Istanbul 2004 
Ww sizes. Our 
nside a small 
a minimum 
of grey level 
of grey level 
| distribution 
bution. 
g periphery" 
istribution of 
pond to the 
nding to the 
by the mean 
nding to the 
itified by the 
lowever this 
/ the variance 
d, it does not 
the different 
iew condition 
/ distributed 
| periphery" 
uirement the 
orresponding 
ure 1). Then, 
he results are 
ases the total 
of a uniform 
periphery" 
ntrated on a 
Itiplication is 
ame range of 
quation 2). 
and of the 
jointly. This 
an “building 
ly" variance 
ilding and its 
ue parameter 
is called 
*(MeanVarS4) 
(2) 
n the side |, 
Iding body” 
hows a high 
y variance of 
So the DRV 
m other land 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
2.5 Application 
The described methodology is applied to an IKONOS 
panchromatic image in order to extract the building centre. The 
study site is firstly presented. The proposed methodology is 
detailed into three steps: (2.5.2) variance values computing; 
(2.5.3) DRV values computing and (2.5.4) building extraction. 
2.5.1 Study site: A small image (388 by 190 pixels) 
extracted from an IKONOS image of Sherbrooke City (Quebec, 
Canada) is used. It is acquired on May 20, 2001 at 10h50 (local 
time). Only the panchromatic band (1 metre resolution) is used 
(image (a), figure 2). The methodology is tested on this site 
because it contains buildings with similar sizes and orientations 
but with different roof colours. Moreover a ground truth is 
available. 
2.8.2 Variance computing: The variance is computed over 
the image by a moving window of 3 x 3 pixels. For each 
window position the computing result (detailed in equation 1) is 
assigned to the central pixel. This operation is carried out on 
ENVI software (3.6 version, Copyright © 2002, research 
Systems, Inc). The calculated variance image contains a wide 
range of variance values (0 to 13918) (image (b), figure 2) 
  
+ 
bd Et 3 + 
: original image 
  
  
   
  
ad Br 
(b) : variance image 
(darker pixels have lower values) 
  
- ^ B 
(c): binary variance image : 
( white: high values) 
  
  
—À 
  
Figure 2. Variance computing 
o 
With very high spatial resolution satellite images in urban arca, 
the local variance can be very important due to noise or 
"foreign" elements (metallic chimney on a roof, vehicle on a 
road...) or very low due to the shadowing. In order to avoid 
these potential problems, the variance values are reduced into 
two values (image (c), figure 2). A binary process is applied to 
the variance image by a simple threshold operation. The used 
threshold limit corresponds to the median value of the variance 
image. It is important to note that this step of processing does 
not need any intervention of the operator, so the whole process 
can be automatically realised. 
2.8.3 DRV values computing: The global process of DRV 
values computing that can be applied to all the building sizes is 
presented. In this example only the case of a 15 by 21 pixels 
"search zone" is showed, corresponding to the building size of 
13 by 19 pixels. 
From the binary variance image, the mean variance values are 
calculated for each of five interest zones (the body and the 4 
periphery sides). This is generated by a simultaneous 
convolution of the five masks (corresponding to the first zone of 
the “building body" and to the four zones for the "building 
periphery"). The operation is carried out over the whole image 
and the DRV is then computed (figure 3) as indicated in the 
equation (2) for each image point and assigned to the central 
pixel of the "search zone" (red square, figure 1). 
In the case of presence of several different building sizes, this 
DRV computing process should be repeated for each 
corresponding "searching zone" size. 
  
Figure 3. DRV image 
(13 by 19 pixels size buildings; darker pixels have lower values) 
2.5.4 Building extraction: The DRV values can be directly 
used to extract buildings without any additional spatial or 
spectral information or feature. 
First of all, the low DRV values are eliminated by a threshold 
operation in order to erase pixels that do not correspond, 
unquestionably, to a building. Threshold value is selected by 
"training". 
A new image containing only DRV values higher than the 
threshold value is obtained. Then, the local maxima DRV 
values (figure 4) are extracted from this image. These local 
maxima are supposed to indicate building's centres. It should be 
noted that, as for the previous steps, these local maximum DRV 
values are associated to a specific "search zone" size.