tope type 
| 287 
] 4 
] 5&9 
| 6 
1e 
ur und 
hotograph 
for 
| - 189. 
Bereich 
anung, 
FEATURES AND THEMES 
OVERVIEW OF FEATURE EXTRACTION AND SPECTRAL CLASSIFICATION 
AND 
THE IMPACT OF HIGH SPATIAL RESOLUTION SATELLITE IMAGE DATA 
Lihua Li & Bruce C. Forster 
School of Geomatic Engineering 
The University of New South Wales 
Sydney, 2052, Australia 
ISPRS Symposium, Budapest, 1998 
Commission VII, WG 3 
KEY WORDS: 
vision, GIS 
ABSTRACT 
Feature extraction, spectral classification, high resolution, photogrammetry, remote sensing, computer 
Acquiring data from digital images is sometimes referred to as feature/object extraction or image classification by 
practitioners in the fields of computer vision, photogrammetry and remote sensing. Algorithms for feature extraction have 
been developed during the last few decades by researchers in all three fields, however with different focuses or approaches. 
In this paper, the authors review the feature extraction techniques and evaluate the possibility of integrating them for use in 
high resolution images for GIS applications. 
1. INTRODUCTION 
Remote sensing technology has been used for over a 
century by the user community to acquire data of the 
earth surface. It provides a variety of data with a range of 
spectral, spatial and temporal resolutions, that can be 
used in the management and policy making process. GIS 
technology is now commonly used throughout 
government organizations and industry for managing 
land-related or spatial data and providing information for 
decision making. The "real-world" features or entities for 
which data is collected and stored in these systems can be 
either natural or built. Remote sensing and GIS 
integration, by either using remotely sensed data for 
updating GIS information or using information stored in 
GIS as an auxiliary data set for improving the information 
extraction potential of remote sensing data, is proven to 
be an advantage to both technologies. However, there are 
certain issues affecting the integration due to the 
historical development of these two technologies. Firstly, 
the availability of data, especially of high resolution 
remote sensing data, suitable for urban applications, has 
been one of the major obstacle to the integration. 
Secondly, there are certain technical issues to be resolved 
before real integration can take place, these are: error 
propagation, computational and algorithms improvement, 
and most importantly, automatic data acquisition 
techniques from remote sensed imagery. 
The new high resolution satellite remote sensing systems 
will have a significant impact on the user community. 
Data for urban applications will be readily available, new 
products will emerge from the synthesis of electro- 
optical, multispectral, hyperspectral, radar and infrared 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
imagery. However, new algorithms for capturing data 
from the high resolution images are needed to handle the 
increased level of both spatial and spectral details. 
Acquiring data from digital images is sometimes referred 
to as feature/object extraction or image classification by 
practitioners in the fields of computer vision, 
photogrammetry and remote sensing. Algorithms for 
feature extraction have been developed during the last 
few decades by researchers in all three fields, however 
with different focuses or approaches. Research in both 
photogrammetry and computer vision have mainly 
concentrated on detecting edges of the objects or features 
of interests by defining the geometric properties of and 
the relationships between the object classes, using mainly 
single band digital images of high resolution, whilst in 
remote sensing most algorithms have been developed 
based on spectral signatures or spectral characteristics of 
the object classes, using images of coarser resolution, at 
most 10 meter resolution. Previous studies on existing 
image systems have shown that resolution has either a 
positive or negative impact on spectral classification 
accuracy. Dependent on the degree of spatial variability 
in the spectral response associated with classes under 
consideration, the conventional multispectral 
classification techniques cannot adequately classify the 
land-cover or land-use from high resolution satellite data. 
In this paper, the authors review the feature extraction 
techniques developed by researchers in these three areas 
and evaluate the possibility of integrating them for use in 
high resolution images for GIS applications. 
  
245