I n s t r u m e n t s

a n d

M e t h o d s

according to the algorithm ofShi and Dozier(1996).For
this project, theoutput of this algorithm which is based on
a decision tree classificationis an image that is partitioned
into four classes:water, bareground/short vegetation,forest,
and snow (figure 1).
For validation, a Landsat Thematic Mapper (TM) image
acquired on April 14,1994 overthe same area was used.
We performed two types of snow mapping on this image.
Forourprimary validation image,we used the algorithm
of Rosenthal andDozier (1996) to map sub-pixel snow frac-
tion for the scene (figure 4) [note: to facilitate comparison,
thefiguresarenot inorder].Thisalgorithmhasdemon-
strated accuracies on parwith those obtainable from air-
photo interpretation (Rosenthal and Dozier, 1996). Adding
to ourconfidencein this method is knowledge that its ac-
curacy assessment was performed overMammoth Moun-
tain, eliminatingquestionsabout the geographicportability
ofthealgorithm.Asasecondaryvalidationimage,we
classified theTM sceneinto the above four categoriesusing
adecisiontreeclassifier(figure2).Thisclassification
remainsunverifiedsoanyconclusionsdrawnfrom
informationderivedthroughitmustbetempered.To
facilitate intercomparison, wecoregistered andwarpedthe
TM products to the SIR-C ground range geometry.
The primary validationrequired a method of compar-
ing the output of the snow fraction image (quantitative or
ratiodata) to the output of the SIR-C algorithm (qualitative
ornominaldata).These are two fundamentally different
datatypes.Inordertocomparewe needed totransform
one of the data types to the other. In other words, we could
eithersynthesize a fractional snow coverimage from the
SIR-C classification, or we could generatea nominalimage
of snow-covered/snow-freefrom the TM snowcoveredarea

map.The transformation from fractionalSCA to a binary
SCAimagewouldrequiresettingsomesnowfraction
threshold above which pixelsare classified as snow and
below which pixels are classified as snow-free. Sincethis
process wouldinvolve a significant loss of information and
the choice of a threshold seemed arbitrary, we pursued the
synthesized fractional SCA SIR-C image option.
OnemethodforsynthesizingafractionalSCAimage
from a nominal snow/snow-free image is to select a win-
dowsizeinwhich tocalculatethe percentage ofpixels
that are classifiedas snow. This window size is bounded,
at one extreme,by a window size of one in which all out-
put pixels are either100% or0% snow and,at the other
extreme,byawindowsize equaltotheentireimagein
which the output is simply the fraction of the image that is
classified as snow cover. Thus wehavea situation in which
we must weigh the value placedon cell resolution against
thevalue offractionalprecision.We selected awindow
size of 10 by 10 pixels. This choiceseemedreasonablesince
it maintained a moderate resolution with a large number
of samples in the scene and the conversion to percent is
conceptuallysimple.Sincethesquareof10is100,the
percentis simply the number of pixels in the window that
were classified as snow (figure 3)..
Withtwo coregisteredimages of fractional SCA, all that
remained wassimply to createa differenceimage inwhich
everypixel inthe image istheSIR-C fractionminusthe
TMfraction.Positivenumbersonthisimage,therefore,
represent areaswhereSIR-CfindsmoresnowthanTM
while negative numbers on this image are areas where TM
finds more snow than SIR-C (figure 5).
In addition to this image, we deemedit useful to perform a

130