International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
  
  
  
Figure 4: local maximum DRV image. 
(13 by 19 pixels size building) 
In the present work, the extraction process is limited of three 
building sizes (13*19, 19*13 and 13*13 pixels) corresponding 
to the approximate majority building sizes located on the 
original image. All the accurate building sizes and orientations 
are not taken into account. Moreover, just one study zone, is 
tested. Thus, the results should be considered as preliminary. 
The local maxima DRV associated to the three building sizes 
(13*19, 19*13 and 13*13 pixels) are individually extracted. 
Then, they are gathered to build a single local maxima DRV 
image. If different local maxima DRV exist for the same pixel, 
the highest DRV value is conserved. Then a final number of 77 
local maxima DRV points are obtained (figure 5). 
2.6 Evaluation and discussion 
In order to evaluate the preliminary DRV building extraction 
results, the DRV image is plotted over the ground truth image 
(figure 5). 
  
  
  
5 = = a = " 
= a n = a = 
3 s 5 5 . » "o 
local maxima DRV ; building (19*13 pixels) 
Bü local maxima DRV ; building (13*19 pixels) 
Bl local maxima DRV ; building (13*13 pixels) 
[1 building (ground truth) 
  
  
ME 
Figure 5. Building extraction results 
Some comments can be derived from figure 5: 
e 53 of the 69 buildings (77 %) are extracted. Some errors 
can be noticed on extracted building size and building accurate 
location. For example one 13*19 is extracted as “13*13” pixels 
building size, while some local maxima RDV are not accurately 
located in the centre of building (confusion between building 
borders and building shadow borders). 
* 16 of the 69 buildings (23 %) are not extracted. Parts of 
these 16 buildings have different sizes than the three selected 
“search size”. It is supposed that these buildings would be 
extracted by an adapted "search zone" size. 
e 23 of the 77 maxima variances values (30 %) are not 
associated to buildings on the ground truth image (13 of these 
23 commissions errors correspond to roads). Moreover, further 
tests showed that these commission errors increase with the 
number of "search zone" sizes used for the building extraction 
process. 
These preliminary results show that the DRV is an interesting 
tool to locate building's centres. The major problem seems to be 
the high number of local maxima DRV points, which are not 
associated to building location. In addition, this extraction 
process is limited to the basic building shapes (square, 
rectangle) with specific orientations. 
Indeed this building extraction approach is only based on 
panchromatic variance values. It is supposed that the use of 
other image features would improve significantly the results, 
especially by decreasing the commission errors. 
The methodology using the DRV directly computed from the 
original variance image (image (b), figure 2) has been also 
tested. Results are similar to these presented here. However, 
small differences are noticed. The binary image variance allows 
a better extraction of buildings with different roof colours. At 
the opposite, the original image variance allows a better 
extraction of buildings with low contrast borders. 
Additional tests have to be carried out in order to define the 
variance image (original or binary) to be used for DRV 
computing. 
3. CONCLUSION 
This paper deals with buildings extraction from very high 
spatial resolution satellite images. An original detection 
approach of building’s centres is proposed, only based on a 
parameter (DRV) taking into account jointly the variance of 
building and its close neighbourhood. This parameter is tested 
to extract buildings centres from a panchromatic IKONOS 
image in an urban area. Preliminary results show that the DRV 
is an interesting tool for building detection, although the 
methodology needs additional developments. The major 
problem is the high commissions errors, which could be reduced 
by using additional spatial and spectral information or features. 
REFERENCES 
Bianchin A., Bravin L., “land use in urban context from 
IKONOS image : a case study”, The International Archives Of 
The Photogrammetry, remote Sensing and Spatial information 
Sciences (CD-ROM), Vol XXXIV-7/W9, Regensburg, 
germany, 27-29 June 2003. 
Guindon, B “A Framework for the Development and 
Assessment of Object Recognition Modules for High 
Resolution Satellite Images”. Canadian Journal of remote 
Sensing, vol 26, No 4, 2000, p 334-348. 
Hofmann, P. "detecting urban features from IKONOS data 
using an object-oriented approach”, joint Worshop urban 2001, 
May 22 ‚2001. 
Jacobsen, K. “Mapping with IKONOS images”. Proceedings of 
22nd EARSeL Symposium, Prague, 2002, 7 p 
Lhomme, S., He D.C., et Morin, D. « Évaluation de la qualité 
d'une image Ikonos pour l'identification du báti en milieu 
urbain », Télédétection vol. 3, no 5, 2004. 
    
   
  
  
  
  
  
  
      
   
    
   
   
   
   
   
    
   
   
    
   
    
   
   
   
  
   
      
   
    
   
    
   
   
    
  
    
   
  
   
  
     
  
  
Inte 
Pes 
[EE 
20€ 
Pui 
Res 
Inte 
Sot 
resi 
Ma 
201 
Tee 
Mo 
Res 
Tec 
On. 
No 
Var 
infc 
sub 
Tra