all that is necessary for the mapping at scales of 1:50,000 and smaller 
achieved in this study. 
For the next step -- overlaying and geometric correction of the data 
set -- checkpoints are selected from each frame with a light pen on a video 
display to establish references for a LARS-developed computer program which 
matches picture elements (pixels) from two dates and produces a new tape 
with eight channels of data for each pixel, four from each date. Another 
tape is then created that represents a correction for sampling scale change, 
rotation, skew, and output scale to complete the pre-processing. 
The analysis makes use of this overlain and geometrically corrected 
tape as it proceeds through the following steps. Throughout the process 
there is very close analyst interaction with the machine through the LARSYS 
programs and with forms of ground truth including maps, air photos, and 
field observation. 
The first step in the analysis consists of predicting spectral con 
flicts and selecting major land-use and land cover groupings which exhibit 
spectral and functional similarities. Then, by considering each group 
separately and relative to each other, boundaries in measurement space 
between these groups are sharpened by the following steps in the process 
called cluster analysis. 
In the Indianapolis example these groups included residential, commer 
cial/industrial, forest, and agricultural land. In the case of the residen 
tial group, 11 samples averaging 157 pixels each were selected. 
The LARSYS clustering function was then used to sort these pixels into 
the most spectrally distinct clusters and establish a statistical base for 
further analysis and eventual classification. The number of clusters re 
quired varies with the complexity of the landscape; while 30 were necessary 
to differentiate spectral types within the more complex agricultural group, 
only 10 were needed for the forest group. 
The LARSYS clustering algorithm produces a map of each sample that is 
then spectrally interpreted by the analyst. Each pixel is represented by 
an alphanumeric symbol indicating the cluster of which it is a part. Using 
available maps and photos, with a hand magnifier and a light table, the 
analyst then determines the predominant land cover each cluster represents. 
Certain of the residential clusters from Indianapolis covered older neighbor 
hoods, others appeared in the newer suburbs, still others showed up where 
heavy tree canopy partially obscured houses. The clustering sorted this 
spectrally complex environment into its components. After examining the 
clusters, the analyst deletes from further consideration those defined by 
too few points or with extreme variances as well as those that seem mostly 
to represent contaminant land covers that happened to be included in the 
sa.mpl ing.