fluxes. The view zenith angle
dependentbias arises becauseof (1) view zenith angle
dependenterrors in sceneidentification,(2) the nonuniform
spatialdistributionof clouds,and(3) thegrowthof the field of
surroundedby fields of view that contain broken clouds and
was thus likely to have more cloud contaminationthan a field
of view identified as clear at the limb.
At the limb the field of
view is large. It was shownthat large regionsidentifiedas
beingclearwereoftensurrounded
by regionsthat werealso
clear.Consequently,
largeregionswerelikelyto be lesscloud
contaminated.Evidently,theselargeclearfieldsof view were
partsof vastregionsthat werecloudfreethroughout.Similarly,overcastscenesat nadirweremorelikely thanovercast
scenesat the limb to be contaminated
by breaksin the cloud.
Becauseclear scenesreflectsunlightand emit radiationthat
variesmorestronglywith anglethanthatfor overcastscenes,
theeffectof the clustering
of cloudson certainspatialscales
andtheERBEinversion
method
causes
clearscenes
to appear
less anisotropicthan they should and overcastscenesto
view size from nadir to limb.
appearmoreanisotropic
thantheyshould.Consequently,
the
anisotropy
of
the
radiative
fields
obtains
an
apparent
link
to
The nonuniformspatialdistributionof cloudscontributesto
the bias becauseregionsthat are the size of an ERBE scanner the sizeof the regionbeingviewed. Becausethe field of view
field of view at nadir-(40 km)2 haverelatively
large size for the ERBE scannergrowsfrom nadir to limb, the
viewzenithangle
frequencies
of clearand overcastconditionsat the expenseof ERBEfluxesarelikelyto havea satellite
relativelyinfrequentoccurrences
of partly cloudyand mostly dependentbias.
In part 1 such a bias was found for clear oceanscenesbut
cloudyconditions.Largerregionshavesmallerfrequencies
of
clearand overcastconditionsand largerfrequenciesof partly not for the otherscenetypes. Instead,it was foundthat the
cloudyand mostlycloudyconditions.In part 1 it was shown ERBE maximumlikelihoodestimate(MLE) sceneidentificathat a scene identified as being clear at nadir was often tion method[Wielickiand Green, 1989] oftenincorrectly
identified
partlyandmostlycloudyscenes
at thelimbasbeing
overcast.The dependence
of the frequency
of scenetypeon
SNow
atCIRES,
University
ofColorado.
Boulder.
Copyright1996 by the banericanGeophysical
[Jnion.
field of view size was determined from observations for the
nadir fieldsof view by combiningthe fieldsof view to obtain
observationsfor regionsof varioussizes. The ERBE scene
Papernumber96JD01157.
identification scheme was assumed to be correct for the nadir
014õ-0227/96/96JD-01
fieldsof view. The frequency
of overcastscenesat the limb
! 57509.00
21,253
21,254
YE AND COAKLEY: RADIATION BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
was foundto be largerthanexpectedgiventhesizeof the limb evidence that the new scene types have the cloud cover
fieldof view. Conversely,
thefrequencies
of partlycloudyand associated
with the ERBE scenetypes: clear,05Ac<0.05;
mostlycloudyscenesat the limb weresmallerthanexpected. partlycloudy,0.05<Ac<0.5; mostlycloudy,0.5<_Ac<
0.95;
Becauseof the incorrectidentificationthe anisotropyof andovercast,
0.955Ac51. Nevertheless,
thethresholds
were
reflected and emitted radiation could not be determined for
partlycloudy,mostlycloudy,andovercastscenes.
Herethe ERBE sceneidentification
errorsfoundin part 1
areavoidedby developing
a newmethodfor identifyingcloud
scene types from the ERBE scannerradiances. The new
methodproducespopulations
of scenetypesat eachsatellite
view zenith angle that are more consistentwith the field of
view sizeat the particularangle. The new methodis basedon
simple thresholdsappliedto the shortwaveand longwave
radiancesobtainedwith the ERBE scanner. Anisotropic
factorsfor reflectedsunlightand emittedradiationare derived
based on the radiances obtained with this new scene identifi-
cation. Comparisonsof the anisotropicfactors with those
consistentwith the featuresof the ERBE MLE method.Larger
shortwave reflectivities and smaller emitted longwave
radianceswere associatedwith largercloudcoverfractions.
In addition to the FFOV
threshold method a second set of
thresholdswas developedusing the "constantsize field of
view" (CFOV) observationsintroducedin part 1. CFOV
observations
were constructed
by groupingneighboringERBE
scannerfields of view so that the size of the region being
viewed was almostinvariantwith satelliteview zenith angle.
