AUTOMATIC CATEGORIZATION OF CLUSTERS 
IN UNSUPERVISED CLASSIFICATION 
Sunpyo HONG, Kiyonari FUKUE, Haruhisa SHIMODA, Toshibumi SAKATA 
Tokai University Research and Information Center 
2-28-4 Tomigaya, Shibuya-ku, Tokyo 151, JAPAN 
ABSTRACT: 
À cluster categorization method is necessary when an unsupervised classification is 
used for remote sensing image classification. It is desirable that this method is 
performed automatically, because manual categorization is a highly time consuming 
process. 
In this paper, several automatic determination methods were proposed and evaluated. 
They are 1) maximum number method, which assigns the target cluster to the category 
which occupies the largest area of that cluster; 2) maximum percentage method, which 
assigns the target cluster to the category which shows the maximum percentage within 
the category in that cluster; 8) minimum distance method, which assigns the target 
cluster to the category having minimum distance with that cluster; 4) element ratio 
matching method, which assigns the local region to the category having the most similar 
element ratio of that region. From the results of experiments, it was certified that 
the result by the minimum distance method was almost the same as the result made by a 
human operator. 
Key Words: unsupervised classification, post processing, categorization, clustering. 
1. INTRODUCTION (3) It is time consuming when the number 
of cluster is large or there are many 
With the launch of second generation high small clusters. 
resolution sensors like LANDSAT TM and 
SPOT HRV, clustering method has been 3. AUTOMATIC CATEGORIZATION METHOD 
revaluated recently. However, the main 
problem of clustering for practical use To solve the above problems, several 
is that clustering is an unsupervised automatic categorization methods are 
classification. That is, clusters gener- considered as follows. In all methods, 
ated by clustering are defined in feature training category areas(TCA) are first 
vector space, not in image data. There- extracted from the target image similar 
fore, in order to use the classified to supervised trainings. 
result for a meaningful reference map, it 
is necessary to determine the relation of (1) Maximum Number Method 
clusters and categories, and to label the 
classified result with the categories. In this method, the number of pixels in 
each TCA for each cluster is calculated. 
Conventionally, this relation has been Then the category having the maximum 
determined mainly by interpretation of an number is assigned to that cluster. 
operator. However, this process is time 
consuming and is not objective. (2) Maximum Percentage Method 
The purpose of this research is to try In this method, for each cluster, the 
several methods of automatic categoriza- percentage(occupation rate) of that clus- 
tion and find out the most useful method. ter in each TCA is calculated. Then the 
In this paper, 4 methods have been exam- category having the maximum percentage is 
ined. assigned to that cluster. 
2. PROBLEMS OF CONVENTIONAL METHOD Fig. 1 shows a comparison of these two 
methods in a simple case. Suppose that 
In this method, each classified cluster cluster k is composed of three categories 
is overlaid with the target image data on A, B and C. As shown im Fig. 1(a); cate^ 
the display, and that cluster is inter- gory À occupies the largest area in 
preted by an operator to determine the cluster k and C occupies the minimum 
category. Therefore, it can be thought area. In the maximum number method, 
that the obtained result is natural and cluster k is always assigned to category 
reliable. A. However, this figure does not show the 
difference of areas of each catesgory. 
However, since everything is determined Fig. 1(b) shows the case that the total 
by an operator in this method, there are area of each category is the same and (c) 
many problems as follows. shows the case that the total area of 
each category is different. As shown from 
(1) The result depends on the skill of this figure, categories which occupy 
an operator. small areas in the image tends to be 
(2) Objective and quantitative evalua- neglected in the maximum number method. 
tion dis difficult. On the contrary, small area categories 
139