323 
The number of unclassified and grossly misclassified pixels ’ 
is high outside of the six sub-areas from which sample points 
were taken, indicating the importance of obtaining sample 
points from all parts of the study area. Many of the smaller 
channels are completely dropped from the images although 
scattered pixels of "water" or "point bar" sometimes roughly 
indicate their location. As can be seen in the two classi 
fied images, "point bars" were detected along virtually all 
of the major waterways. This of course did not correspond 
to airphoto observations. Because of the small size of 
many of the point bar sample areas, and because the physical 
form of the point bar often consists of a mixture of unvege 
tated ground and small bodies of water, it would be more 
accurate to label this class "land-water interface". Thus 
large areas are classified as "point bar" in the shallowest 
part of Mamawi Lake and along the narrow channels of the 
Delta. Because of the small size of the sample areas of 
"point bar", many of the sample points included both land 
and water. 
Certain tendencies become apparent upon examination of 
accuracy figures for the three years. Firstly, accuracy of 
classification of each unit varies greatly between years. 
For example, in the ungrouped analysis, the accuracy of 
classification of levees (UNIT = 1) was 38.8% in 1974, 47.5% 
in 1978 and 63.2% in 1980. Mud flats varied from 49.5% in 
1974 to 76.0% in 1978 and 13.5% in 1980. One possible 
reason for the latter remarkably high range is that mud flats 
were generally quite wet in 1974 (a high water year), they 
were distinct in 1978 (a dry year following several wet 
years) and in 1980 (a dry year following two dry years) 
succession to grasslands had progressed far enough to alter 
the spectral signature although vegetation was still 
sparse enough to look like mud flats on the airphotos. Thus 
confusion with vegetated classes in 1980 was a primary 
reason for the low accuracy of classification of mud flats. 
Secondly, the analyses of grouped data had much less varia 
tion in unit accuracies than ungrouped data, although the 
misclassification of mud flats in 1980 is an exception. 
The accuracy of classification of the largest group (levees, 
muskeg and shrubbed basins) ranged only 0.5%. This suggests 
that this type of analysis may be best suited to analysis of 
relatively simple systems and improved through the use of 
larger sample sizes. 
CHANGE DETECTION 
The result of a series of low water years is obvious when 
the images in Figure 3 are compared. The class "mud flats", 
which was fairly extensive in 1974, had been replaced almost 
entirely by vegetation in 1980. Water was classified much 
more consistently for 1980 than for 1974, when many pixels 
located in lakes or rivers were unclassified. The vegetated 
unit was more evenly distribute in the 1980 image than in 
the 1974 image. The extent of water bodies had also been 
considerably reduced. However, because of the reduced number 
of units which had to be used to obtain reasonable accuracies