ih ES M 
M. 
do 
on 
of 
xn 
els 
30- 
are 
nts 
ıld 
ith 
to 
its) 
ion 
ata 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXX V, Part B7. Istanbul 2004 
  
additional information from a GPS (Global Positioning System) 
and IMU (Inertial Measurement Unit) device are recorded at a 
very high temporal resolution, providing information about the 
exact position and orientation of the sensor at the time of 
acquisition. This so-called direct georeferencing is crucial, as 
due to the dynamic acquisition principle and due to 
unsteadiness of the airplane movement each scanline has its 
individual exterior orientation and a conventional orientation 
procedure would not be able to deliver the required accuracy. 
The typical process of this direct georeferencing is described in 
Fricker (2001). Next, after automatic tie-point measurement 
triangulation can be applied. It is recommended to use some 
GCPs at least at the block corners, but as shown in the excellent 
results from Scholten and Gwinner (2003) it is not obligatory, 
since the collected orientation data from the GPS and IMU unit 
seem to be sufficient. 
Once the images are accurately oriented (multi-image) 
matching algorithms can be applied for deriving a DSM. 
Finally, the images and the DSM can be used to create an 
orthophoto of the scene, which will also be used as input for the 
building extraction algorithm. 
At the DLR (German Space Agency) the processing chain for 
handling airborne three-line scanner imagery has been 
automated to such a degree, that human interaction is only 
needed at the data acquisition step (Scholten and Gwinner, 
2003). Once the data is downloaded onto the computers the 
whole process from image correction and rectification, to 
orientation and triangulation, and to the DSM and orthophoto 
production can be done fully automatically. 
4.2 Building Recognition 
The most important input for the creation of a Digital City 
Model is the DSM. But as mentioned previously, the resolution 
and accuracy if derived from spaceborne or airborne line 
scanning imagery is rather low (compared to Lidar-DSMs). 
Hence the images or orthophotos if available, will be used to 
proceed to building extraction. 
The first step is to create a so-called normalized Digital Surface 
Model (nDSM), which is the difference-model between the 
DSM and the DTM. Hence all objects in the nDSM stand on 
elevation height zero. 
There are multiple approaches for getting the nDSM. An 
approved method is to filter the existing DSM-data by applying 
a skewed error distribution (Kraus, 2000). This procedure is 
fully automated and the user needs only to input the parameters 
of the skewed error-distribution in order to calculate the nDSM, 
in which vegetation and buildings are eliminated. Of course, 
when the area of interest is highly urban and nearly no terrain 
points are available in the DSM this technique might not deliver 
the desired result. In such a case it is necessary to manually 
digitise points on the ground and interpolate or approximate a 
DTM-surface through them. 
The next and maybe most critical step is to find the regions of 
interest, i.c. the regions of potential candidates for buildings. 
Until today, either the buildings were digitised in stereo or the 
building corners were localized in the orthoimage or the user 
had to tell the system where a building is. 
Rottensteiner (2004) proposed to create a 'building mask’ using 
a certain elevation threshold. This mask would point out areas 
where objects higher than the threshold are located, and 
everything below that height is considered as 'non-building!. 
'Non-building' can be low vegetation, cars or other objects 
positioned on the terrain surface, but not being high enough to 
be treated as building candidates. Unfortunately, it is not 
seldom that vegetation, especially trees or even forests, have 
the same height as buildings do. 
Figure 2 shows such a typical case. One can easily recognize 
that the dense trees to the left are at least as high as the 
buildings. In such a case simple thresholding would not make it 
possible to distinguish between objects of interest (buildings) 
and vegetation. 
  
Figure 2. Image and DSM subset from the HRSC-Bern data set 
The algorithm proposed for finding regions with potential 
buildings is the following: 
Set a certain elevation threshold of minimal building- 
height according to the current scene, then 
- mark the potential candidate-regions of the nDSM in 
the image and 
— apply a check for homogeneity in the image to refine 
the candidate selection. 
This makes sense, since most of the times building-roofs are 
very homogeneous e.g. one colour, or one regular pattern of 
tiles, in comparison to treetops or other vegetation. To use the 
first derivative of the nDSM as a homogeneity parameter (a 
very popular method) is not recommended, because the derived 
nDSMs from high-resolution line scanning systems are much 
too coarse. 
4.3 Building Extraction 
Once the regions of building candidates are found, one has to 
extract the shape of the building. The coarseness of the nDSMs 
does not allow the determination of the roof shape. Even more, 
it is not even possible to get the building corners from these 
nDSMs. Therefore additional information is taken from the 
orthophotos. Since they provide higher geometric quality, they 
will be used to obtain the location of corners and hence the 
shape of the buildings. 
For the segmentation in intensity images four approaches are 
common, namely threshold techniques, boundary-based 
methods, region-based methods, and hybrid techniques which 
combine the latter two (Adams and Bischof, 1994). 
The proposed algorithm for building extraction is an ‘adaptive’ 
region-based method. The growing process of a certain region 
depends on the assigned threshold. If the difference between the 
value of the starting point (seed) and the value of examined 
neighbouring, adjacent pixels is equal or smaller than the given 
threshold, the neighbouring pixels will be aggregated into the 
region. The whole procedure is described in the following 
steps: 
I. Inside the potential building area called seed region 
(Figure 3a) start with the region growing process and 
1035