For the CFOV observations, longwave and shortwave
thresholdswere adjustedso that the populationof scenetypes
within an ERBE satelliteview zenithanglebin was equivalent
to the population obtained for the CFOV observationsat
derived with the ERBE MLE scene identification method
show the effects of scene identification errors and cloud
nadir.
clustering
on the angulardistribution
of reflectedsunlightand
emittedlongwaveradiation. The anisotropicfactorsderived
with the new sceneidentification
are shownto be relatively
insensitiveto field of view size, a necessaryconditionfor
obtainingradiativefluxes from scannerobservations
using
methodslike thoseemployedby ERBE.
cloud cover from nadir to limb found with the ERBE
2. Analysis Methods
The ERBE MLE scene identification method was assumed
Like the FFOV threshold identification method, the
CFOV identificationmethodremovesthe growthof apparent
MLE
scene identification scheme. In addition, the CFOV observa-
tions reduceeffectsdue to the spatial scaleson which clouds
congregate.Becauseof their constantsize, the CFOV observationsshouldbe subjectto relativelyconstantcloudand clear
contaminationat all satelliteview zenithangles.
Figure1 illustratesthe procedures
usedto derivethe longwave and shortwaveradiancethresholdsfor a particular
satelliteview zenith angle. The figureillustratesthe division
of the longwave-shortwaveradiance pair domain for a
particularSun-scene-satellite
geometry.The crosses
give the
to providethe correctidentificationfor fields of view at nadir.
The dependence
of thepopulationof cloudscenetypeson field
of view size was determinedby simulating nadir fields of
means of the longwave and shortwaveradiancesassociated
view
used in the ERBE
of the sizes that matched
those associated
with
the
with clear,partlycloudy,mostlycloudy,and overcastscenes
MLE
scene identification method.
The
ERBE satellite view zenith angle bins. The required ellipses are meant to illustrate the standarddeviationsand
groupingswere describedin part 1. The averagesof the linear correlationsfor the longwaveand shortwaveradiances
shortwaveand longwaveradiancesfor the scannerfields of based on the values used in the ERBE MLE scene
view that contributed to a simulated field of view were taken
identification
method.The figureshowsa hypothetical
casein
to be the radiances for the simulated field of view. The scene
which the ERBE MLE methodincorrectlyidentifiessome
type was basedon the nominal cloud coverderivedfrom the mostlycloudyfields of view as overcastand someclear fields
cloud scenetypes of the contributingscannerfields of view.
The assignmentof scenetype wasalsodescribedin part 1.
As in part 1, the observationswere from the Earth
RadiationBudget Satellite(ERBS) scannerfor the monthsof
Hot
September,October,and November 1986. Only oceanand
overcastscenetypeswere considered,as the purposeof this
study was to determinewhetherthe anisotropyof the radiances obtained with the new scene types showedreduced
sensitivityto field of view size. While the range of scene
typeswas limited, theyconstituted
approximately
70% of the
observations
obtained
with
the ERBS
scanner.
Thus
the
findings presentedhere are likely to prove relevant for all
scene types.
The field of view size dependenceof the populationof
scenetypeswas usedto derivetwo setsof thresholds.For the
first set, thresholdswere adjustedso that the populationof
scene types associatedwith the field of view size at a
particularanglewereidenticalto thatfor the samesizefield of
view
at nadir.
These thresholds
are referred to as the "full
resolution field of view" (FFOV) thresholds. Since the
Cold
Dark
Bright
Shortwave Radiance
thresholds were based on the ERBE scene identification
Figure 1. Diagram illustratingproceduresusedto obtainthe
results,the scenetypesderivedby usingthe thresholdswere longwave and shortwave radiance thresholds based on
takento be the sameas their ERBE counterparts:clear,partly frequenciesof occurrencederived from simulated nadir
cloudy,mostly cloudy, and overcast. There is, however,no observations.
YE AND COAKLEY: RADIATION BUDGET,CONSISTENTSCENEID AND ANISOTROPY
SZA=25.8-36.9
VZA=O.O-15.0
initial shortwavethresholdwas arbitrarily set at the mean
AZA-9.0-30.O
120
100
oo•½•OOo/O
:/' ' ß
:--..... •ooo•',o/*.,
?'
,
60
**
**:**
*
**
,
•. * , %4--
,•*
J*
,
.•
+
i
ß
*
+
*
ß ,
+4-
ß
_ __ •. -
,
*
_
++ +
+
+
,
+
i
+
+
+
40
CFOVTHRESHOLD
i
/
deviations
of
the
shortwave
sr-• fortheovercast-mostly
cloudy
boundary
and0.1 Wm-2
sr-• for the clear-partly
cloudyand partlycloudy-mostly
+
i
'
standard
thresholdwas reduced(moveddownwardin Figure 1), and
the initial shortwaveradiancethresholdwas increased(moved
right in Figure 1). The ratio of the changein the longwave
radiance to the change in the shortwaveradiance was set
equal to the slope of the longwave-shortwaveradiance
relationship derived from a linear leastsquaresfit for the
scannerfields of view identified as being overcastby the
ERBE MLE sceneidentificationmethod. The step size used
for the shortwaveradiancethresholdwas takento be 1 Wm-2
*
o c9•,• * •/
* •** ,*
o /o, /;_L..,•.?__.,_
,. .... ,___,_.___,.*_
.....................
,* **
two
fields of view as overcast,then the initial longwaveradiance
+OVERCAST
,
?, • ,.
O_o?
....
o •
minus
radiances.If basedon the populationof scenetypesexpected
/"'/ ' ' x•Ct'EAR'SKY'
'
for the size of the field of view at the particularsatellite
,' /
o
PARTLY
CLOUDY
/ /
, MOSTLY
CLOUDY viewingzenithangle,the initial thresholds
identifiedtoo many
••oøøo
,' /
'•-•%•;'o•
•,,,/
- •,•
80
value
' I.....
