In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
173
a bed
Figure 4. Comparison of SAR and optical viewing geometry under the assumption of locally flat terrain (a); optical data (b) overlaid
with cadastral building footprint; optical data (e) and LIDAR data (d) overlaid with detected building comer
from SAR data is set to one third. Such weights are determined
empirically and lead to good results. However, further research
has to be done in order to support this choice with reasonable
statistics.
A quality measure is assigned to each region and initially set to
1. In the first evaluation part each primitive is evaluated based
on the optical feature vector. Each time a feature does not
completely support the building hypothesis, the quality measure
is reduced by multiplication with a value between 0 and 1. The
exact reduction value for each feature was learned on manually
labelled training data. Such reduced quality measure is again
multiplied with another reduction value if another feature partly
rejects a building hypothesis. The final quality measure based
on the optical feature vector is weighted with 0.666.
A second region evaluation is conducted based on the comer
line primitives extracted from InSAR data. First, all building
object hypotheses are enlarged by two subsequent dilation
operations. In this manner, a two-pixel wide buffer,
corresponding to 0.6 meters in ground geometry, is added to the
original region since building roofs may be shifted away from
the corner line. Thereafter, it is checked if the corner line
crosses this enlarged region with a certain minimum length. The
initial quality measure is multiplied with a reduction value like
in the optical case if this is not the case. The resulting quality
measure based on the corner line is multiplied with a weighting
factor of 0.333.
Finally, the overall quality measure is obtained by summing up
the optical and the InSAR quality measures. In case neither an
optical feature nor an InSAR feature has decreased the quality
measure, both quality measures sum up to one. All regions that
have a quality measure greater than an empirically determined
threshold are classified as building objects. Such threshold was
set to 0.6. As a consequence, a region may be classified as
building region even if there is no hint from the InSAR data, but
strong evidence from the photo. The reason is that some
buildings do not show corner lines due to an unfavourable
orientation towards the SAR-sensor (see the gabled-roofed
buildings in the lower right corner of Fig. 5b) or occlusion of
the potential corner line region by plants. On the contrary, a
region cannot be evaluated as building region based merely on
the corner line which are strong hints for buildings but may also
caused by other abrupt height changes in urban areas.
an effective baseline of approximately 2.4 m. The mapped
residential area in the city of Dorsten in Germany is
characterized by a mixture of flat-roofed and gable-roofed
buildings and low terrain undulation.
Results of the presented approach for building recognition by
means of feature combination from optical imagery and InSAR
data are shown in Fig. 5. In Fig. 5a building recognition results
based solely on optical features are displayed. All parameters
where specifically adjusted in order to achieve the lowest
possible false alarm rate while still detecting buildings. Less
than 50% of the buildings contained in the displayed scene are
detected. In addition, false alarms could not be avoided
completely. Results are rather poor due to the assumption that
roofs do not split up into more than two regions during the
region growing step, which is not met for the data at hand. As a
consequence, several gable-roofed buildings with reddish roofs
in the lower right corner of the image could not be recognized.
Some big flat-roofed buildings in the upper part of the image
are not detected because their colour and shape are similar to
such of street segments. Thus, their evaluation value does not
exceed the threshold.
Fig. 5b shows the corner lines extracted from the InSAR data
superimposed onto one SAR magnitude image. An InSAR
corner line could be detected for almost all buildings in this
scene. Some lines are split into two parts because the
corresponding building was partly occluded by, e.g., plants.
Some corner lines in the lower right diagonally cross buildings
which is not plausible. Most likely this effect is an artefact
introduced by too large tolerances applied in the merging and
prolongation steps of adjacent line segments. The final building
recognition result using both optical and InSAR features is
shown in Fig. 5c. The overall building recognition rate could be
significantly improved to approximately 80% by integration of
the InSAR corner lines into the classification procedure.
Additionally, all false alarms could be suppressed. However, the
gable-roofed buildings in the lower right comer stay undetected
although InSAR corner lines are present. Such missed
detections are due to the over-segmentation of the rather
inhomogeneous roof regions in the optical image.
6. CONCLUSION AND OUTLOOK
5. RESULTS
The InSAR data used in this project was recorded by the AeS-1
sensor of Intermap Technologies. The spatial resolution in
range is about 38 cm while 16 cm resolution is achieved in
azimuth direction. The two X-Band sensors were operated with
In this work, first building detection results from combined
optical and InSAR data on feature level were presented. A
rather simple approach for feature fusion was introduced
leading to a significantly improved building recognition rate.
Additionally, the number of false alarms could be reduced
considerably by the joint use of optical and InSAR features.
Corner lines from InSAR data proved to be essential hints for
