a TI Rs, 
  
   
  
  
   
  
  
   
  
  
  
  
  
  
  
  
  
  
  
  
    
  
  
  
  
  
  
  
  
   
   
  
   
   
   
   
  
  
    
  
  
    
  
   
   
  
   
   
  
  
   
  
OBJECT-BASED CLASSIFICATION OF 
    
IKONOS DATA FOR RURAL LAND USE MAPPING 
M. Mori**, Y. Hirose*, Y. Akamatsu*, Y. Li® 
* Kokusai Kogyo Corporation, Hino Technical Center, 3-6-1 Asahigaoka, Hino, Tokyo 191-0065 Japan — 
(masaru_mori, yoko_hirose, yukio_akamatsu)@kkc.co.jp 
? Japan Space Imaging Corporation, 8-1 Yaesu 2-Chome, Chuo-Ku, Tokyo 104-0028 Japan — 
yunli@spaceimaging.co.jp 
KEY WORDS: Land Use, Object, Segmentation, IKONOS, Classification, Rural, Spectral, Texture 
ABSTRACT: 
Land uses in rural areas of Japan, which typically consist of small agricultural fields, complex vegetation covers, and sparsely 
distributed residential arcas, have been problematic in terms of land use mapping using satellite remote sensing data due to the 
complexity of the spatial structure. Increasing availability of very high resolution satellite (VHRS) data and the advancement of 
object-based image classification techniques in recent years provide a new opportunity for mapping detailed land uses from space. 
To utilize VHRS data and object-based image classification for creating land use maps, the correspondence between image objects 
extractable from VHRS data and land use classes observable on the ground, based on the spectral, spatial, and contextual 
relationships among image objects, must be established, because neither pixels nor segmented image objects of VHRS data directly 
correspond to land use classes into which we want to classify geo-spatial features on the ground. In this study, we conducted 
quantitative analyses of the spectral and spatial properties of image objects extracted from an IKONOS pan-sharpen image of a rural 
town in southwestern Japan, and proposed a classification framework for detailed land use mapping. Firstly, we extracted image 
objects from the IKONOS data using a multi-scale segmentation technique. Based on the statistical analyses of image objects, a 
hierarchical classification scheme for object-oriented land use classification was proposed. 
1. INTRODUCTION 
Very high-resolution satellites (VHRS) including IKONOS 
provide a useful way for us to periodically monitor detailed 
land use in broad areas. Pixel sizes on the ground of VHRS 
sensors are small enough to capture geometrical details of 
common land use patches, and most geographic features on 
VHRS images, which we want to map on land use maps, are 
represented as clusters of multiple pixels. For mapping land use 
from VHRS images, object-oriented classification methods 
utilizing image segmentation and contextual rule-based 
labelling techniques are expected to be more suitable than 
traditional pixel-by-pixel classifiers (Blaschke, T., Strobl 2002). 
To utilize VHRS data and object-based image classification for 
creating land use maps, the correspondence between image 
objects extractable from VHRS data and land use classes 
observable on the ground must be established, based on the 
spectral, spatial, and contextual relationships among image 
objects, because neither pixels nor segmented image objects of 
VHRS data directly correspond to land use classes into which 
we want to classify geo-spatial features on the ground. 
In this study, we conducted quantitative analyses of the spectral 
and spatial properties of image objects extracted from an 
IKONOS pan-sharpen image of a rural town in southwestern 
Japan, and proposed a classification framework for detailed 
land use mapping. Firstly, we extracted image objects from the 
[KONOS data using a multi-scale segmentation technique. 
Based on the statistical analyses of image objects, a hierarchical 
classification scheme for object-oriented land use classification 
was proposed. 
2. LAND USE IN RURAL PARTS OF JAPAN 
Figure 1 shows a schematic view of the land use pattern often 
seen in the rural parts of Japan, depicted on a panchromatic 
image taken from the IKONOS-2 satellite. A typical land use 
pattern in rural parts of Japan consists of houses (usually with 
backyards), agricultural fields (rice, wheat, etc.), forests, and 
networks of roads. 
In a macroscopic view, houses in a rural town in Japan tend to 
aggregate in certain areas of the town, and form residential 
areas. Residential areas are usually surrounded by agricultural 
fields growing rice, wheat, vegetables, etc. The boundaries 
between residential areas and agricultural fields are relatively 
clear for human interpreters. Road networks are not in a grid- 
like arrangement, and the shapes and sizes of parcels are not 
uniform. 
A house in the countryside usually has a backyard or a small 
plot of ground for growing vegetables. That is, image objects 
corresponding to the houses and agricultural fields in rural 
towns in Japan share relatively similar constituents. 
The relationship between houses and residential areas can be 
seen as a hierarchical structure of the land use pattern in the 
view of landscape ecology, and pixel-by-pixel-base image 
classifiers hardly produce optimal land use maps without 
intensive editing by human photo-interpreters. 
  
  
Providin 
Japan, ! 
hardly | 
without 
the simi 
among € 
the pro 
propertii 
spatial r 
mapping 
classifie 
based le 
mapping 
In the p: 
have bi 
classific 
study, 
Schápe