J
• J,'•
21,255
;
cloudy boundaries. Similar adjustmentswere made for the
thresholdsseparatingthe clear and partly cloudy fields of
--- FFOVTHRESHOLD
100
SHORTWAVE
200
300
view.
After the radiance boundaries for overcast and clear fields
RADIANCE
Figure 2. An exampleshowingthe effectof full resolution
field of view (FFOV) and constantsizefield of view (CFOV)
thresholdson sceneidentification. The arrows in the figure
pointto observations
whichwereidentifiedaseitherclearsky
or overcastby the Earth Radiation Budget Experiment
(ERBE) maximumlikelihoodestimate(MLE) sceneidentificationbut were reidentitledby the CFOV thresholdmethod.
Radiances
arein Wm-2sr-•.
of view were determined,the boundaryseparatingpartly and
mostly cloudy fields of view was fixed by adjustingboth the
shortwaveand the longwavethresholdsso that the boundary
remained normal to the line connectingthe means of shortwave and longwaveradiancesassociatedwith the partly and
mostly cloudy fields of view identified by the ERBE MLE
scene identification method. The final position of the
boundarywas determinedby the frequenciesof occurrence
obtained from the simulated nadir observations.
of view as partly cloudy. In this case,the populationof
overcastand clear sceneshas to be reducedaccordingto the
Figure 2 showsan exampleof the determinationof FFOV
and CFOV thresholdsfor the ERBE angularbin in which the
solarzenith anglesrange from 25.8ø to 36.9ø, the view zenith
frequenciesof occurrencederived from the simulatednadir
anglesrangefrom 0.0ø to 15.0ø, and the azimuthalangles
rangefrom 9.0ø to 30.0ø. The markingsgive the MLE scene
The following procedureswere usedto obtain the desired identification:
clear (crosses),
partlycloudy(circles),mostly
populationof scenetypes. First, thresholdswere adjustedto cloudy (asterisks),and overcast(pluses). Fields of view
distinguish between partly cloudy and clear scenes and identifiedas either clear or overcastby the MLE methodbut
between mostly cloudy and overcastscenes. For overcast reidentitledby the CFOV thresholdmethodas partly and
scenes,an initial longwavethresholdwas arbitrarilysetat the mostlycloudyare indicatedby arrows. For the angularbin
mean value plus two standarddeviationsof the longwave shown in the figure, as was often the case, the CFOV
radiances
for the fieldsof view identifiedas beingovercastby thresholds are more restrictive than the FFOV thresholds for
the ERBE MLE sceneidentificationmethod. Similarly,an clear and overcast scenes.
observations.
Table 1. Frequencies
of Occurrencefor SimulatedFull ResolutionERBE
Scanner Observations
SatelliteView Zenith Angle
Frequencyof Occurrence,%
Field of
Bin
Angular
Range,
View
Size
Number
deg
104km2
1
2
3
4
5
6
0 - 15
15 - 27
27 - 39
39 - 51
51 - 63
63 - 75
0.16
0.19
0.25
0.40
0.79
1.35
Clear
Partly Mostly
Cloudy Cloudy
Ocean Ocean OceanOvercast
17.2
15.6
13.1
9.6
8.7
7.8
28.6
27.9
32.6
35.8
37.2
37.1
32.5
36.6
36.8
40.9
41.6
43.4
21.7
19.9
17.5
13.7
12.6
11.8
The frequencies
wereobtainedfrom ERBS scannerobservations
for September,
October,
and November 1986. ERBE, Earth RadiationBudgetExperiment;ERBS, Earth Radiation
BudgetSatellite.
21,256
YE AND COAKLEY: RADIATION BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
CLEAR
SKY
50.0
i
•K
o .....
i
PARTLY
CLOUDY
o
20
50.(
i
N<.
FFOV OBSERVATIONS
-o CFOV OBSERVATIONS
37.5
z
u
0
25.0
I,I
12.5
n/
25.•
12.
o
0
MOSTLY
0
5o.o
i
i
i
20
40
60
8O
i
8O
OVERCAST
CLOUDY
i
50.(
i
0
0
37.5
I,I
25.0
25.•
I,
12.5
12.
o
0
20
40
60
80
0
20
40
60
VIEWING ZENITH ANGLE (DEG)
Figure 3. Frequencies
of occurrence
for FFOV and CFOV observations
identifiedusingthe FFOV
thresholds. The resultsare basedon Earth RadiationBudgetSatellite(ERBS) scannerobservations
for
September,
October,andNovember1986.
