International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
| one can say the feature space is segmented into distinct regions 
which leads to a many-to-one relationship between the objects 
and the classes. As a result each object belongs to one definite 
class or to no class. Classic classifiers in remote sensing (e.g., 
maximum-likelihood, minimum-distance or parallelepiped) 
thereby assign a membership of 1 or 0 to the objects, expressing 
whether an object belongs to a certain class or not. Such 
classifiers are usually also called hard classifiers since they 
express the objects’ membership to a class only in a binary 
manner. In contrast, soft classifiers (mainly fuzzy systems 
and/or Bayes classifiers) use a degree of membership/a 
probability to express an object’s assignment to a class. The 
membership value usually lies between 1.0 and 0.0, where 1.0 
expresses full membership/probability (a complete assignment) 
to a class and 0.0 expresses absolutely nonmembership / 
improbability. Thereby the degree of membership/probability 
  
But the | depends on the degree to which the objects fulfill the class- 
fication, | describing properties/conditions (Baats, 2002). 
sers feel | 
images 3. EXPERIMENT 
] image | 
s image 3.1 Study area and data 
yrmation 
iculture, The study site is located around Daejeon city in middle part of 
ing data | the Korean peninsular as shown Figure 1. The area measures 
nd gives approximately 575 [1 and comprises rural areas, agriculture 
areas, forest areas and different areas. 
Classification for the area was performed using Landsat TM 
acquisitions of 13 March 2000 (Figure 1). And, large-scale 
Figure 2. Land cover map as reference of study area 
  
  
  
objects. (1:50,000 scale) land cover map which ministry of environment 
A ia 2 : e om Path/row Date Purpose 
for the produced in Korea was used as reference map (Figure 2). It is I3M 
5, image comprised of seven classes (Rural, Forest, Grass, Agriculture, Landsat TM 115/35 so Classification 
y can be Wetland, Barren, Water) and manufactured based on 2 s : 
texture, September 1998 of Landsat TM. Table 1 show used data in this Land cover map - e Reference 
spectral study. 
  
  
  
  
  
  
Table 1. Used data in this study 
3.2 Feature database construction 
s long as 
nains is 
ised to a 
ized and 
eneity is 
ity in the | 
We must select training data in supervised classification. As 
images are classified based on training data, we select training 
sites within image that are representative of the land cover 
classes of interest. The training data should be of value if the 
environment from which they were obtained is relatively 
homogeneous. However, if the land cover conditions should 
channels) > E 
y in the | change dramatically across the study area, training data of 
deviation partial in study area would not be representative of spectral 
Es ... es ~ 
pends on conditions. So, we have to select training data carefully and 
widely in study area. And, the general rule is that if training 
data are being extracted from n bands then >10n pixels of 
training data are colleted for each class. This is sufficient to 
compute the variance-covariance matrices required by some 
classification algorithms (Jensen, 1996). 
It is indispensable step to extract training data in supervised 
classification, but it requires substantial human operations. Such 
substantial human operations make training data selection a 
time-consuming and laborious processing. The specially, as 
many satellite images are processing, thus operations will be 
troublesome more and more. In this paper, feature database is 
introduced to reduce a laborious processing and obtain high 
classification accuracy. Feature database has statistics 
ir Ot calculated training data. As feature database constructed by 
| | specialists provide to non-specialists, they can have the 
ct's total 
ween the 
object’s 
um for a 
s are not 
ects to a 
hereby, a 
perties or 
1 become | 
- have not 
language 
  
  
  
^ : R A advantage of convenience and accuracy. 
Figure 1. Location of the study area and Landsat TM image advantage of convenience and : Sv. 
€ 
1€ author.