My assessment of the status of processing techniques is as follows: 
1. A variety of techniques have been demonstrated to be feasible in 
many applications under limited conditions appropriate to showing 
feasibility but not generally appropriate to prototype operational 
conditions. 
a. The accuracy achieved in some applications is acceptable, in 
others it needs to be improved before operational use will be 
undertaken. 
b. Little or no time constraint for processing has yet been imposed. 
c. Too much ground observation has been used. 
2. These limitations are being lessened to the point where operational- 
prototype information systems are feasible in some applications. 
3. Further processing technique development is necessary. 
The long operation of the ERTS-1 multispectral scanner has allowed users 
to observe areas of interest repeatedly at 80 meter resolution. Because the 
ERTS MSS has four rather broad spectral bands, extractive processing using 
only spectral information frequently yields imperfect separation of classes 
of materials of interest. Partially to overcome the degraded performance of 
ERTS spectral channels relative to those available from typical aircraft 
scanners, users have exploited the spatial information inherent in the ERTS 
along with the spectral data. Further, using the repetitive coverage 
capability of ERTS, some users have used the temporal variation of spectral 
data as inputs to the pattern recognition processors. 
ERIM has explored both applications of information from ERTS, with 
promising results. In one of the approaches we have tried to spatial-spectral 
processing, the key step is the formation of spatial "features" or quantified 
attributes of the scene. Spatial features were formed from ERTS data over 
Michigan as shown in Table 1 (Kauth, 1974). These spatial features were 
formed by measuring variations in ERTS band MSS-7 signal level in a 9 x 9 
array of pixels with the pixel of interest at the center. Then both spectral 
and spatial signatures were extracted for terrain categories in the 
Ann Arbor-Brighton, Michigan area. 
An optimum feature selection was made based on an algorithm which selects 
the feature which, along with the features already selected, minimizes the 
average pairwise probability of misclassification between pairs of signatures. 
The results of the optimum feature ordering are presented in Table 2. Note 
that 2 of the first 4 features are spatial features.