3. Results
Table 1 lists the frequenciesof occurrence
obtainedfrom
the simulatedfields of view at nadir. The frequenciesof
occurrence decreased for clear and overcast scenes and
increasedfor partlyand mostlycloudyscenesas the field of
not includedin the analysis. As describedin part 1, these
procedures
wereusedto insurethe statisticalindependence
of
each sample. Differencesshown in Figure 3 betweenthe
FFOV and the CFOV frequencies
for the largestsatelliteview
zenith angle are typical of the sampling errors. A second
cause of the residual variation is that the FFOV
thresholds do
viewsizegrewfromthe-(40 km)2 forthesatellite
viewing
zenithanglebin at nadirto -(115 km)2 for thesatellitenot allow for effectsdue to cloudclustering. Scenesidentified
viewingzenithanglebin at thelimb, whichherewastakento
be at 64 ø. Clear and overcast conditions occurred about 20%
with the FFOV thresholdsare likely to have more cloud
contaminationat nadir than at the limb. Consequently,the
CFOV frequenciesfor clear scenesat small satelliteview
Correspondingly,
partly and mostly cloudy conditions zenithanglesare likely to be higherthanthosefor largeview
zenith angles. The same trend is expectedfor the overcast
occurred 60% of the time at nadir but about 80% of the time at
of the time at nadirbut only about10% of thetime at the limb.
scenes. The variations shown in Figure 3 for the CFOV
observations,however, appear to be within errors due to
establish the FFOV and CFOV thresholds described in the
sampling.
previoussection.
Figure4 showsthe frequenciesof occurrence
for the four
Figure3 showsthe frequencies
of occurrence
obtained
by
cloud
categories
identified
with
the
CFOV
threshold
method
applyingthe FFOV thresholds
to FFOV (solidlines) and
for
FFOV
(solid
line)
and
CFOV
(dashed
line)
observations.
CFOV (dashedlines) observations.For the FFOV observa-
the limb. The frequenciesshownin Table 1 were usedto
tionsthefrequencies
of occurrence
showedtheexpected
trends Comparedwith theresultsobtainedwith theFFOV thresholds
with satelliteviewing zenithangle: the frequencies
of clear (Figure 3), almost constantfrequenciesof occurrenceare
and overcastfields of view decreasedwith increasingview
obtained for the CFOV observations. For the FFOV observa-
zenithangleandthe frequencies
of partlyandmostlycloudy tionsthe CFOV thresholdsproducefrequenciesof occurrence
fields of view increased.
For the CFOV
observations the
near nadir for clear ocean and overcast fields of view that are
frequencies
for all scenetypeswererelativelyconstant.The smaller than thoseproducedby the FFOV thresholds. The
variationsin thefrequencies
werecausedin partby sampling. smallerfrequenciesare expectedbecausefor clear and overThe frequencies
to whichthe thresholds
wereadjusted(Table cast scenes the CFOV thresholds are often more restrictive
1) were obtainedusingall fields of view. The thresholds, than the FFOV thresholds,as was, for example, the case
however,were obtainedfor fieldsof view that were separated shown in Figure 2. Scenesnear nadir, which the FFOV
by at least40 scanlinesfor a givenscenetype,solarzenith, thresholdsidentify as clear or overcast,are oftenin the midst
satellitezenith,andrelativeazimuthangularbin. In addition, of broken clouds. With the CFOV thresholds, some of these
bins for which therewere fewer than eight observations
were fieldsof view areidentifiedaspartlyor mostlycloudy.
YE AND COAKLEY: RADIATION BUDGET, CONSISTENT SCENE IX) AND ANISOTROPY
CLEAR OCEAN
PARTLY CLOUDY OCEAN
50.0
50.0
37.5
37.5
25.0
25.0
12.5
12.5
-
e, .....
o ....
0
i
i
i
20
40
60
MOSTLY CLOUDY
50.0
I
0
80
0
i
i
i
2O
4O
6O
8O
OVERCAST
OCEAN
50.
i
:•
• FFOV
OBSERVATIONS
CFOV
OBSERVATIONS
<,---•
37.5
37.
25.0
25.
12.5
12.
>z
I
I
I
2O
4O
6O
0
8O
0
I
I
I
2O
4O
6O
8O
VIEW ZENITH ANGLE (DEGREE)
Figure4. Frequencies
of occurrence
forFFOV andCFOV observations
identified
withtheCFOV threshold
identificationmethod. The resultsare basedon ERBS observations
for September,October,and November
1986
SZA-25.8MLE 20
I
36.9
FFOV THRESHOLD
I
I
SW MEAN
I
I
x CLEAR
MLE i
•KY I
o PARTLY CLOUDY
-
i
k,
"o
_ ,,' '-.,.•, '..•
-
•,/ /
/
'"C'.,.
,,. .....
..,
'.•.
"...,
_
w
L•
i
ß
,, '
_
_
n•
i
•..-""...
,4.
-
w
i
+ OVERCAST
%/'
w
i
_
* MOSTLY CLOUDY
-
FFOV THRESHOLD
i
LW MEAN
I
-20
I
I
o
w
t
_
I
I
I
80
40
MLE 20
z
I
CFOV
I
I
THRESHOLD
I
I
I
I
I
I
o
4o
MLE -
CFOV
THRESHOLD
I
- LWMEAN
SW MEAN
_
-
w
t•j,'•
, /
i i'
,5 -
/',,/
I/
/ •---- --•/
//.'/
,0
-20
80
I
o
I
I
I
4o
I
I
VIEW ZENITH ANGLE (DEGREE)
Figure 5. Percentdifferences
in the meanradiancesfor scenesidentifiedby the ERBE MLE, FFOV, and
CFOV methods.The meansobtainedusingthethreshold
methodsweresubtracted
fromthoseobtainedusing
the ERBE MLE method.ERBS scannerobservations
for September,
October,andNovember1986 wereused.
