are treated favorably in the maximum 
percentage method as shown in Fig. 1. 
(3) Minimum Distance Method 
In this method, first, TCA is overlaid 
with the target image and the centroid of 
ecah category is calculated. Secondly, 
the distance between each category and 
each cluster is calculated. Then the 
category having the minimum distance is 
assigned to that cluster. For the dis- 
tance, Euclidian distance was used in 
this experiment. 
In the case of the maximum number method 
and the maximum percentage method, the 
result is dependent upon the size and 
location of training area. In the minimum 
distance method, it needs less number of 
pixels for TCA compared to other 3 meth- 
ods but needs caution on spectral featute 
of ICA. 
(4) Element Ratio Matching Method 
In this method, based on the idea that 
the category can be represented by the 
Set of clusters, the relation of catego- 
ries and cluster is determined by the 
ratio of clusters which compose a catego- 
ry. For defining the element ratio, area 
is necessary. Therefore, the process is 
performed by local region as follows. 
At first, TCA is overlaid with the result 
of clustering, which is classified by the 
clusters, and the element ratio of each 
category is calculated. Secondly, for 
each local region that. is fixed on the 
result of clustering, element ratio is 
calculated and matched with the element 
ratio of each category precalculated. 
Then the category having the most similar 
element ratio of concerned region is 
assigned to that region. 
4. EXPERIMENTS AND RESULT 
4.1 FLOW OF EXPERIMENTS 
In order to evaluate the proposed methods 
described in chapter 3, following LAND- 
SAT TM data was used in the experiment. 
Sagami River basin was seleted ‘for target 
area to pereform the quantitative evalua- 
tion. This area includes the test site 
area which landcover is already investi- 
gated and categorized in 52 categories. 
sensor : LANDSAT TM 
date * Nov. 4, 1984 
path-row ^: 107-35 
area : Sagami River basin in Japan 
pixel size: 25m X 25m 
image size: 512 X 480 pixels 
At first, clusters were generated by a 
hierarchical clustering using Ward meth- 
od. Since the image data in remote sens- 
ins is very large, usually clustering is 
performed with sampled data. In this 
experiment, 2500 samples(about 1% of 
entire image data) were used to generate 
66 clusters. Based on the 66 clusters, 
the target image data was classified by a 
maximum likelihood method. Secondly, 
140 
representative area of each category in 
the target image(training category area) 
was selected. 14 categories were selected 
as shown in Table 1. Finally, the rela- 
tions of clusters with categories were 
determined by 4 methods described in 
chapter 3. 
To evaluate quantitatively, classifica- 
tion accuracy was estimated based on the 
test site area. The classification accu- 
racy was calculated over 5 major catego- 
ries as shown in table | to adjust the 
selected 14 categories in target image to 
52 test site categories. 
4.2 Results of Experiments 
(1) Maximum Number Method 
Fig. 2 shows the result by this method, 
and Table 2 shows the classification 
accuracy estimated with the test site 
data. As expected, this result shows the 
tendency that clusters were labeled with 
the categories having large area such as 
urban. 
(2) Maximum Percentage Method 
Fig. 3 shows the result by this method, 
and Table 2 shows the classification 
accuracy estimated with the test site 
data. As expected, this result shows the 
tendency that clusters were labeled with 
the categories having small area such as 
other. 
(3) Minimum Distance Method 
Fig. 4 shows the result by this method, 
and Table 2 shows the classification 
accuracy estimated with the test site 
area. This result is very similar to the 
result(Fig. 2) by a human operator. In 
other words, a good result was obtained 
with roughly selected TCA. 
(4) Element Ratio Matching Method 
Fig. 5 shows the result by this method, 
and Table 2 shows the classification 
accuracy estimated with the test site 
area. In this method, 5 X 5 sized local 
region was used. 
(5) Conventional Supervised Method 
For the purpose of comparison, conven- 
tional supervised method was executed. In 
this method, clusters were categorized by 
a human operator. 66 clusters were la- 
beled with 14 categories (Table 1) as 
Shown in Fig. 6. Table 2 shows the clas- 
sification accuracy by this method with 
the test site area. 
5. CONCLUSIONS 
Four categorization methods for clusters 
were evaluated by experiments using LAND- 
SAT TM data. From the results of experi- 
ments, following conclusions were ob- 
tained. 
(1) In the case of the maximum number 
method, the clusters are hardly labeled 
with 
con 
per 
harc 
rie: 
thes 
area 
(2) 
weise 
elem 
i n f« 
los! 
cate 
loca 
(3) 
meal 
meth