MINIMUM RESIDUAL CLUSTERING 
Haruhisa SHIMODA, Kiyonari FUKUE, and Toshibumi SAKATA 
Tokai University Research and Information Center 
2-28-4 Tomigaya, Shibuyaku, Tokyo 151, Japan 
ABSTRACT 
A new high accuracy clustering algorithm named a minimum 
residual clustering is presented in this paper. This 
algorithm can eliminate the disadvatages of a conventional 
maximum likelihood method assuming normal distributions 
and a clustering using a  conbinational method while 
keeping advantages of these method. Thus the minimum 
residual clustering achieves an automatic classification 
with a high classification accuracy. 
Simulation experiments of the algorithm and case studies 
using this algorithm were conducted. In case studies, a 
landcover classification and a forest type classification 
were examined using Landsat MSS data and "multisensor data 
set: lof: IAPR :TC-7, As the results, the expected 
characteristics of the minimum residual clustering were 
confirmed. Further, processing times were about 1/5 
compared to a conventional maximum likelihood method. 
1. INTRODUCTION 
A classification is a principal and time consuming 
processing phase in image processings for remote sensing. 
Two kinds of classification methods are usually used. The 
first one is a supervised method; typicaly a maximum 
likelihood method assuming a normal distribution of the 
samples are used and it has the advantage on the 
classification accuracy. The other is a unsupervised 
method; typicaly a clustering method using combinational 
algorithm is used and it has the advantage in an automatic 
processing. 
These methods, however, have following disadvantages. 
Classification accuracy of clustering method is not 
sufficient in most applications, while the classification 
speed of a maximum likelihood method is slow and a very 
carful training data selection is necessary in order to 
maintain a high theoretical accuray. 
Authors developed a table look-up maximum likelihood 
method/1/, which can decrease classification times one 
fourth to seventh. However the total procedure from 
training data seletions to classifications consumes much 
more time, because of the second problem of training data 
- 623 -