ROAD NETWORK EXTRACTION FROM SAR IMAGERY SUPPORTED BY CONTEXT 
INFORMATION 
B. Wessel 
Photogrammetry and Remote Sensing, Technische Universitaet Muenchen, 80290 Muenchen, Germany 
birgit.wessel@bv.tum.de 
Commission III, WG III/4 
KEY WORDS: Mapping, SAR, Object, Extraction, context-based, Automation. 
ABSTRACT 
This paper deals with automatic road extraction from SAR imagery. In general, automatically extracted road networks are not complete, 
i.e., gaps remain in the erxtracted network. Especially in SAR imagery many objects occlude road sections and cause gaps, due to the 
side looking geometry of the SAR sensor. In this paper an approach for automatic road extraction is proposed that is optimized for 
rural areas by using additional explicitly modeled knowledge about roads and the context of roads. Roads are modeled as a network. 
Context of roads is hierarchically structured into a global and a local level. Local context objects like trees or vehicles can interfere 
road extraction due to the layover effect or the motion, but they can also support it. It is shown that the incorporation of local context 
objects into the extraction improves the results by bridging smaller gaps. Though the approach is restricted to rural areas, other global 
context regions can provide additional information, too. Here, urban areas are used to deliver confident seed information for the road 
network generation, because it is the characteristic and function of roads to connect urban areas with each other. With this information a 
more complete network is extracted. Furthermore, a new approach for highway extraction is proposed based on a multi-scale modeling. 
Because of the larger dimensions of highways and the more salient substructures, like the crash barriers, a more detailed model and 
extraction strategy is needed. Finally, examples and results are given, showing the potential of using context information and explicit 
modeling of roads for automatic road extraction from SAR imagery. 
1 INTRODUCTION 
Up-to-date road network data are required by many applications, 
e.g. topographic mapping, traffic monitoring, or navigation. Ex- 
tensive research has been done to automate the extraction and up- 
date of road networks. For automatic road extraction from images 
with low resolution the most common techniques rely on the de- 
tection and following of lines. Because SAR is (almost) indepen- 
dent from weather and daylight conditions, the extraction of road 
networks from SAR imagery received notable attention over the 
past years. Beside the advantages of SAR, road extraction from 
this type of imagery has also to cope with a lot of difficulties, es- 
pecially with well-known SAR inherent effects like layover and 
shadow caused by adjacent objects. 
In this paper, an approach for automatic road extraction from 
SAR imagery is presented, that is based on the extraction of lines 
and explicitly modeled knowledge of roads. The approach is ex- 
tended by the use of context information and, moreover, by a 
more detailed model for highways. Context information means 
knowledge about the road and its spatially neighbored objects. 
The relations are modeled on a local and on a global level. Lo- 
cal context objects influence the appearance of a road section, 
e.g., vegetation or buildings with its corresponding layover and 
shadow. The global context region residential areas are always 
connected by roads. Therefore, the contours of residential areas 
allow to provide confident seed information for the road extrac- 
tion. 
In this paper the road extraction approach is successively ex- 
tended by three add-ons. In the first extension, different local con- 
text objects are exemplarily introduced into the extraction. Then, 
in the second extension, global context regions are introduced. 
In the last extension, the approach is extended by a more detailed 
model for the road class "highway". Because of the larger dimen- 
360 
sions of highways and the more salient sub-structures (e.g. crash 
barriers) special focus is on multi-scale modeling and incorporat- 
ing reflection characteristics. The paper starts with a model for 
each extension (Sect. 2). Then, extraction methods are outlined 
(Sect. 3) and finally, the results are evaluated for each extension 
(Sect. 4). 
2 MODEL FOR ROADS AND CONTEXT 
INFORMATION 
The appearance of roads in digital imagery strongly depends on 
the sensors spectral sensitivity and its resolution in space. The 
proposed model is restricted to SAR imagery with a resolution of 
about 2 m per pixel. The used imagery are multilook X-band data 
with a ground resolution of about 2 m (E-SAR, DoSAR) and up 
to | m resampled AeS-1 data, originally 0.5 m. The model below 
describes in the first part the characteristic properties of roads 
(Sect. 2.1). The second part defines different context information 
levels (Sect. 2.2) and in the third part a complex model for the 
object "highway" in SAR imagery is given (Sect. 2.3). 
2.1 Roads in open rural areas 
In a resolution of about 2 m roads appear mainly as lines with the 
following characteristics: 
e Radiometric characteristics: low and homogeneous reflec- 
tion (because the smooth surface leads to total reflection) 
e Geometric characteristics: low curvature, constant width, 
elongated segments (because of road construction constraints) 
e Topological characteristics: roads form a network 
   
   
    
  
   
  
   
    
  
   
  
  
  
  
  
  
  
  
  
  
  
  
  
    
   
    
   
    
   
   
   
      
     
   
   
   
   
   
   
   
   
   
   
  
    
  
  
     
    
    
    
   
   
    
  
Internat 
  
(a) Lay 
road se 
In open 
ably we 
be extra 
scene ir 
pear fre 
regions, 
directio 
an extre 
rather v 
count. 
22 C 
The mo 
Context 
the obje 
chler, 1¢ 
like in ( 
e Lo 
roc 
e Gl 
jec 
et 
fec 
Here, b 
mining 
imagery 
extractic 
The fals 
carefull; 
2.2.1 | 
objects 
followir 
® 10d 
e bri 
bai 
act 
e lar, 
Many o! 
tering c: 
dividual 
These p 
in AeS- 
positive 
to be ex