The resultspresented
in thefigureareazimuthallyaveraged.Theyarefor solarzenithanglesbetween25.8ø
and 36.9 ø.
21,257
21,258
YE AND COAKLEY: RADIATION BUDGET, CONSISTENTSCENEID AND ANISOTROPY
Clearly, the mean radiances, standard deviations,and
correlationsbetweenshortwaveand longwaveradiancesfor
the scene types identified with the FFOV and CFOV
0.70
thresholds differ from those obtained with the ERBE
0.65
MLE
sceneidentification. Figure 5 shows an example of the
percentdifferencesin the meansof the shortwaveand longwave radiances.The differencesare obtainedby subtracting
the radiancesobtainedusingthe thresholdmethodsfrom those
obtainedusing the ERBE MLE sceneidentificationmethod.
For scenesidentified as clear by the FFOV thresholds,the
0.60
means of the reflected shortwave
0.55
radiances are smaller than
thosefor the scenesidentifiedas clearby ERBE MLE method.
The correspondingmeans of the longwave radiancesare
larger. Conversely,for scenesidentified by the threshold
.......
*
• FFOV THRESHOLD
o .........
.e CFOV THRESHOLD
0.50
methods as overcast, the means of the reflected shortwave
radiancesare largerand the meansof the longwaveradiances
are smaller. Sincecloudsgenerallyhave higherreflectivities
than the oceanbackgroundand are at lower temperatures,
the
clearfieldsof view identifiedby the thresholdmethodswould
appearto have lesscloudcontamination
than thoseidentified
by the ERBE MLE method. Likewise, the overcastscenes
would appearto be lesscontaminated
by breaksin the cloud.
Thesedifferenceswere, of course,forcedby the designof the
thresholdmethods. The designwas to reducethe growth in
the numberof overcastsceneswith satelliteview zenithangle
obtained with the ERBE MLE
0.45
,
0
I
20
,
I
40
,
I
60
,
80
VIEWZENITHANGLE(DEGREE)
Figure 6. Fractionalcloud cover obtainedwith the ERBE
MLE method(heavysolidline), the FFOV thresholds
(solid
line), andthe CFOV thresholds
(dashedline). Observations
are from the ERBS scannerfor September,
October,and
November
1986.
scene identification method.
Comparedwith the radiancesobtainedwith the FFOV
thresholds, the reflected shortwave radiances obtained with
fractionalcloudcoverderivedfromboththreshold
methods,
the CFOV thresholdsare smallerfor clear and partly cloudy
oceanscenes.The shortwaveradiancesare largerfor mostly
cloudyoceanand overcastscenes. The longwaveradiances
obtainedwith the CFOV thresholdmethodare larger than
usingthenominalcloudfractions
for theERBEscenetype,
wererelativelyconstant
with increasing
view zenithangle
those obtained with the FFOV
comparedwith the increaseobtained with the ERBE MLE
scene identification method.
threshold method for clear
Using the procedures
describedin part 1, FFOV and
oceanscenes. Evidently,clear oceanscenesidentifiedwith
CFOV pseudoanisotropic
factorswereobtainedfor theFFOV
and CFOV thresholds.Figure 7 showsdifferences
in the
anisotropicfactorsobtainedwith the FFOV shortwaveand
longwavethresholds
appliedto the FFOV and CFOV obser-
the CFOV
threshold method are less cloud contaminated than
those obtained with the FFOV thresholds. The reflectivities of
overcast scenes identified with the CFOV
thresholds are
generallyhigherthanthosefor scenesidentifiedby the FFOV
thresholdmethod and the emitted longwave radiancesare
lower. Evidently, the overcastscenesidentified with the
CFOV thresholdmethodare lesscontaminated
by breaksin
vations.The figureshowsthe percentdifferences
(FFOV-
CFOV)forthesolarzenithanglebinswhichhadthelargest
numbers
of satellitezenithandazimuthanglebinsfor which
thedifferences
in theanisotropic
factors
weresignificant
at the
the clouds.
90% confidencelevel (shadedregions). Proceduresfor
Becausethe shortwave-longwave
radiancepairs associated determining
theconfidence
levelswerealsodescribed
in part
with the thresholdscenetypesno longermatch thoseof the 1. Figure 8 shows the same differencesbut for the solar
ERBE scenetypes,the conventional
identities, clear,partly zenithanglebins which had the smallestnumbersof satellite
cloudy,etc., are usedhereonly to simplifythe discussion.No zenithand azimuthanglebins in which the differences
were
claim is made that the thresholdsdid a better job of statistically
significant.Compared
with similarresultsshown
identifyingclear scenesor overcastscenes. The thresholds in part 1 obtained
with the ERBE MLE sceneidentification,
simplydividedthe longwave-shortwave
radiancedomaininto thedifferences
between
theFFOVandtheCFOVanisotropic
regionsin which certainspecifiedpercentages
of the fields of factorswerereduced
bothin magnitude,
whenthedifferences
view were to be found. In making thesedivisions,it was werestatistically
significant,andin the numberof binsfound
assumedthat eachregionof the longwave-shortwave
domain to havestatistically
significantdifferences.For binsin which
was occupiedby a distinctscenetype. This assumption,
while large percentagedifferencesoccurred,the differenceslacked
commonto all thresholdmethods,was not tested. The goal statisticalsignificance. In addition,comparedwith their
was to determinewhether dividing the longwave-shortwaveERBE counterparts,the differences obtained with the
radiancedomain so that the populationof scenetypeswas threshold
methods
appeared
to be morerandomlydistributed
solelya functionof field of view sizewouldleadto an angular in the viewingzenithand azimuthangledomain. Table 2
dependencefor reflected sunlight and emitted longwave givesthepercentage
of thenumberof angularbinswhichhad
radiation that, unlike the ERBE observations, was differenceswhich were significantat the 90% confidence
independent
of field of view size.
