OVERALL SYSTEM BLOCK DIAGRAM 
variations, where transference of signatures from one frame to another is 
desired, or where spectral features can be enhanced. There is no clear-cut 
means of examing a data set to decide, in a priori manner, which of several 
possible transforms is needed for a particular scene. The method used would 
normally be one in which each of several transforms are performed on the 
training sets and the resulting data is tested for the optimal probability 
of correct classification using the training set and test set data to choose 
the transform to be employed. 
The purpose of the linear transform (Crane, 1973) is to provide 
a new set of data in which the spectral data is combined in such a manner 
that the dimension-reduced, transformed data has essentially the same 
discriminability for the classes of interest to the user. This has the 
desirable effect that the classifier can perform a classification operation 
in which the accuracy of classification using the smaller number of dimensions 
is equivalent to that obtained with a larger number of untransformed 
dimensions. 
The classifier performs a maximum-likelihood decision, assuming a multi 
modal Gaussian multivariate distribution. This assumption has been well 
justified at this time by over 100 experiments using multispectral data at 
ERIM (Marshall, 1973) and, as time goes on, by more and more experience 
at NASA and other centers. Although simpler algorithms can perform well