Satellite Image Analysis using Integrated Knowledge Processing 
K.-J. Schilling and T. Vögtle 
Institute for Photogrammetry and Remote Sensing 
University of Karlsruhe 
Germany 
schi@ipf.bau-verm.uni-karlsruhe.de 
http://www-ipf.bau-verm.uni-karlsruhe.de/Personen/schi/schi.html 
Commission Ill, Working Group 3 
KEY WORDS: Extraction, Knowledge-Base, Analysis, Automation, Land-Use, Satellite Image Analysis, Semantic Model, 
Integrated Knowledge-Based Processing 
ABSTRACT 
A new approach to satellite image analysis — based on integrated knowledge processing — for improvement of the (up to 
now often dissatisfying) results and automation of the analysis process will be presented. An expansion of feature base to 
object-oriented non-spectral features (shape, size, structure, relations) and a systematic structuring of knowledge in semantic 
networks (SN) are the main components of this concept. Digital topographic databases which are still created in many 
countries (e.g. ATKIS in Germany) can offer the necessary external knowledge for the semantic modelling of image contents. 
The analysis is based on a comparison of two specific SN containing knowledge about the topographic data and the image 
objects, respectifely (generic models 'database' and 'image"). In a first step unchanged objects are verified, while non-verified 
objects have to be related to their semantic meaning by a general classification procedure. The results of this image analysis 
will lead in a future step to a change detection and a consecutive update process of the digital database. 
KURZFASSUNG 
Ein neuer Ansatz zur Satellitenbildanalyse — basierend auf einer integrierten Wissensverarbeitung — soll hier vorgestellt werden, 
mit dem Ziel der Verbesserung der (bisher oft unbefriedigenden) Auswerteergebnisse einerseits und einer Automatisierung des 
Auswerteprozesses andererseits. Eine Erweiterung der Merkmalsbasis hin zu objektorientierten nicht-spektralen Merkmalen 
(Form, GróBe, Struktur, Relationen) und die systematische Strukturierung des Wissens in Semantischen Netzen (SN) sind die 
Hauptkomponenten dieses Konzeptes. Digitale topographische Datensatze, wie sie im Augenblick in vielen Làndern aufgebaut 
werden, können das notwendige externe Wissen fiir eine semantische Modellierung des Bildinhaltes liefern. Der Analyseprozeß 
stützt sich auf den Vergleich zweier spezifischer Semantischer Netze, wobei eines das spezielle Wissen über die topographische 
Datenbasis, das andere über die Bildobjekte enthàlt (generisches Modell 'Datenbasis' und ’Bild’). In einem ersten Schritt 
werden unveránderte Objekte verifiziert, wahrend nicht-verifizierten Objekten ihre semantische Bedeutung in einem allge- 
meinen KlassifizierungsprozeB zugewiesen wird. Die Ergebnisse der Bildanalyse sollen in einem zukünftigen Schritt zu einer 
Anderungsdetektion und einem anschlieBenden FortführungsprozeB der digitalen topographischen Datenbasis führen. 
  
1 MOTIVATION 
spectral features * Signature 
Conventional methods for satellite image analysis, such as * texture 
multispectral classification, to achieve a semantic descrip- 
tion of image objects, have undoubtedly reached their limits. 
  
This may be caused by the restriction to only one feature, non-spectral features * shape 
the spectral signature, and the strictly (single-)pixel-based * Size 
processing without taking the spectral behaviour of neigh- * Structure 
bouring pixel into account. First approaches were done using x relations 
local neighbourhood information, for instance by introducing 
texture or relaxation techniques. But a significant progress 
can be reached by an expansion of feature base to more then 
one and the introduction of additional non-spectral object- 
oriented features. 
  
  
  
  
Image objects can differ significantly in shape and size, e.g. 
rivers, settlement or forest areas. Also different object struc- 
tures may occur. For instance, urban areas get a specific 
Beside this improvment of satellite image analysis the au- structure by the street system and the building alignment, 
tomation of the analyzing process should be the second aim which is totally different to the structure of ’natural’ objects 
of this work. A new approach, which combines both postula- like forest areas. Relations between objects (e.g. settlement- 
tions by integrated knowledge processing, will be presented. streets, farm-agricultural areas) can be expressed in rules, 
which can improve ambiguous classification decisions. How- 
2 CONCEPTION ever, all this features are more or less fuzzy or uncertain in- 
formation, so this has to be taken into account during the 
In a first step the spectral feature base is expanded to non- decision process. 
spectral features, which are relevant to improve the satellite 
image analysis ((VoEGTLE, ScHILLING 1995)). These fea- 
tures will be object-oriented: 
To fulfill the requirements of automation these features 
should be extracted without interaction of a human oper- 
ator. Therefore, the representation and integrated processing 
752 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
   
  
    
  
   
   
   
    
  
   
  
  
  
  
    
   
   
   
   
  
   
  
  
  
  
  
    
   
    
    
   
    
    
   
    
   
   
    
    
   
   
    
    
  
  
   
   
    
   
  
of e 
nece 
nal | 
mea 
imag 
Nort 
or w 
surv 
grap 
ent c 
Kart 
one 
mati 
are | 
Lanc 
1:2 
(:: 
used 
the i 
data 
avail 
In a 
DLM 
settle 
eas ( 
a-pri 
for tl 
lated 
appre 
analy 
In co 
extra 
quite 
tor cl 
durin 
terpr 
exper 
turba 
inside 
edge 
tion c 
to ex 
but t 
Ther: 
ried c 
tolerz 
major 
are v: 
Becaı 
featu 
image 
candi 
featui 
for th 
knowl 
can b 
datab 
3.1 
Ther 
growi