AUTOMATIC LAND-COVER CLASSIFICATION OF LANDSAT IMAGES 
USING FEATURE DATABASE IN A NETWORK 
G. W Yoon*. S. L Cho, G. J. Chae, J. H. Park 
ETRI Telematics Research Group, Daejeon, Korea — (gwyoon, chosi, cbase, jhp)@etri.re.kr 
KEY WORDS: Classification, Land cover, Database, Networks, Segmentation, Object, Software 
ABSTRACT: 
In order to utilize remote sensed images effectively, a lot of image classification methods are suggested for many years. But the 
accuracy of traditional methods based on pixel-based classification is not high in general. And, in case of supervised classification, 
users should select training data sets within the image that are representative of the land-cover classes of interest. But users feel 
inconvenience to extract training data sets for image classification. In this paper, object oriented classification of Landsat images 
using feature database is studied in consideration of user's convenience and classification accuracy. Object oriented image 
classification, currently a new classification concept, allows the integration of a spectral value, shape and texture and creates image 
objects. According to classification classes, objects statistics such as mean value, standard deviation, tasselled cap transformation 
and band ratio component were constructed as feature database. The feature of seven classes (Rural, Forest, Grass, Agriculture, 
Wetland, Barren, Water) was constructed in this study, it will be served in a network to user for image classification training data 
sets. Proposed method will be higher classification accuracy than that of traditional pixel-based supervised classification and gives 
convenient environment to users. 
1. INTRODUCTION 
The remote sensing technology is currently being offered a 
wide variety of digital imagery that covers most of the Earth's 
surface. This up-to-date image data is a promising tool for 
producing accurate land cover maps. To maximize the benefit 
of such data, automatic and efficient classification methods are 
needed. To achieve this objective, pixel-based classification has 
been extensively used for the past years. Currently the 
prospects of a new classification concept, object-based 
classification, are being investigated. Recent studies have 
proven the superiority of the new .concept over traditional 
classifiers (Each, 2003; Darwish, 2003; Mitri, 2002; Niemeyer, 
2001; Sande, 2003). The new concepts basic principle is to 
make use of important information (shape, texture and 
contextual information) that is present only in meaningful 
image objects and their mutual relationships. 
In order to obtain image objects, classification software is 
developed by ours. It gives convenient environment to non- 
specialists, because operated automatically. And, feature 
database is constructing for automatic land cover classification. 
Feature database has information of seven class (water, rural, 
barren, wetland, grass, forest, agriculture) features in Landsat 
images. Proposed method will be higher classification accuracy 
than that of traditional pixel-based supervised classification and 
gives convenient environment to non-specialist users. 
2. OBJECT ORINTED CLASSIFICATION 
The object oriented classification concept is that important 
semantic information necessary to interpret an image is not 
represented in single pixels, but in meaningful image objects 
and their mutual relations. Image analysis is based on 
contiguous, homogeneous image regions that are generated by 
initial image segmentation. Connecting all the regions, the 
o 
  
image content is represented as a network of image objects. 
These image objects act as the building blocks for the 
subsequent image analysis. In comparison to pixels, image 
objects carry much more useful information. Thus, they can be 
characterized by far more properties such as form, texture, 
neighbourhood or context, than pure spectral or spectral 
derivative information (Baatz, 1999). 
2.1 Segmentation 
Adjacent, similar pixels are aggregated into segments as long as 
the heterogeneity in the spectral and spatial domains is 
minimized in this step. Neighbouring segments are fused to a 
new segment if the resulting heterogeneity is minimized and 
below a specified level. The definition of heterogeneity is 
flexible and consists of a trade-off between homogeneity in the 
spectral domain (e.g. backscatter values in various channels) 
and form/shape in the spatial domain. Homogeneity in the 
spectral domain is defined by a weighted standard deviation 
over the spectral channels. Homogeneity of shape depends on 
the ratio of an object’s border length to the object’s total 
number of pixels (compactness), and the ratio between the 
lengths of an object's border to the length of the object’s 
bounding box (smoothness). Compactness is minimum for a 
square; smoothness is minimum if the object borders are not 
frayed (Benz, 2001). 
2.2 Classification 
Usually classifying means assign a number of objects to a 
certain class according to the class’s description. Thereby, a 
class description is a description of the typical properties or 
conditions the desired classes have. The objects then become 
assigned (classified) according to whether they have or have not 
met these properties/conditions. In terms of database language 
* Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author. 
564 
Interne 
one ca 
which 
and th 
class « 
maxim 
thereb 
wheth: 
classif 
expres 
manne 
and/or 
probat 
memb 
expres 
to a 
improl 
depen 
descril 
3.1 8 
The st 
the Ki 
appro: 
areas, 
Classi 
acquis 
(1:50, 
produ: 
compr 
Wetla 
Septeı 
study.