33. Istanbul 2004
ition, about 6m
Isewhere. There-
also bridges, get
line. Other ob-
ple, alleys appear
forth represented
signs only appear
tomatic) interpre-
are added to the
The same is true
[hey are modeled
ral terminals that
1 now, context re-
est areas from the
the computation
| segments, which
1s. For this task a
extracted directly
full-polarimetric
orest areas, based
ds.
nformation in the
ontours of urban
same way as de-
reas to be a seed
g the contour line
traction attempts
ide. Second, the
network without
the contours are
ecause often, the
n strategy for high-
eses formation in
sh resolution, (3)
ration. (1) To cre-
ind wide lines are
re weighted with
idth, length, cur-
form for straight,
t lines according
1 the high resolu-
d, i.e., candidates
between. To get
nclosing a bright
ted according to
ion, according to
ines. (3) All hy-
rategy. Thereby,
highest weights.
h-based grouping
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
4 RESULTS
The potential of using context information is demonstrated by
two test sites: a rural-industrial test site near Munich, Germany
(AeS-1 data) and a 1 1 km x 7,5 km extended rural test site in South
Germany, Ehingen (E-SAR data), both X-band data. A quantita-
tive evaluation of the results according to the evaluation scheme
in (Wiedemann and Ebner, 2000) has been carried out with man-
ually plotted reference data. As summarized in Table | and Ta-
ble 2, by means of quality measures for completeness, correct-
ness, and geometrical accuracy the extraction results are evalu-
ated with and without introducing context information. The eval-
uation shows that the results are relatively complete, especially
for main roads (> 93 96). In both test sites the completeness and
correctness could be improved by using context information.
(c)
2
(d)
Figure 4: Extraction results (a) SAR image (b) Extraction result
Without context (c) Extraction result with local context objects (d)
Introduced local context objects: vehicle (line), trees and junc-
tions (linking point)
| | AeS-1 | AeS-1 with context |
Completeness 66.7 9/o 69.7 %
highways 63.8% 64.1 %
main roads 94.9 % 95.5%
secondary roads | 64.9% 71.3%
Correctness 71.8% 72.8 %
RMS 2.2m 2.2m
Table 1: Evaluation of extraction results for Munich
without with with local and
local global context
Completeness 84.6% | 85.3% 88.5%
main roads 93.9% | 96.9% 97.2%
secondary roads | 81.3% | 81.2% 85.4%
Correctness 73.6% | 73.7% 73.6%
RMS 2.2m 2.1m 21m
Table 2: Evaluation of extraction results for Ehingen
4.1 Results for context objects
In the Munich test site more secondary roads have been extracted
by the use of local context objects, especially due to the intro-
duction of bridges. For highways, only a small improvement is
reached. Obviously, introducing a traffic sign as a short line seg-
ment is not sufficient. To cope with disturbances caused by reflec-
tions at metallic structures, a feedback loop to the SAR process-
ing would be necessary, e.g., a specialized technique to suppress
the side lopes.
In the case of the Ehingen test site some gaps in the road network
could be closed by the aid of individual trees as potential road
segments. As depicted in Figure 4 one small gap in the upper left
part could be closed. The central gap isn’t totally caused by local
context objects. Apart of layover of a large tree and a moving
vehicle, the low contrast in between seems responsible for the
missing extraction. Note, that only local gaps are supposed to be
closed by local context information.
Another aspect of modeling context is the higher robustness of the
extraction. Influences of non-modeled objects are usually tried to
be overcome by relaxing some of the parameters involved (e.g.,
parameters for grouping lines). On one side, this may lead to
a more complete result; on the other side, the result is typically
less correct since relaxed parameters cause more misdetections.
However, when adding information about context objects during
the extraction, the amount of gaps is usually less, so that the pa-
rameters can be set much more restrictives.
4.2 Results for context regions
By introducing the contour line of urban areas the algorithm can
start the network generation with this segments. Usually, roads
nearby urban areas are more influenced by adjacent buildings or
vegetation. That is the reason why these road segments are often
missing in a conventional extraction. In Figure 4.1 the improve-
ment by using global context is shown. Regarding the left village,
two more outgoing branches could be extracted. Furthermore, the
contour line might be used as basis to connect loose road parts
to a topological correct road network. (To avoid confusion, the
two lines running downwards from the village are just ways and
therefore not part of the reference data (Figure 5(b)).)
4.3 Results for highways
We applied the extraction system to some test sites which con-
tain highways. The results for a test site north of Karlsruhe are