level. The percentage
is that of the total numberof bins for
The FFOV and CFOV thresholdswere applied to the which there were at least eight statistically
independent
ERBS scanner observations. Figure 6 shows that the observations
as described
in part 1. Clearly,with theFFOV
YE AND COAKLEY: RADIATION
BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
21,259
PERCENTDIFFERENCE(FFOV - CFOV)
CLEAR
PARTLY
SKY OCEAN
25.8
-
36.9
CLOUDY
84.3
171
g
180
0
51
27
39
0
27
15
15
MOSTLY CLOUDY
66.4-
51
39
63
171
9
180
0
75
90
75
90
51
63
27
0
39
15
53.3-
72.5
0
15
63
75
90
63
75
90
60.0
9O
0
27
39
51
39
OVERCAST
OCEAN
180
51
27
15
9O
75
90 63
90.0
9O
9O
75
90
63
-
OCEAN
27
15
51
39
63
75
90
180
0
75
90
63
51
27
0
39
27
15
15
51
39
Figure 7. Percentdifferences
betweentheFFOV andtheCFOV pseudoanisotropic
factorsfor the solarzenith
anglebinswhichhadthelargestnumberof satelliteview zenithandrelativeazimuthanglebinsfor whichthe
differences
weresignificant.The scenes
wereidentifiedwith theFFOV thresholds
appliedto boththe FFOV
andthe CFOV observations.
The solarzenithanglesfor eachscenetypeare givenin the figure. The radial
axisis for view zenithangle. The polaraxisis for therelativeazimuthangle. The incrementof the contoursis
2.5%. Regionsin whichthedifferences
werepositiveandsignificantat the90% confidence
level are shaded.
Thosethatwerenegativeandsignificantarehatched.
Table 2. Percentageof AngularBinsWhich Had Differencesin theCFOV
andFFOV AnisotropicFactorsthatWere Significantat the90% Confidence
Level
Clear
Partly
Cloudy
Ocean
Ocean
Shortwave
Ocean
Overcast
ERBE MLE
FFOV threshold
CFOV threshold
61
13
3
77
13
15
48
15
12
64
19
14
ERBE MLE
FFOV threshold
59
26
72
35
61
29
57
31
CFOV threshold
22
25
27
23
Scene
Identification
Method
Mostly
Cloudy
Longwave
The percentages
areforERBS scanner
observations
for September,
October,andNovember
1986. CFOV, constantsizefield of view; FFOV, full-resolutionfield of view; MLE, maximum
likelihood estimate.
21,260
YE AND COAKLEY: RADIATION BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
PERCENTDIFFERENCE(FFOV - CFOV)
CLEARSKY OCEAN
84.,3
-
PARTLYCLOUDYOCEAN
90.0
78.5-
9O
9O
171
171
9
180
75
90 63
27
0
39
27
15
MOSTLY
5!
15
CLOUDY
45.6-
39
63
9
180
0
5!
0
75
90
63
75
90
5!
27
39
0
OCEAN
15
53.3
25.8-
180
75
90
36.9
171
0
27
63
9O
9
0
59
5!
39
OVERCAST
171
51
27
15
9O
75
90
65
84.,3
27
15
51
15
39
65
9
180
75
90
0
75
90 63
51
27
39
0
15
27
15
51
39
63
75
90
Figure 8. Sameas Figure 7 but for the solarzenithanglebins which had the smallestnumberof satellite
view zenithand relativeazimuthanglebins for whichthe differences
were significantat the 90% confidence
level.
SZA
CLEAR
1.3
•o
FFOV
e--o CFOV
1.2
-
90.
SKY
PARTLY
i
ß---,
0.0
CLOUDY
i
MODEL
MOD
1.1
i--
o
•
0.9
i
0
MOSTLY
20
40
60
0
80
CLOUDY
20
40
60
i
i
i
20
40
60
80
OVERCAST
Q•
1.5
.3
._1
1.2
.2
1.1
1.0
.0
0.9
0
20
40
60
80
0
80
VIEWING ZENITH ANGLE (DEGREE)
Figure 9. Azimuthallyaveragedshortwave
anisotropic
factorsfor FFOV andCFOV observations
obtained
with the FFOV thresholds.The azimuthallyaveragedanisotropicfactorswere averagedfor all solar zenith
anglebins.
