In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C„ Tournaire O. (Eds). 1APRS. Vol. XXXV1I1. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
3.3 InSAR features 
abed 
Figure 2. (a) flat-roofed building signature in magnitude data of 
InSAR pair (range from left to right), extracted double-bounce 
lines overlaid to (b), the coherence image, (c) the 
interferometric heights, (d) the magnitude image. 
A post-processing step is accomplished in order to reduce the 
number of mismatches, which occur due to multiple matches of 
individual lines. Finally, the stereo line segments are 
reconstructed exploiting the intersection of the stereo image 
rays. The stereo lines overlaid to a small part of the orthophoto 
are shown in Fig. 3(b) and a sketch showing the mapping 
geometries is given in Fig. 3(a). It can be seen that most of the 
stereo lines are located along the boundaries of the roofs 
particularly in case of flat roofs (see building A). In case of 
gable roofs some parts of the roof ridges are also extracted (see 
building group B). 
In order to derive meaningfi.il and consistent features we 
normalize the heights of the stereo lines. First, the local ground 
height is determined for each training and test image (of 310 m 
x 310 m size) assuming locally flat terrain. This assumption can 
readily be made because the test area is relatively flat. Second, 
the individual ground height of each image was subtracted from 
the heights of the stereo lines. Then, based on the assumption 
that the minimum building height is three meters, all stereo lines 
below this threshold are discarded. Then, we simply check if an 
image patch intersects with a line. In case it does the patch 
value is set to one and all other patches are set to zero (Fig. 
4(c)). We compute this feature in all three scales. 
Buildings in InSAR data appear differently compared to optical 
data due to the active illumination, the different wavelength, the 
side-looking viewing geometry, and the distance measurement. 
Furthermore, relevant building features occur in both magnitude 
and in phase data. An example is given in Fig. 2. It shows a 
typical magnitude signature of a flat-roofed building in (a) 
dominated by layover, double-bounce scattering, and shadow. A 
more in detail explanation considering different building types 
and illumination directions is provided in Thiele et al. (2010). 
Focusing on the coherence (b) and interferometric height data 
(c), especially the double-bounce line shows characteristic 
distributions. The high coherence value indicates high signal-to- 
noise-ratio in the InSAR data of this region. Furthermore, the 
interferometric height distribution at this line enables to 
discriminate between building lines and bright lines due to 
other effects. This double-bounce line is part of the building 
footprint, which is shown in Fig. 3(a). All these attributes make 
the double-bounce lines the most reliable building feature in 
urban areas and thus we extract features based on them. 
First, those double-bounce lines are extracted as proposed by 
Wegner et al. (2009) based on the magnitude image, the 
coherence, and the InSAR heights in slant range. Those lines 
(given in (b), (c), and (d)) are projected from slant to ground 
b c 
Figure 3. (a) Geometries of orthophoto, optical stereo images, 
and InSAR, (b) Buildings in orthophoto with flat roofs (A) and 
gable roofs (B) overlaid with 3D stereo lines, (c) same region as 
(b) overlaid with InSAR double-bounce lines 
4. CONDITIONAL RANDOM FIELDS 
High-resolution optical and InSAR data provide detailed infor 
mation of urban area objects (see Fig. 2(a) and 2(e)). Single 
trees, gardens, and streets are mapped. Those objects, their typi 
cal spatial distribution and interrelations with buildings can be 
exploited in order to improve classification through context 
integration. 
Conditional Random Fields, similar to Markov Random Fields 
(MRF), provide the possibility to integrate this context know- 
ground geometry Pi] Jj? 
InSAR data level 
lyT 
7 
double-bounce line 
stereo image level 
L/ - 
stereo lines 
orthophoto level 
.-jtmi 
roof segement 
p r p 2 
P 3 *P4 
projection using the local mean interferometric height at the line 
position. A schematic comparison of the extracted building 
hints of orthophoto, stereo images, and InSAR data is given in 
Fig. 3(a). In Fig. 3(c) the double-bounce lines of a flat-roofed 
(A) and a gable-roofed (B) building are superimposed to a small 
part of the orthophoto. 
Again, double-bounce lines may not be introduced in vector 
format directly since we deal with image patches. Thus, we 
apply a segmentation to the orthophoto and overlay segments 
and double-bounce lines. All intersecting segments are set to 
one, all others to zero. Finally, a distance map is generated and 
minimum and maximum values within each patch are 
computed. This feature is only generated for the highest 
resolution (i.e., the smallest patch size) (Fig. 4(d)). 
optical sensor 
(central projection) 
SAR sensor 
(slant projection)