587 
The reason of conclusion (1) can be 
attributed to the unbalance of areas of 
each category. The reason of conclusion 
(2) can be considered as follows. 
Generally speaking, larger number of 
clusters can bring higher classification 
accuracies. However, if number of 
clusters are too much compared to sample 
size, generated clusters are statistcally 
unstable, and classification accuracies 
will decrease. If these assumptions are 
true, the peaks(c1 uster numbers 
corresponding to maximum classification 
accuracy) in each sample size column will 
move to larger number of clusters 
according to increasing sample size and 
the maximum classification accuracy will 
be obtained at the maximum sample size. 
Actually, the former estimation seems to 
be realized in Table 3(c), but for the 
latter estimation, the maximum 
classification accuracy is obtained when 
sample size is 45 x 45. 
The classified result of the maximum 
classification accuracy is shown in 
Fig.6. If you compare Fig.3(a) 
(supervised learning case) and Fig.6, it 
is evident that high density urban areas 
are almost disappeared in Fig.6. This 
phenomena cannot be observed from 
classification accuracies, because high 
density urban areas and residential areas 
are merged in the classification accuracy 
assessment. The reason that high density 
urban areas have disappeared is evident. 
The spectral signatures of high density 
urban areas and residential areas are 
very similar, so there occurs much 
misclassifications between these two 
categories. However, as the areas of 
residential areas are far larger than 
high density urban areas, most of 
clusters corresponding to high density 
urban areas are assigned to residential 
areas because of the nature of area 
assignment. 
The reason of this phenomena were 
considered to be the unbalanced test 
site. To improve the balance of areas 
between categories, new test sites were 
added to test site a. 
4.2 Results for test site b 
Table 4(a) and (b) shows the results for 
test site b. Table 4(a) corresponds to 
area weighted mean and Table 4(b) 
corresponds to an arithmetic mean. From 
Table 4(a), the following conclusions can 
be obtained: 
(1) The absolute accuracies have 
increased about 10% compared to test site 
a. 
(2) The variation among classification 
accuracies has increased a little (about 
10%), but if we omit very low accuracies 
(in the case of 10 clusters), variations 
are still very small (about 5%). 
(3) The tendency of peaks seems more 
apparent compared to the case of test 
site a. This tendency is also shown as a 
graph in Fig.5. However, it is very 
difficult to say that it is statistically 
significant. 
(4) The maximum classification accuracy 
is still obtained at the case of sample 
size 45 x 45. 
In order to acquire more distinct 
results, simple average were calculated 
and is shown in Table 4(b). From this 
result the following conclusion were 
obtained: 
(1) The absolute accuracies are now 
almost the same with Table 4(a). 
(2) The tendency of peaks also exists, 
but it is broken at sample size 50 x 50. 
(3) The variations of classification 
accuracies are still very small and it is 
difficult to obtain statistically 
meaningful conclusion. 
(4) The maximum classification accuracy 
is obtained when sample size is 45 x 45 
and number of clusters is 83, but the 
accuracy in the case of when sample size 
is 30 x 30 and 28 clusters were almost 
the same (only 0.3% difference). 
(5) The maximum classification accuracy 
does not increase with the sample size. 
In order to avoid this deficit, 
percentage assignment has been applied. 
The classified result are shown in Fig.7. 
From this figure, the inverse effect can 
be observed. That is, clusters are tend 
to be assigned to small area categories. 
For instance, if there is a category 
which has only 3 pixels in the test site, 
and 2 pixels in the same cluster are 
classified to that category, the 
percentage of that cluster to that 
category is 66% and this cluster will be 
assigned to that category even when other 
100 pixels of that cluster are classified 
to another category which has 200 pixels. 
From the above results the following 
conclusions were obtained: 
(1) It is natural that classification 
accuracies increase with the number of 
sample size, but this assumption has not 
been certified in this experiments. 
(2) There exists the optimal number of 
clusters according to the sample size. 
(3) Variations among classification 
accuracies were very small and classified 
results were worse than the case of 
supervised learning though the estimated 
classification accuracies using test site 
data were better. 
(4) Most of the above conclusions were 
mainly dominated by the category 
assignment procedure, and a better 
procedure is necessary to obtain more 
distinct conclusion.