YE AND COAKLEY: RADIATION BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
21,261
PERCENTDIFFERENCE
(FFOV - CFOV)
CLEAR SKY OCEAN
PARTLY CLOUDY
25.8 - 36.9
84.3 - 90.0
90
9O
171
180
171
9
75
90 65
51
59
27
0
15
OCEAN
15
27
59
51
75
9
180
0
0
75
90 65
65 90
51
27
59
0
15
51
59
65
75
90
OVERCAST
MOSTLY CLOUDY OCEAN
72.5-
27
15
78.5-
78.5
84.5
9O
9O
171
180
75
90 65
51
59
27
15
0
15
27
59
51
75
0
65 96
180
75
90 65
9
51
59
27
15
0
15
27
59
51
65
75
90
0
Figure 10. Sameas Figure7 but the FFOV andCFOV pseudoanisotropic
factorswereconstructed
from
scenesidentified with the CFOV thresholds.
thresholds
thesepercentages
weresmallerthanthoseobtained identificationshowed less evidencefor a view zenith angle
with the ERBE MLE.
dependentbias.
Becauseof cloud clusteringand the growth of the field of
Figure9 showstheazimuthallyaveragedanisotropy
for the
FFOV
and CFOV
observations obtained with the FFOV
view size from nadir to limb, a varying degree of cloud or
threshold
scene
identification
method.
Aswasdonein part1 clear contaminationwas expected in the FFOV threshold
the azimuthallyaveragedanisotropic
factorswereaveragedfor scene identification. Even though the FFOV threshold
all solar zenith angles. Becauseof the effects of cloud method appearedto reduce effects due to errors in scene
obtained
with
the
ERBE
MLE
scene
clustering,the FFOV observations
identifiedas clear were identification
expected
to havemorecloudcontamination
thanthe CFOV identificationmethod, as indicated by the nearly constant
observationsand the overcastsceneswere expectedto have fractional cloud cover at all satellite view zenith angles
more clear contamination.Consequently,for clear skiesthe (Figure 6), the derived anisotropystill appearedto be a
CFOV observations
wereexpectedto showa higherdegreeof .functionof spatialscale. The residualdependence
was due to
anisotropy,
andfor overcast
conditions
theywereexpected
to cloud clustering. Clear and overcastfields of view were
showa higherdegreeof isotropythantheFFOV observations sometimesextractedfrom the midst of brokencloudy fields.
showed. Figure9 showsthat for clearoceanscenesthe
CFOV
observations
had a constant size field of view from
averageanisotropicfactorsfor the CFOV observations
were nadir to limb; so the degreeof cloudand clearcontamination
smaller than those for the FFOV observations at nadir and in the CFOV observations
was expectedto remainconstant
largerat the largestview zenithangle. For overcastscenes from nadir to limb. AS a result, the CFOV thrcsholds
obtainedfor the CFOV observationswere expectedto reduce
Differencesin the azimuthallyaveragedanisotropicfactors thespatial
scaledependence
of thederived
anisotropic
factors.
were reducedfrom the levelsof 2- 10% found in part 1 for
FFOV and CFOV anisotropicfactors were constructed
the E-•E
MLE scene identification to about 0.5% overall.
with the CFOV threshold scene identification. Figure 10
The anisotropicfactorsderivedusing the FFOV threshold showsthe percentdifferences(FFOV - CFOV) for the solar
sceneidentificationmethodshowedless spatial scale zenithanglebins whichhad the largestdomainof significant
dependence
thanthoseobtainedusingthe ERBE MLE scene differencesat the 90% confidencelevel (shaded regions).
identificationmethod. Consequently,
comparedwith the Figure 11 showsthe samedifferencesbut for the solarzenith
E•RBEMLE sceneidentification,the FFOV thresholdscene angle bins which had the smallestdomain of significant
this trend was reversed.
2i,262
YE AND COAKLEY:
RADIATION
BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
PERCENTDIFFERENCE(FFOV - CFOV)
PARTLY CLOUDY
CLEAR SKY OCEAN
66.4-
72.5
72.5
171
171
9
75
90 63
51
39
2.7
0
15
MOSTLY
15
CLOUDY
55.5
-
2.7
39
51
63
75
90
9
180
0
0
75
90 63
51
27
0
39
0.0
60.0
0
0
15
63
75
90
-
25.8
90
9
27
39
OVERCAST
171
$9
51
15
OCEAN
180
51
27
15
90
75
90
63
- 78.5
90
90
180
OCEAN
27
15
51
39
63
75
90
171
9
180
0
75
90
51
63
27
39
0
15
27
15
51
39
75
63
90
Figure 11. SameasFigure8 but theFFOV andCFOV pseudoanisotropic
factorswereconstructed
from
scenesidentified with the CFOV thresholds.
scenetype was usedto assignthe combinedfields of view to
one of the ERBE cloudscenetypes. Thresholdsceneidentification methodswere developedto give the populationof
CFOV observations identified with the CFOV thresholds are
scene types at a given satellite view zenith angle
to the field of view size at the particularangle.
