CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
170 
abed 
Figure 1. Flat-roofed (a) and gable-roofed (c) building in optical image overlaid with corresponding regions after segmentation (b,c) 
countries rooftops look usually grey, reddish or brownish but 
almost never green. Roof types can roughly be subdivided into 
flat roofs and gable-roofs. Flat roofs coincide often with rather 
homogeneous image regions (Fig. la) while gable-roofs 
sometimes appear less homogeneous. Chimneys and shadows 
cast by chimneys may further complicate roof extraction if 
homogeneous planes are fit to roofs (Fig. 1 c,d). Due to similar 
colour of adjacent roof and street regions, such entities are 
sometimes hard to be told apart even for human interpreters 
(Fig. 1 a,b). 
In this work the focus is on fusion of building primitive 
hypotheses delivered by approaches from the literature, tailored 
to the specific constraints that are determined by the 
particularities of the optical and microwave realm, respectively. 
With respect to the visible domain, a robust model-based roof 
detection approach introduced in Mueller and Zaum (2005), 
known to deliver good results, was used. It is based on an initial 
region growing step yielding homogeneous segments. As a 
consequence of the previously outlined diverse appearance of 
building roofs in optical imagery, such segmentation may 
sometimes lead to suboptimal results if contrast between roof 
regions and adjacent regions is very low (Fig. la). Thus, the 
region growing step can lead to erroneous roof segments (Fig. 
lb). Gable-roofs usually split up into at least two segments if 
they are not oriented along the sun illumination direction (Fig. 
1. c,d). Sometimes, gable-roofs may split up into even more 
than two segments and only parts are evaluated as roof regions. 
In such cases, the introduction of building hints from SAR data 
can highly improve building detection. 
2.2 Feature Extraction 
The building roof extraction approach consists of a low-level 
and a subsequent high-level image processing step (Mueller and 
Zaum, 2005). The low-level step includes transformation of the 
RGB image to HSI (Hue Saturation Intensity) representation, a 
segmentation of building hypotheses in the intensity image and 
the application of morphological operators in order to close 
small holes. Region growing, initialized with regularly 
distributed seed points on a grid, is used as image segmentation 
method. Seed points that fall into a grid cell which either 
consists of shadow or features a greenish hue value are erased 
and no region growing is conducted. Adjacent roof regions 
having a significant shadow region next to them are merged. 
This step is important for gable-roofed buildings because 
sometimes the roof is split at the roof ridge due to different 
illumination of the two roof parts. However, gable-roofs that 
were split up into more than two segments are not merged to 
one single segment which is the main reason for undetected 
buildings later-on in the process. 
Features are extracted for each roof hypothesis in order to 
prepare for classification. Four different feature types are used, 
based on geometry, shape, radiometry, and structure. Geometric 
features are the region size and its perimeter. The shape of a 
building region is described by its compactness and length. 
Right angles, distinguishing roofs from trees in the real world, 
are not used as a shape feature since the region growing step 
may lead to segments that are not rectangular although they 
represent roofs (Fig. lb). Radiometry is used in order to sort out 
regions with a high percentage of green pixels. Structural 
features are for example neighbouring building regions and 
shadows cast by the potential building. Shadows are good hints 
for elevated objects. In order to not take into account shadows 
cast by trees, only shadows with relatively straight borders are 
considered as belonging to buildings. 
Finally, a classification based on the previously determined 
feature vector takes place (see chapter 4.2 for details). All 
necessary evaluation intervals and thresholds were learned from 
manually classified training regions. 
3. ANALYSIS OF INSAR DATA 
3.1 Appearance of Buildings 
The appearance of buildings in InSAR data is characterized by 
the oblique illumination of the scene and therefore the image 
projection in slant range geometry. Furthermore, it depends on 
sensor parameters, on properties of the imaged object itself, and 
on the object’s direct environment. 
In Fig. 2 an example of flat-roofed buildings in optical (Fig. 2a) 
and InSAR data (Fig. 2 b,d) is given. The appearance of 
different building types and effects that occur if the scene is 
illuminated from two orthogonal flight directions have been 
comprehensively discussed in Thiele et al. (2007 and 2008). 
The magnitude profile of a building is typically a sequence of 
areas of various signal properties: layover, corner reflector 
between ground and building wall, roof, and finally radar 
shadow (Fig. 2c). The layover area is the building signal 
situated the closest to the sensor in the image because its 
distance is the shortest. It usually appears bright due to 
superposition of backscatter from ground, façade, and roof. The 
layover area ends at the bright so-called corner reflector line. 
This salient feature is caused by double-bounce reflection at a 
dihedral corner reflector spanned by ground and wall along the 
building. This line coincides with a part of the building 
footprint and can be distinguished from other lines of bright 
scattering using the InSAR phases (see Fig. 2d and profile in 
Fig. 2e). The single backscatter signal of the building roof is 
either included in the layover mixture or scattered away from