In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
Lwjv 
0 20 40 60 80 100 
Slant-Range (pixel] 
e 
Figure 2. Appearance of flat-roofed buildings in optical data (a), in SAR magnitude data with illumination from right to left (b,c) 
and InSAR phase data (d,e) 
the sensor depending on roof structure and illumination 
geometry. Ground behind the building is partly occluded by the 
building shadow leading to a dark region in the image. 
A building also leads to specific patterns in the interferometric 
phase data (Fig. 2d and Fig. 2e) because the phase value of a 
single range cell results from a mixture of the backscatter of 
different contributors, such as ground, façade, and roof in the 
layover area. Again, the appearance is characterized by a 
layover region and a homogeneous roof region (in Fig. 2 not 
observable because of the narrow building width). The phase of 
the terrain enclosing the building is displayed slightly darker. A 
similar phase value is calculated at the building comer location, 
which is used for the detection of building footprints. Since no 
signal is received in shadow area, the related InSAR phase 
carries no useful signal but noise only. 
3.2 Feature Extraction 
This approach of building recognition in InSAR data is based 
on the detection of parts of the building footprint. First, the 
segmentation of bright lines is carried out in the magnitude 
data. Based on this set of lines, only the ones caused by a 
dihedral corner reflector spanned by ground and building wall 
are used as building hints. In order to exclude all lines that do 
not fulfil this criterion, the local InSAR heights are analysed. 
Finally, the filtered corner lines are projected into the same 
ground range geometry as the optical data. 
3.2.1 Comer Line Segmentation 
As previously discussed, the bright corner lines are very useful 
hints to buildings since they provide information about the true 
location of a part of the building footprint. The full process of 
corner line detection is shown in Fig. 3, upper row. 
The line detection is carried out in slant range geometry based 
on the original magnitude images (Fig. 3 “Magnitude”) by using 
an adapted ratio line detector according to Tupin et al. (1998). 
This template detector determines the probability of a pixel of 
belonging to a line. In our case, eight different template 
orientations are considered. The probability image for the 
vertical template orientation is shown in Fig. 3 “Line”. 
Thereafter, line segments are assembled based on the eight 
probability images and their respective window orientation. The 
resulting segments are fitted to straight lines and edges, 
respectively, by linear approximation and subsequent 
prolongation (yellow lines in Fig. 3). 
3.2.2 Geocoding of Building Features 
After line extraction, the interferometric heights are calculated 
as described in (Thiele et al., 2007). Results are shown in 
Fig. 3 “Heights”. Local InSAR heights are investigated in order 
to discriminate lines caused by direct reflection and lines due to 
double-bounce reflection between either ground and wall or 
roof and substructures. For this filter step, the height difference 
between Digital Surface Model (DSM) and Digital Terrain 
Model (DTM) is used. 
The DSM is given by the calculated InSAR heights. In order to 
derive the DTM from it, a filter mask is computed to define the 
DSM pixels which are considered in the DTM generation. Only 
pixels with a high coherence value (Fig. 3 “Coherence”) and an 
InSAR height close to the global mean terrain height are 
considered in equation 1 (Fig. 3 “Mask”). 
1 - if x coh,i ^ 0-5 and (x h i -x h )< ±<j- (]) 
0, else 
Based on this mask and the InSAR heights, a DTM height value 
is calculated over an area of 50 m x 50 m in ground range 
geometry (Fig. 3 “DTM”). Thereafter, the height differences 
(i.e., a normalized DSM) between DSM and DTM are 
calculated (Fig. 3 “Height difference”). 
In the following line filtering step, lines are considered as real 
building corner lines if their neighbouring pixels show a low 
mean height difference value (Fig. 3 “Height difference”, 
rescaled for visualization). The filtered real corner lines are 
displayed in Fig. 3 (red lines). Final geo-coding of these comer 
lines is carried out using the InSAR heights. The resulting 
geographic position of the corner lines superimposed onto the 
optical image is displayed for the entire test site in Fig. 5b. 
4. FUSION OF EXTRACTION OUTCOMES 
In order to accurately combine features from InSAR data and 
the optical image, different sensor geometries and projections 
have to be considered carefully. It is required that both feature 
sets are projected to the same geometry, i.e., all data have to be 
transformed to a common coordinate system (Thiele et al., 
2006). In addition, a fusion and classification framework for 
combining the detection outcomes from the optical image and 
from the InSAR data has to be set up.