694 - 
techniques employing a 3 x 3 pixel array in which information from 9 pixels is 
used in classifying the "central" pixel only. Improvement over single pixel 
classification is consistently achieved. Work on automatic boundary detection 
and sample classifiers is also promising greater accuracy as well as efficiency. 
TABLE 3. PERFORMANCE OF CLASSIFIER USING SPECTRAL/SPATIAL FEATURES 
m 
in 
< 
TRUE 
CATEGORY 
WATER 
SOIL 
URBAN 
FOREST 
AG 
WATER 
19 
1 
0 
0 
0 
SOIL 
0 
0 
0 
2 
4 
URBAN 
0 
0 
21 
0 
0 
FOREST 
0 
1 
1 
18 
9 
AG 
0 
1 
1 
3 
35 
UNCLASS 
2 
2 
1 
1 
1 
AVERAGE 
PER 
CLASS 
50% 
17% 
87% 
75% 
71% 
OVERALL 
AVERAGE 
CORRECT CLASSIFICATION 
76% 
Because of the periodic coverage of ERTS, the temporal variation of 
spectral signatures can be utilized to help discriminate objects. This 
capability is enhanced with ERTS data over the previously available aircraft 
data because the periodic coverage is obtained from a sun-synchronous orbit, although 
separate radiometric correction of each pass may be needed. Further, the data are 
obtained from a relatively stable platform and at small scan angles from the nadir 
(typically +5.5°). Both facts greatly ease the data registration problem-that of 
merging data collected at two different times so that pixels are precisely overlaid. 
To illustrate the advantage of multitemporal spectral processing, 
consider the following example of mapping of natural vegetation in northern 
Michigan. (The work being discussed here was performed for NASA under grant 
NGR23-005-552). 
Table 4 shows the signatures of a mixture of hardwoods, conifers, and 
grass and shrub swamp on two different dates — June and March. Shown are the 
mean values of the signatures, with the standard deviations in parenthesis. 
Notice that in June, the two classes 1 signatures overlap appreciably in each 
ERTS channel — the mean difference between classes is less than the standard 
deviation in each channel. A pattern recognition device hopelessly confuses 
these two classes in June. 
But in March, the two classes are more separable, as shown in the second 
half of Table 4. This separation of some classes at one time of year and 
not at others can be exploited to improve maps of natural vegetation areas. 
Now the challenge is to understand how vegetation signatures change with 
time, so that we may specify exactly when data are to be collected.