413 
cover-type classes had similar density values in all three channels, 
which resulted in classification errors. Even after canonical trans 
formation (Fig. 2), which was done to improve class separability, 
several class densities were still too close for adequate discrimina 
tion. The confusion occurred mainly with features of high reflectance, 
such as light-colored roof tops, bare soil, concrete, and the sludge 
storage and thickener tanks. 
Figure 2. Classification of the color scene. 
After construction of the best possible classification maps, the two 
classified scenes were digitally compared pixel-by-pixel. The compari 
son map was confusing and contained many identification errors (Fig. 3). 
The results from this method were influenced by several problems. First, 
the difference in sun angle between the two scenes introduced two prob 
lems. Because the photos were exposed at different seasons and at dif 
ferent times of the day, the shadows differed in the two scenes. There 
fore, although the shadows were often in unchanged areas, these areas 
would map out as having changed. The sun angle also affected the 
reflectance of many features, resulting in different densities for the 
Figure 3. Digital comparison of Figures 1 and 2. Light gray 
represents changed areas.