RESULTS 
Tables III, IV, and V summarize the classification performance 
obtained when spatial features are employed and when signatures estimated 
from one date (July 10) are used for classification at a later date (August 
15). In determining the average classification accuracy, the individual 
category accuracies were weighted by the number of samples in each category. 
For both July and August, classification accuracies between 80 and 100 per 
cent were achieved on the training data. There does not appear to be a 
significant difference in the performance on these two dates. It is seen 
from Table V, however, that it is not possible to maintain these accuracies 
when July 10 signatures are applied to August 15 data. In particular, 
dryland farming is confused with rangeland and urban and irrigated farming 
are confused. 
Tables VI and VII show the performance estimates obtained from 
pixel by pixel spectral classification for the two dates. Performance was 
slightly better for the August 15 overflight indicating that greater spectral 
differences among the vegetation land-use categories existed. Irrigated 
farmland was nearing a mature state and was less easily confused with either 
rangeland or urban categories. However, urban classification accuracy was 
strongly degraded. The overall accuracy obtained when July 10 spectral 
signatures were applied to August 15 data was 61 percent indicating that 
neither spectral nor spatial signatures may be extended for these two dates. 
Tables VIII and IX summarize the results that can be achieved when 
either spectral average features for all MSS bands or a combination of 
spatial/spectral average features are employed. If all four MSS bands are 
employed, greater accuracies are obtained with spectral average features 
than with spatial features derived from only one band. Using only the 
spectral average feature from one band, however, yields an accuracy of only 
68 percent. Adding the spectral average feature from one band to the spatial 
feature vector derived from another band yields results comparable to that 
obtained when all four spectral average feature components are used.