yields spatial feature vectors of dimension 15 which correspond to spatial 
frequencies of 0.39 cycles/km to 5.86 cycles/km along the scan direction 
and to 0.56 cycles/km to 8.46 cycles/km along the satellite path. For 
classification purposes the logarithm to base ten of the power spectra com 
ponents was taken and a multiplicative scale factor applied to the data 
values in order to format the resulting values into the 0 to 255 range 
required by the classification programs. Scanning the training areas indi 
cated in Figure 1 with a 32 by 32 cell size compressed the data by a factor 
of 1024 and yields the sample sizes indicated in Table I. In effect, each 
cell which contains 1024 pixels of four spectral channels each is replaced 
by 1 pixel with fifteen spatial channels. 
Spectral Average Features - Spatial/Spectral Average Combined 
Features : A second set of feature vectors for the July 10 image 
was obtained for the 32 by 32 pixel cells utilized in calculating the spatial 
features by averaging the spectral response in each MSS band for the 1024 
pixels contained in the cells. This process yielded a spectral average 
feature vector of dimension four for each cell. In order to obtain a com 
bined feature set representing both the spatial and spectral information 
contained in a cell, the 15 dimensional spatial components for each cell 
were merged with the four spectral average components for the cell to yield 
a 19 dimensional feature vector. 
Spectral Features : In order to calculate the mean spectral response 
and covariance matrix of individual pixels for each category, training fields 
of dimension 10 by 10 pixels were selected. Within the large test sites 
denoted in Figure 1, five or six such individual fields were extracted for 
each category as noted in Table I. These fields were chosen in such a 
manner as to sample the reflectance patterns within the larger test stages. 
Spectral signatures were calculated for these training fields for both July 
and August, and an individual pixel by pixel classification of the training 
sample carried out. 
Classification : Table II is a summary of the classification 
comparison applied to each image using the feature sets described above. In 
all cases a supervised likelihood ratio test assuming multivariate normal 
distributions for the data was employed (Smith, 1972). For each comparison, 
the corresponding sample points were used to estimate the mean vector and 
covariance matrix for each category as a function of the feature type employed. 
These same sample points were then classified using the likelihood ratio 
test.