EXTRACTION OF BUILDINGS 
FROM HIGH-RESOLUTION SATELLITE DATA AND LIDAR 
G. Sohn 
Dept. of Geomatic Engineering, University College London, Gower Street, London, WCIE 6BT UK - 
gsohn@ge.ucl.ac.uk 
Commission WG III/4 
KEY WORDS: IKONOS, LIDAR, Building Reconstruction 
ABSTRACT: 
Acquiring accurate detection and description of buildings is a difficult object recognition problem due to a high complexity of the 
scene content and the object representation. Since most urban scenes deliver very rich information, a robust separation of fore 
(objects to be reconstructed) from background (irrelevant features) is an essential process in object recognition system, but is 
difficult to achieve since objects in the scenes normally show a large amount of geometric or chromatic co-similarity across them. 
In addition, most mapping applications require building shapes to be reconstructed with a high degree of geometric freedom. 
However, information extracted from remotely sensed data is usually incomplete for reconstructing a full description of building 
objects due to limited resolving power of the sensor used, object complexity, disadvantageous illumination condition and 
perspective views. The research illustrates that the problems outlined above can be resolved by fusing multi-data sources, where 
2D linear features extracted from Ikonos images is attributed with a high-quality of 3D information provided by airborne lidar. 
1. INTRODUCTION 
With the recent advent of a series of commercialized high- 
resolution satellite, the potential of Ikonos imagery in 
topographic mapping has been investigated and highlighted 
by many researchers (Holland et al, 2002; Holland & 
Marshall, 2003). However, the success of fully automated 
reconstruction of building objects from the Ikonos imagery is 
still far to reach, and only partial solution in constrained 
environments have been reported (Kim & Muller, 2002; Lee 
et al, 2003). This research aims to develop a building 
extraction system which automatically reconstructs prismatic 
building models in an urban environment. In particular, two 
research interests have been exploited in this study; building 
detection (separation of objects to be reconstructed from 
irrelevant features) and building description (reconstruction 
of generic shape of building boundaries in a combination of 
data-driven and model-drive cues). 
2. DATA CHARACTERISTICS 
2.1 Ikonos image 
A “pan-sharpened” multi-spectral (PSM) Ikonos image 
covering the Greenwich industrial area was provided by 
Infoterra Co. for this research. The Ikonos PSM image is 
produced by combing the multi-spectral data with the 
panchromatic data, and resampled with l-metre ground 
pixel. The image was orthorectified by Space Imaging Co. to 
satisfy the positional accuracy (~1.9 metres) of Precision 
Product of Space Imaging. Figure 1 shows the Greenwich 
Ikonos PSM image, in which the red channel is replaced 
with the near-infrared channel while the green channel as 
red channel respectively. The sub-scene image is 681 x 502 
pixels with 1m resolution whose dimension of the image is 
approximately 341,862 m. 
     
a WALI TR 
P ACE S ase ur p 
Figure 1. Greenwich Ikonos PSM image 
   
2.2 Lidar DSM 
Figure 2 shows a lidar DSM which was also provided by 
Infoterra Co., which covers a sub-site of Greenwich 
industrial area with the size of 305,523 m’. The lidar DSM 
was acquired by the first pulse of OPTEC 1020 airborne 
laser sensor. The data has been converted from OSGB36 
(plan) and OSD Newlyn (height) to UTM/WGS84. The lidar 
DSM contains a total of 30,782 points, which corresponds to 
a point density of 0.1 (points/m?), i.e., one point per 3.2 X 
3.2 (m°). The height of the study area varies from 1.4 m to 
26.3 m. The terrain in the Northern part is higher than the 
Southern part, and the highest terrain height can be found in 
the North-West corner in figure 2. The Greenwich LIDAR 
DSM shows a typical urban environment, where a number of 
industrial buildings with different sizes spread over the 
study area. In particular, figure 2 shows the point density of 
the OPTEC 1020 LIDAR is not enough to properly represent 
the shape of those small houses though they are formed in 
planar roof surfaces. 
   
  
  
  
  
   
    
  
  
   
  
   
  
  
  
  
  
  
  
   
  
   
  
     
  
   
  
   
  
  
   
  
  
  
  
    
  
  
  
   
  
   
  
   
   
  
   
   
   
  
  
  
  
  
   
   
  
    
Interi 
The 
redui 
singl 
dete 
sequ 
not r 
31 
A Ik 
filter 
poin 
RTF 
optir 
gene 
Ihe 
plan: 
sub- 
cann 
desc 
Dow 
dete 
terra 
filtei 
lidar 
32 
Witl 
DTN 
than 
DTN 
“Tow 
3.3 
Sine 
às V 
The 
com 
char 
proj 
cons 
iden 
than 
3.4 
Isol: 
indi