332 
marine silts; bedrock; till; and gravel. From the class statistics, 
a check was made on the validity of the training sets by examining 
a plot of class separability (figure 3) which is a projection plot 
of the decision regions used by the classification program. Also 
examined was the divergence matrix ( a value of about 50 shows that 
the classification was about 85% accurate for a particular pair of 
classes) and the confusion matrix, in which a set of synthetic data 
samples having the same spectral intensities as the training set 
data, are run through the maximum likelihood classifier under the 
multivariate normal assumption (figures 4 and 5)* 
ki lometres 
Figure 2. Study area, Arrowsmith River, N.W.T. 
The results were displayed on full screen to check on 
possible unclassified areas. The additional water unit was noted, 
and at least one unclassified land unit. However, it was felt at 
this 
the 
der 
to a 
all 
ot c 
ion 
phot 
the 
ated