I n s t r u m e n t s

a n d

M e t h o d s

Fig.3

Fig.4

conventional errormatrix comparison using the TM clas-
sified data (oursecondary validation image).This isthe
conventionalmethodbywhichaccuracyassessmentis
performed(Congalton,1991).Forsnowcoveraccuracy
assessment we aggregated categories so that we were left
withabinaryimageofsnowandsnow-freecases.We
calculated three types of accuracies. The overall accuracy
is the likelihood of a pixel being classified into the same
category by both classifiers. Producer's accuracyis thelike-
lihood of a pixel classified as snow in the referenceimage
(TM) being classified as snow by the experimental (SIR-C)
image. It is sensitive to errors of omission. Another type of
accuracythatisreported,user'saccuracy,expressesthe
likelihood that a pixel classified by the experimental sen-
sor as snow was classified as snow by the referencesensor.
User'saccuracyissensitivetoerrorsofcommission
(Congalton,1991).

RESULTS

The algorithm's accuracy at mapping snow is highly vari-
able across the scene. On the positive side, it is quite suc-
cessfulinthe extremecases.That is,forareasin which
TM reports no snow at all, SIR-C is very likely to report no
snow (only 13% of TM snow free pixels had snow on the
SIR-Cimage)and areas thatarefoundtobecontinuous
snow fieldsare also correctly mapped as snow bySIR-C
forthe mostpart (63% of TM pixels with 98% orgreater
snowfractionwereclassifiedassnowbySIR-C).This
phenomenon is illustrated in figure 6,which depicts the

absolute error in percentsnow cover estimation over a range
of TM snow fractions.
On the other hand, the overall predicted snow fraction
by SIR-C is much less than that predicted by the TM algo-
rithm (30% vs. 62%). This underprediction is most preva-
lent in forested regions.We note that 70% ofregions that
are underpredictedby SIR-C by 60%or more are classified
by SIR-Cas forest.Going back to figure 6, the peak errors
occurforTM snow fractions of around 70%. The bulk of
these pixels iscomposed ofmixtures of forest and snow
that the TM mixture algorithm interprets as 70%snow and
30%forest.SIR-Ctendstoclassify theseareas asforest.
This is the mixed pixel problem inherent in any nominal
classification. Regions on the ground that compose a pixel
are more often than not made up of several different cover
types.A nominal classifiermust pick only one of them to
represent the entire pixel while a mixture model can se-
lect a fraction of each of the land cover types to character-
ize the pixel.In comparison to woody vegetation,snow's
signalinthe microwave portionofthespectrumisvery
weak. Thus, it is likely that pixels that are mixturesof snow
and forest will be classified as forest by such a system.
Further obscuring the detection of snow by radar is the
side looking nature of radar sensors. Whileoptical systems
typically look very near nadir, this SIR-C image has a look
angle around 25degrees.Ina forestedregion,itisclear
thathigherlookangleswillresultinseeinglessofthe
surfaceunderneath forestcanopy,while nadirandnear
nadir systems will be able to see much more of the under-
lying surface.

131