twiler, D., 1995. 
2-D and 3-D im- 
Computer Vision, 
-oximations with 
ature extraction. 
4), pp. 349—376. 
ires: À differen- 
n Conference on 
92. Multi-scale 
n without edges. 
Aboratory, Cam- 
  
SEGMENTATION OF OPTICAL SATELLITE IMAGERY USING SPATIAL SUBPIXEL ANALYSIS* 
Joachim Steinwendner 
Institute for Surveying and Remote Sensing 
Universitat fiir Bodenkultur 
(University of Agriculture, Forestry and Renewable Natural Resources) 
Austria 
joachim@mail.boku.ac.at 
Commision Ill, Working Group 2 
KEY WORDS: Pixel, Satellite Image, Segmentation, Spatial Subpixel Analysis 
ABSTRACT 
Due to low resolution of satellite imagery in relation to the size of observed objects the problem of mixed pixels arises. 
These pixels have spectral signatures being a combination of two or possibly more pure spectral signatures of objects (e.g. 
agricultural parcels, roads, etc.). In this contribution, a satellite image is segmented using standard segmentation algorithm 
to obtain mean pixel values of regions. The second step uses “edgel chains” to obtain region boundaries to subpixel accuracy. 
These two steps deliver the parameters for correcting the mean pixel values of regions by subpixel analysis. 
KURZFASSUNG 
Die niedrige Auflosung von Satellitenbildern in Relation zu den zu untersuchenden Objekten verursacht das Problem von 
Mischpixeln. Die spektralen Signaturen dieser Pixel bestehen aus einer Kombination von zwei oder mehr reinen Signaturen 
von Objekten (z. B. Landwirtschaftsflàchen, StraBen, etc.). Dieser Beitrag behandelt dieses Problem in zwei Schritten. Das 
Satellitenbild wird mit herkommlichen Segmentierungsalgorithmen segmentiert. Im zweiten Schritt werden sogenannte "edgel 
chains" verwendet, um Objektgrenzen in Subpixelgenauigkeit zu erhalten. Die Kombination dieser beiden Ergebnisse liefert 
die Parameter für die Verbesserung der mittleren Pixelwerte von Objekten durch ràumliche Subpixelanalyse. 
1 MOTIVATION AND INTRODUCTION 
In [1], a physical model for the image acquisition process 
is discussed and formulated. The inversion of the model is 
used for remote sensing image understanding. The model 
transforms the reflectance values of objects (regions on the 
terrain surface) to pixel values in the image. It is advisable 
not to work on a pixel-by-pixel basis, but rather to segment 
the image into "region tokens" (i.e. sets of pixels belonging 
together in some sense) prior to the inversion process. A 
feature vector for each region token is produced containing 
necessary image parameters for the physical model. Possible 
parameters are mean pixel intensity, pixel variance, etc. 
There are at least two advantages of the segmentation 
approach: The amount of information to be processed in 
the model inversion step is reduced, and the "mixed pixel 
problem" can be managed if a proper subpixel segmentation 
procedure is employed. 
New sensors with higher resolution are being developed ques- 
tioning the need for spatial subpixel information. However, 
the following arguments speak for the usefulness of spatial 
subpixel analysis: 
e Availability of satellite images is a problem when used 
operationally considering bad weather conditions or 
other impairments. It may thus be necessary to be 
  
* This work is supported by a grant from the Austrian “Fonds zur 
Fórderung der wissenschaftlichen Forschung" (No. S7003). 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
able to use images from any sensor available. 
e The objects to be examined can be smaller, e.g. trees, 
houses, etc. 
In the first part of this contribution, standard segmentation 
algorithms well known to the computer vision community are 
applied to remotely sensed imagery. There is no classifica- 
tion involved. The segmentation is based solely on spectral 
homogenity properties of pixels in a region. 
The second part deals with finding object borders with 
subpixel accuracy. Edgel chains provide the means to find 
those borders. Spatial subpixel analysis is applied to correct 
the mean pixel values of regions. 
The segmentation and the production of edgel chains do not 
influence each other so that they can be executed in parallel 
to increase the speed of the process (see also figure 1). 
2 STANDARD SEGMENTATION ALGORITHMS 
FOR SATELLITE IMAGERY 
A pixel in the interior of a region has only small intensity dif- 
ferences to the other pixels in the region. Thus, spatial sub- 
pixel analysis of interior pixels is not practical. Even though it 
might be a mixed pixel, spatial subpixel analysis achieves no 
improvement. Segmentation of the image delivers boundary 
pixels neighboured to pixels with high intensity differences in 
their spectral signatures.