same feature in the two scenes. This problem was encountered with the 
roof tops, and with the sludge storage and thickener tanks. 
A second problem was caused by seasonal and climatic changes. For ex 
ample, the deciduous trees which dominated the forested areas, still had 
their leaves in the October 1972 scene. However, the same semiforested 
areas had not achieved spring leafout in April of 1978, when the later 
scene was taken. This seasonal change also affected the classification 
of grass areas in both scenes. 
Another natural change that affected classification was the saturation 
of the water-containing features. In the early scene they appear rela 
tively dry, while in the later scene they appear to be very wet. These 
changes, like the shadows, are shown as areas of change in the final 
results, but do not define changes in land use/land cover. 
The second method, involving the classification of a merged data set, 
was subject to the same problems as the first technique and gave similar 
results. The merged four-channel data set was classified into 15 clas 
ses. Several of the classes represented areas of known land cover 
change. These density classes were often too similar to other land 
cover type classes for good discrimination during classification. As a 
result, the final map contained a large number of misclassifications 
(Fig. 4). Here again, canonical transformation assisted in separating 
Figure 4. Classification of the merged data. 
out some classes but others were still too similar in multivariate den 
sity to separate. For example, the area associated with the addition 
to the incinerator building, which was grass in the 1972 scene, falls 
into the same classification as the areas that went from grass in 1972 
to bare soil in 1978. Also, the forested area was not clearly differ 
entiable from grass. Some of these misclassifications can be attrib 
uted to the sun angle and temporal change problems discussed in the 
first method. 
In the third method, the norms of the two-channel data set of ratios and 
differences were density sliced to highlight the extremes in the data. 
The density slices were selected on the basis of a frequency distribu 
tion table. Five density slices are displayed in Fig. 5. Although the 
sun angle and the temporal change problems continued to influence the 
results, the third technique did succeed in highlighting areas of 
414