shown in Figure 12. Again the azimuthally averaged corresponding
anisotropic
factorswere averagedfor all solarzenithangles. These sceneidentificationmethodsreducedthe angular
of the sceneidentificationerrorsobtainedwith the
As expected,for overcastscenesthe CFOV observations
were dependence
more isotropic than were the FFOV observations. The ERBE MLE scene identification method. The thresholdto be defined
differences,however,were smallerthan thoseshownin Figure derivedscenetypes,however,are acknowledged
9. Table 2 also showsthat the percentages
were smallerthan by the thresholdsrather than by some characteristicfeature
differences.Comparedwith Figures7 and 8, the domainsof
significantdifferences
in Figures10 and 11 are smaller. The
azimuthallyaveragedanisotropicfactorsfor the FFOV and
those
obtained
with
the
FFOV
thresholds
for both
the
shortwaveandthe longwaveradiances.
4. Conclusions
such as fractional cloud cover. Nevertheless, the thresholds
were designedto mimic procedurescommonlyusedto obtain
cloudcoverfrom satelliteimagery. The thresholdsweresetso
that clear sceneshad the largestlongwaveradiancesand the
smallest reflected shortwave radiances; overcast scenes had
Errors in scene identification coupled with the spatial the largest reflected shortwave radiances and the smallest
scaleson which cloudscongregategive rise to a field of view longwave
radiances.
The scene identification method based on thresholds that
size dependencefor the anisotropyof emittedand reflected
radiances. In addition, as was noted in part 1, the ERBE were designed
to recoverthe population
of scenetypes
MLE sceneidentification
methodappearsto haveidentified
commensurate with the areas of each of the ERBE satellite
manymostlyand partlycloudyscenesat the limb as overcast. viewing zenith angle bins is referredto as the full resolution
field of view (FFOV) threshold method. Fields of view
This tendencywas revealedby assumingthat the ERBE MLE
scene identification method was correct for nadir observations
identified with this method producedreflected and emitted
and then determiningthe relationshipbetweenthe population radiances
with angulardependencies
that wererelatively
of scenetype and the field of view size from the nadir obser- insensitiveto field of view size. Isolated clear fields of view
vations. Fields of view of varioussizes were constructedby drawnfrom the midstof partlycloudyscenes,
however,
combiningneighboringfields of view at nadir. A composite tended to exhibit effects due to cloud contamination, and
YE AND COAKLEY: RADIATION BUDGET, CONSISTENT SCENE ID AND ANISOTROPY
SZA
0.0
-
90.
PARTLY
CLEAR
SKY
ß--,
FFOV
MODEL
•-•
CFOV
MODEL
21,263
CLOUDY
1.3
1.2
1.0
i
0.9
0
20
vlOSTLY
1.3
i
40
i
60
8O
i
0
20
40
60
80
60
80
OVERCAST
CLOUDY
i
i
1.2
1.0
0.9
0
•
•
I
20
40
60
i
80
0
20
i
40
VIEW ZENITH ANGLE (DEGREE)
Figure 12. Azimuthallyaveraged
shortwave
anisotropic
factorsfor FFOV andCFOV observations
obtained
withtheCFOV thresholds.
The azimuthally
averaged
anisotropic
factorswereaveraged
for all solarzenith
anglebins.
overcastscenesdrawn from the midst of mostly cloudy
regions likewise showed signs of breaks in the clouds.
Evidence for these types of contaminationcame from
comparisonswith fields of view identifiedwith constantsize
field of view (CFOV) thresholdsin which the size of the field
of view was setequalto the sizeof the ERBE scannerfield of
view at a view zenithangleof about64ø.
The field of view size dependencein the anisotropyof
reflectedand emittedradiancesrevealedherecoupledwith the
increasein field of view size with satelliteview zenithangle
from nadir to limb should produce satellite view angle
dependentbiasesin the radiativefluxes obtainedby ERBE.
While the magnitudesof the biasesare undoubtedly
small (2
- 10%, dependingon satelliteview zenithangle),the results
presentedhere indicate that they are statisticallysignificant
evenfor a samplesize asshortas 3 months.
Acknowledgments.We thank BruceBarkstromfor stimulatingour
curiosityconcerningthe possiblerelationshipbetweenfield of view
size and the anisotropyof the observedradiances.We also thank
Bruce Wielicki and Richard Green for their insight and numerous
valuablesuggestions.This work was supportedin part by NASA
ERBE NAS1-18992 andby NASA CERES NAG-l-1263.
References
Wielicki, B.A., and R.N. Green, Cloud identification for ERBE
radiativeflux retrieval,J. Appl. Meteorol., 28, 1113-1146, 1989.
Ye, Q., and J.A. Coakley Jr., Biases in Earth radiation budget
observations,1, Effects of scanner spatial resolutionon the
observedanisotropy.J. Geophys.Res.,thisissue.
' J'.A.C0akley
Jr. (co.Tesponding
author),
College
ofOceanic
and
Atmospileric
Sciences,
OregonStateUnivcmity,Corvallis,OR 97331.
Q. Ye, CIRES, Universityof Colorado;Boulder,CO 80309-0449.
(ReceivedDecember7, 1994; revisedDecember 1. 1995'
accepted
April 1, 1996.)
Download