ONTEXT
re not complete,
gaps, due to the
is optimized for
d as a network.
es can interfere
of local context
as, other global
ion for the road
is Information a
scale modeling.
iled model and
ion and explicit
ures (e.g. crash
and incorporat-
ith a model for
ds are outlined
each extension
TEXT
gly depends on
in space. The
1 a resolution of
ok X-band data
)oSAR) and up
1e model below
erties of roads
ext information
« model for the
2-3).
s lines with the
seneous reflec-
al reflection)
onstant width,
tion constraints)
work
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
(a) Layover effects of trees on — (b) Blurred vehicle moving in azimuth
road section direction
(c) Junction (d) Bridge and traffic signs
Figure 1: Local context objects
In open rural areas the appearance of roads corresponds reason-
ably well to this model. Lines representing road centerlines can
be extracted in a stable manner. Whenever the complexity of the
scene increases, interactions between roads and other objects ap-
pear frequently, e.g., neighbored objects cause shadow/layover
regions, or moving vehicles cause bright peaks blurred in azimuth
direction. In such situations, the line model holds no longer and
an extraction system relying purely on such a model would be
rather weak. It is necessary to take context information into ac-
count.
2.2 Context information
The model for roads is now extended by context information.
Context, in general, is restricted to knowledge about relations of
the object of interest to other neighboring objects (Stat and Fis-
chler, 1995). For roads, two levels of context can be distinguished
like in (Baumgartner et al., 1999).
e Local context describes knowledge about relations between
roads and spatially neighbored objects (context objects).
e Global context is characterized by the presence of other ob-
jects in the larger region (context regions). In (Baumgartner
et al., 1999) it is used to define an appropriate extraction
technique .
Here, both context levels will be taken into account. For deter-
mining the relevent context objects and regions for roads in SAR
imagery, a detailed analysis was carried out. Therefore, the road
extraction of Sect. 3.1 was applied to indicate the difficulties.
The false alarms and gaps in the road network were examinated
carefully.
22.1 Local context objects Situations in which neighbored
objects make road extraction locally difficult are caused by the
following local context objects. They are illustrated in Figure 1.
e layover and shadow regions caused by buildings and alleys
e blurred bright stripes caused by vehicles moving in azimuth
direction
* road junctions
* bridges (indicated by the high backscattering of the metal
barriers at the sides of the bridge and corner reflector inter-
action with the ground)
* large traffic signs at highways
Many of the context objects are characterized by a high backscat-
tering caused by metallic structures or by multiple bounces: in-
dividual trees, larger traffic signs, bridges, blurred vehicles, etc.
These phenomena exist in all SAR data, in E-SAR data as well as
in AeS-1 data. In order to make it possible to use these objects as
positive evidence for roads during the road extraction, they have
to be explicitly modeled.
361
2.2.1 Contextregions The road extraction depend also on the
region where it is applied, i.e., on the global context. For instance,
roads in urban areas have a quite different appearance from roads
in forest areas or in rural areas. Precisely because the influence
and presence of neighbored local objects and the density of those
objects. For roads three global context regions are distinguished
here, rural, urban, and forest areas. In rural areas the mentioned
context objects are casual distributed, in contrast to forest and
urban areas where trees and buildings are more frequent. Expe-
rience shows that approaches that are suitable for road extraction
in rural areas usually cannot be applied in other global contexts
without modification. Often a completely different approach is
required. Therefore, forest and urban areas are initially mask out
for the extraction of rural roads (Figure 2).
But urban areas are a high indicator for roads: Roads are orga-
nized as a network connecting all areas inhabited and exploited
by human beings. This semantical function of roads can be used,
because it comprehends a reliable seed information for the road
extraction.
(a) SAR image, L-band
(b) Rural area
Figure 2: Global context
2.3 Highways
In contrast to rural roads, highways are mostly wider and for the
side looking radar less occluded by context objects. A separate,
detailed model for highway is set up in the following. Highways
comprise two (anti-)parallel roads that are bordered by crash bar-
riers. To extract such types of objects, the use of a multi-scale
model has proven to be very important (see e.g. (Hinz and Baum-
gartner, 2003)). In the highest resolution, a highway is character-
ized by two parallel dark lines separated by a thin bright line, the
central crash barrier. This crash barrier is a reliable feature, be-
cause metal objects produce a strong radar backscattering. Fig-
ure 3(a) illustrates this and some more typical effects depend-
ing on the viewing angle: (A) Crash barriers orientated approx-
imately in azimuth direction appear very bright, because of the
corner reflector effect. (B) In other orientations, the high reflec-
tion of the crash barrier is still present. (C) There are also some
areas without any reflection. These are caused by radar shad-
ows in case of very high objects nearby. (D) The corresponding
