variety of graphic displays, statistical data, and statistical measures 
of separability which can be used to evaluate the quality of training 
set statistics and the need for additional training sets. Training is 
usually accomplished in an iterative fashion and allows for rapid re- 
finement of training sets, addition of new training areas, and re-analysis 
of all statistics until an acceptable set of training set statistics is 
derived. The procedures used for deriving training set statistics varies 
with the analysis objectives and the classification algorithm employed. 
An extremely fast classification algorithm available on many systems 
is the parallelepiped decision rule. This algorithm is hardwired and 
operates in real time but does not employ rigorous statistical decision 
criteria as maximum likelihood decision rules do. The parallelepiped 
decision rule is based on calculating the minimum and maximum brightness 
values in each spectral band for a given training set. A given pixel 
is assigned to the class represented by the training set if the pixel 
brightness value in each spectral band fall within the minimum and max- 
imum brightness values calculated for the training set. There are several 
adaptations of maximum likelihood decision rules but generally they are 
more rigorous decision rules, require more processing time, and generally 
yield more accurate results for detailed classification. 
Data that has been classified can be displayed in color or black 
and white in geometrically corrected format on a color monitor or dis- 
played on a variety of line printers. One common method is to output 
classification results geometrically corrected on a line printer at a scale 
of 1:24,000. Classification results can also be outputted to tape and re- 
formatted for display on a variety of film recorders from which a colored 
classification map at a user specified scale can be produced. 
Processing time varies considerably with the specific image analysis 
system used but more importantly varies with the objectives of the anal- 
ysis, acceptable accuracy levels, number of classes, number of data chan- 
nels, and the size of the project area. Johnson (1975) indicates that 
typically 20 to 30 categories can be extracted from a LANDSAT scene and 
that a full scene can be analyzed in eight hours with approximately 6.5 
hours for training and analysis and 1.5 hours for generating "categorized, 
geometrically corrected" film products.  Erikson (1975) reports that a 
LANDSAT scene can be clássified into 16 classes in approximately 40 sec- 
onds. He further indicates that processing costs of $500/frame may be 
possible. Clearly, image processing systems, currently available, are 
highly flexible, user oriented systems capable of processing data at 
greatly reduced costs than previously available. These systems have pro- 
vided improved procedures for training and classification of multispectral 
data and subsequent display of results in formats compatible with the user 
community. 
Geometric Correction 
Geometric distortions are inherent in LANDSAT film products and com- 
puter compatible tapes. These distortions greatly influence the plotting 
accuracy of boundaries and extraction of area statistics. A number of re- 
seachers have examined the geometric quality of LANDSAT data. Wong (1975) 
summarizes the research of several investigations and based on his research, he 
reports that geometric distortions in bulk processed LANDSAT MSS film pro- 
ducts may range from + 150m to + 350m. Wong (1975) indicates a relative