3JECTS
Section 2 are given
> images again. Ad-
with sub-pixel reso-
eps of processing are
lization process and
chickler 1992, Busch
)metimes wider than
on process that leads
el. We use a skele-
r linear features. It
akes use of their di-
the topology of the
' concept of evading
r stage of computer
re classified as ends,
n or cross, or simple
g process links line or
de and/or end pixels.
id or node pixel are
tor representation of
and ends now since
linear feature extrac-
fail at nodes because
| to the line or edge
features that have to
nd ends we have real
mage which typically
s, we have to analyse
on process there.
| looking for end pix-
ct is to find items for
s. The number and
ires are helpful crite-
nportance of a node.
ur linear features are
three linear features
d the direction of the
atures which are each
number of incoming
ondence of opposite
s from the direction
may be also pseudo-
res meet. They occur
rithm avoids thinning
lyse the pseudonodes
ning linear objects or
recognized using the
are able to find pairs
ject, i.e. bound one
to find parallel edges
les the geometric ac-
m this, since — due
— the line position is
Figure 2: Detail of a KWR 1000 scene showing two roads
with extracted lines and edges, 235 x 213 pixels, ground
resolution ~ 2m.
Figure 3: Resulting segmentation
affected by different grey levels on the left and right side of
the line. So we can use the edge positions instead which are
more exact. Criteria for evaluating correspondence are the
neighbourhood, the constancy of the line width, and the di-
rection of the linear features. After that we have a geometric
description of the line including its width. We use this for
the segmentation of the lines in an image, too. An example
based on data of the Russian KWR 1000 sensor is shown in
Figure 2 and Figure 3.
4. ROAD, RAILWAY, OR RIVER?
The method described so far is part of low and mid-level
computer vision since no knowledge about the real objects
depicted in the image has been incorporated. So it may be
used to find lines and edges in arbitrary digital images of any
resolution.
We want to apply the method to satellite images to extract
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
objects that are relevant to cartography, e.g. roads and rivers.
Although we know that there are limits due to the spatial
resolution of operational sensors like SPOT and TM, so that
only major roads and autobahns can be detected (McKeown
1994), we think that our example is instructive and shows
the capability of the methods. Extracting lines from a SPOT
or TM satellite imagery results in lots of objects that are not
interpreted, i.e. all that we know about them is that they fit
to our line model of Section 1.2. Besides the objects we are
interested in, there is a large variety of other ones, like open
strips in a wood, long and narrow fields, or long buildings.
In our example we want to use two kinds of knowledge for dis-
criminating objects: knowledge about the width of the objects
and knowledge about their spectral characteristics. Since the
result of the line extraction depends on the size of the image
window used for the least squares fit of the polynomial (2),
we are able to select lines of different width. So it is not pos-
sible to detect lines of large width with a small window, while
a large window is not sensitive to narrow lines due to smooth-
ing. For using spectral characteristics we take advantage of
the fact that the extracted lines are skeletonized as mentioned
in Section 3. Hence, we have a representation of their middle
axis containing only few mixed pixels which constitute the
crucial point in multispectral classification. Therefore, the
detected lines are a good starting point for an object-related
multispectral classification. In our example the knowledge
comes from training areas that have been marked interac-
tively by an operator and that consist of detected line pixels
only. But it is possible to represent the knowlegde about the
spectral characteristics of roads and rivers in a knowledge
base. Additionally, unsupervised classifiers (e.g. Schulz and
Wende 1994) allow further improvement and automization.
The example is based on SPOT XS data (Figure 4). For
extracting the river Main that is flowing from the upper left
corner to the right side of the image we have applied the line
extraction technique described in Section 1.2 to band 1. We
have used a window size of 15 x 15 which is suitable for the
width of the river that varies from 6 to 11 pixels. The signif-
icance level for the robust estimation method of Section 2.2
has been set to 1096. Figure 5 shows the result. The small
part of the river Rhine in the lower left corner of the image
has not been detected because of if its width of more than
25 pixels. This demonstrates that the line model (2) allows
to dinstinguish lines of different width. For all pixels depicted
in Figure 5 we have gathered the spectral information from
the three bands so that multispectral classification has been
applied to these pixels only. The result of the classification
(Figure 7, bold line) illustrates that it has been possible to
select the river from the other linear features.
To find roads we have analysed the SPOT XS image (Fig-
ure 4) setting the window size and the significance level to
5 x 5 and 1096, respectively. The three bands have been
processed independently. Figure 6 combines the results by a
logical "OR" operator and shows all pixels that have been
recognized as line pixels in any of the three bands. This pro-
cedure is different to the one used for the river because the
width of the roads is close to the spatial resolution of the
SPOT XS sensor. Thus, we have needed information from
the three bands, whereas in case of the river it has been suf-
ficient to analyse one band. In Figure 7 we see the result of
the classification together with the extracted river. It demon-
strates that it has been possible to recognize the autobahns
