368 
their shadows in the study site. In the resulting 
classification map we have difficulty in recognizing rivers 
and a park. 
In Figure 10 we illustrate the same classification 
result of the ground albedo data set on Sept.2, 1980 
classified according to the class statistics obtained from 
the ground albedo data set on May 23,1979. Although the 
accuracy of classified result is not so good as compared 
with that in Figure 8, it is much improved when compared 
with that in Figure 9 and we can recognize the rivers and a 
park easily in Figure 10. 
In Figure 11 we illustrate the same classification 
result of the original CCT level data set on Sept. 2, 1980 
classified according to the class statistics obtained from 
the original CCT level data set on May 23,1979. As is seen 
in Figure 11, the accuracy is very poor and the most of 
urban, residential area is misclassified. This suggest that 
the temporal data analysis can not be done in the form of 
original CCT level, as we claimed in the preceeding 
paragraphs. In addition to this ,the above examples indicate 
clearly a possibility of signature extension between among 
data sets taken at different times if those data sets were 
converted in absolute albedo uint. 
In the next section we describe a processing system 
implementing the signature extension of the remote sensing 
data. 
STRUCTURE OF SIGNATURE EXTENSION SYSTEM 
We show the overall data flow in signature extension system 
in Figure 12. There are two phases, i.e., the data base : 
phase and the table look-up phase in a signature extension : 
system. The data base phase consists of three program steps, - 
that is, the AECS step, the classification step produces , 
both classification map and the statistical quantities for 
certain classes if necessary. The accumlated statistics on 
pattern classes are stored in the form of Table of 
Reflectance Pattern Class (TRPC) in computer disk. In the 
TRPC each class is expressed in terms of three digits. The 
first digit gives a basic land cover such as the forest, 
sea, urban, and so on. The next two digits specify more 
detailed category such as evergreen forest, deciduous forest 
and so on. The correspondence between this three digits 
class and the actual land cover should be done through the 
ground truth. The TRPC should contain the following 
information: three digits class number, its mean albedo 
values in bands 4,5,6 and 7, its covariance matrix elements, 
its scene ID number, the date of exposure and the 
reliability bound. The reliability bound gives the seasonal 
reliability of class statistics. 0 j 
The table look-up phase consists of three steps, other 
than the AECS step. In the extraction step an appropriate 
part of the large data base TRPC is extracted according to 
the date when the Landsat data set to be classified was 
taken. If all patterns are normally distributed, it is said 
that the Mahalanobis distance D 2 (i) between a point 
belonging to a class with class number i and its class 
center has a chi-square distribution. In the table look-up 
and link-index step we construct confidence regions for