Height difference
Figure 3. Upper row: steps of building corner segmentation in slant range geometry with illumination direction from left to right;
lower row: steps of the InSAR height filtering and slant range to ground range projection of the building corner lines
The particularities of Synthetic Aperture Radar (SAR) and
optical cameras in terms of sensor principle and viewing
geometry result in very different properties of the observed
objects in the acquired imagery. In Fig. 4a an elevated object P
of height h above ground is imaged by both a SAR sensor and
an optical sensor (OPT). SAR is an active technique measuring
slant ranges to ground objects with a rather poor angular
resolution in elevation direction. Layover, foreshortening, and
shadowing effects consequently occur and complicate the
interpretation of urban scenes. Buildings therefore are displaced
towards the sensor. Point P in Fig. 4a is thus mapped to point
PS in the image. The degree of displacement depends on the
object height h and the off-nadir angle 8 t of the SAR-sensor.
By contrast, optical sensors are passive sensors acquiring
images with small off-nadir angles. No distances but angles to
ground objects are measured. Elevated objects like P in Fig. 4a
that are not located directly in nadir view of the sensor are
displaced away from the sensor. Instead of being mapped to P\
P is mapped to PO in the image. The degree of displacement
depends on the distance between a building and the sensor’s
nadir point as well as on a building’s height. The further away
an elevated object P is located from the nadir axis of the optical
sensor (increasing 6 2 ) and the higher it is, the more the building
roof is displaced. The higher P is, the further away P is located
from the optical nadir axis and the greater the off-nadir angle 9/
becomes, the longer the distance between PO and PS will get.
The optical data was ortho-rectified by means of a DTM in
order to reduce image distortions due to terrain undulations.
Building façades stay visible and roofs are displaced away from
the sensor nadir point since buildings are not included in the
DTM. Such displacement effect can be seen in Fig. 4b to 4d. In
Fig. 4b the building in the optical image is overlaid with its
cadastral boundaries. The building roof is displaced to the right
since the sensor nadir point is located on the left. The upper
right part of the building is more shifted to the right than the
lower left part because it is higher (see Fig. 4d for building
height). Fig. 4c shows the same cut-out overlaid with the corner
line extracted from the corresponding InSAR cut-out. Such
corner line represents the location where the building wall
meets the ground which can nicely be seen in Fig. 4d. Due to
the previously outlined perspective effect the building roof falls
to the right over the corner line. This effect is of high interest
and can be exploited for three-dimensional modelling of the
scene (Inglada and Giros, 2004, Wegner and Soergel, 2008)
because the distance between the corner line and the building
edge comprises height information.
4.2 Joint classification framework
A joint classification is carried out after having projected the
optical and the InSAR primitive objects to the same ground
geometry. In order to combine the building hints from optical
and InSAR data, a fusion step is required. One possibility is
data fusion in a Bayesian framework while another would be
Dempster-Shafer evidential theory (Klein, 2004). Both
approaches are usually requiring an object to be represented
identically in the different sensor outputs, i.e., exactly the same
region is found in both datasets but with slightly different
classification results. This requirement is not met in the case of
the combination of line features from InSAR data with roof
regions from optical imagery.
Hence, combined analysis is based on the linear regression
classifier already used for building extraction from optical data
in (Mueller and Zaum, 2005). All potential building objects
from the optical image are evaluated based on a set of optical
features described in section 2.2 and on the InSAR corner line
objects. The evaluation process is split up into two parts, an
optical part and an InSAR part. Optical primitive objects are
believed to contribute more information to building detection
and hence their weight is set to two thirds. InSAR data is
assumed to contribute less information to overall building
recognition and thus the weight of primitive objects derived
172