776 
by its cleared appearance from the single-date maps. Its ground resolution, 
on the other hand was degraded due to the necessity of registering twelve 
images with resulting discontinuity in reproduction of narrow objects (small 
rivers and roads). 
The multiband and multidate comparisons of the mean vectors (f and 
their dispersions are summarized for selected classes in Table 1. Accuracy 
estimates of classification are presented in confusion matrices (Tables 2-5). 
Only three ground classes (agricultural land, coniferous forest and deciduous 
forest) were included in accuracy analysis because the areas (number of 
pixels) of the remaining four classes were not large enough to enable 
meaningful estimates. The omission errors and overall classification 
accuracies are summarized in Table 6. 
As was expected, classification accuracy varied significantly as a 
function of the date of image acquisition. The lowest overall accuracy of 67 
per cent was obtained for October 6 ERTS scene. Understandably, this was 
caused by low accuracy in classification of forest classes especially the 
deciduous forest with most stands at the peak of fall leaf coloration at that 
time. Fifty-two per cent of pixels classified as deciduous forest in this 
scene were overestimated. The September 5 image yielded the overall accuracy 
of 81 per cent - the highest for single-date images. The multidate 
classification resulted in overall accuracy of 83 per cent thus improving the 
best single-date image by 2 per cent only. However, one could expect a more 
significant improvement if the images included in the multidate classifier 
were selected on the basis of class spectral patterns (Table 1) rather than on 
the basis of their availability. With ERTS in the third year of its operation 
(exceeding all expectations) the selection of suitable images should be easier 
as more of them become available. Presently we plan to repeat the multidate 
classification of the Larose Forest test area with another image combination 
which includes one winter scene. As was said before, further improvement can 
be expected by inclusion of the spatial image signatures. 
It can be concluded that multidate classification of the ERTS 
digital images is both practical and most promising for thematic mapping of 
broad land cover classes. This is especially important in view of greatly 
improved ground resolution in the ERTS-C and EOS satellites planned to be 
launched in 1978 and 1979 respectively. There is little doubt that thematic 
mapping from orbital platforms will have an important role in mapping natural 
resources and monitoring their dynamic changes. 
ACKNOWLEDGEMENT 
Mr. T.F. Potts and Dr. L. Scherk from Computing Devices Company 
Ltd., under contract to the Canadian Forestry Service, were instrumental in 
development of classification software. Their contribution to this study is 
gratefully acknowledged.