3.1. Processing Functions 
It is convenient to develop the discussion of the functional 
requirements which must be met in a general automatic multispectral 
scanner data processing and interpretation system in terms of pattern 
recognition as applied to spectral discrimination. Pattern recogni- 
tion, in general, is concerned with the investigation of adaptive 
and analytic techniques for processing large amounts of data, the 
extraction of useful information to reduce the data, and the classifi- 
cation of the data as required. An extensive review has been given 
by Nagy [27] and discussion suitable to these applications is given 
by Sebestyn [28]. The basis for much of the early work in the field 
was found in "brain" modeling [29]. The significant attributes or 
features of the patterns, which are to be used as the basis for 
classification may be described in one of four different forms: 
physical features, topological features, mathematical features, and 
statistical features. Pattern recognition as applied to automatic 
multispectral processing is discussed in terms of a statistical des- 
cription of features where each pattern or picture element (pixel) 
is considered as a vector in n-dimensional space whose components are 
the simultaneous response in each spectral band from one ground reso- 
lution element. The goal of the recognition system is to define 
partitions in this space such that each region can be identified with 
a class of patterns, through the techniques of statistical decision 
theory [30, 31]. 
Pattern recognition is accomplished by comparing information 
derived from an input pixel with similar data derived from known 
sample patterns (called signatures, paradigms, or prototypes) which 
come from training sets. The specification of these signatures is 
accomplished utilizing a learning algorithm. The important ansatz 
is that these signatures are truly "representative" of the class and all 
of the class dispersion. Based on these comparisons, a decision is 
made as to the nature of the input pattern. 
Before deriving these signatures, it is necessary to format and 
condition the input data to a form suitable for subsequent analysis. 
The purpose in performing this conditioning operation is to enhance 
the features and provide some amount of invariance to ease the 
recognition task. Various techniques for preprocessing or condi- 
tioning the input data are presented below. This is followed by a 
discussion of techniques for extracting features from the condi- 
tioned data and for the design of the recognizer (classifier) and post- 
processor. The purpose of displays is discussed briefly. 
The key to multispectral recognition is invariance. For example, 
in multispectral processing, it is desirable that the classification 
assigned to an object or pattern of interest be independent of the 
position of that object in the field of view, the aspect at which it 
is viewed, the background against which it is seen, partial obscuration 
of the object, minor changes within a class, and changes in illumina- 
tion or atmospheric condition. It is not too difficult to provide 
any one of these invariances. To provide all of the desired invariances 
26