to perform some or all of the following functions without regard to 
order: 
1. Enhancement and strengthening of features. 
2. Provide invariance to scale changes of the input data. 
3. "Noise Suppression" 
(a) Suppression of variability present in the input data 
(variability which obscures the phenomena to be recognized, 
making the task of feature extraction more difficult). 
(b) Normalization of input data. 
Formatting the processed data for acceptance by the 
recognition device. 
5. Provide geometric and radiometric corrections to the 
data. 
The enhancement and strengthening of features through preprocessing 
is the most critical function of multispectral scanner data processing 
for operational systems. Success of even the most powerful automatic 
techniques is at best limited to the vicinity of local areas near 
training sites (tens of kilometers usually) unless account is taken of 
systematic variations imposed on the signatures by the conditions 
of measurement (such as atmospheric haze, cloud shadows, and the position 
of the sun) and parameters of measurement (sensor characteristics) 
which obscure a signature's characterization of the unique or sensible 
attributes of a class. Feature enhancement by preprocessing uses some 
of the knowledge we have about the physical world. Ignoring such things 
as the effects caused by differences in slant path through the atmosphere 
is to ask too much of the classification function and is reflected in 
a costly need for additional ground observations and frequent re-training. 
Invariance to scale changes and systematic variations is a key 
to high accuracy discrimination. Variations which are due to within 
class dispersion are the only variations which the statistical discrim- 
inant analysis techniques are designed to handle. 
The function of "noise suppression" is provided by preprocessing 
in the sense that effects of random noise of both high frequency and. 
low frequency components can be reduced by integration of the signal 
if redundancy exists in the data and by clamping and scaling to stable 
reference signals, which also normalizes the data. 
The fourth function to be performed by the preprocessor is the 
formatting of the processed data for acceptance by the recognition 
device. It is frequently advantageous to transform the output of the 
sensor to facilitate further processing of the data, to reduce the 
dimensionality of the data through selection of a subset of spectral 
channels, and to ease the task of feature extraction. The decision as 
30