Topographic maps and aerial images are available covering
extendedregions of Switzerland. Since 1995,the Swiss Feder-
al Institute for Forest, Snow and Landscape Research (WSL)
has used a complete set of scanned topographic map sheets.
Maps with scales 1:25*000, 1:50*000, 1:100*000, 1:200'000
and 1:500°000 [L+T 1994] and with a maximum resolution of
508 dpi have been published by the Swiss Federal Office of
Topography (L+T).
Specifically-tuned methods for the automatic detection of
specific map symbols have been developed for the L+T maps,
in order to shorten the data collection process.
Figure lc: tree symbols
2. Shape descriptors
Instead of trying to resolve the map pattern recognition pro-
cess as a whole, the methods that are applied here are focused
on the detection of one specific symbol at a time, following
the ,,divide-and-conquer* principle. Due to this simplification
and also reasons of efficiency, the recognition process does
not rely on any operator interaction.
As individual map symbols are drawn by hand at L+T, their
geometric properties are non-rigid (Figure 2). In practice they
vary considerably in shape and quality and therefore cannot be
described with a single geometric shape measure.
Complex symbols such as forest boundaries, dry channels and
avalanche obstructions differ even more from symbol to sym-
bol (Figures 1a-1c), due to the local surface properties, and do
not have salient features such as key-points or an object cen-
tre.
A suitable recognition method therefore needs to be sufficient-
ly robust and scale-, rotation- and ,distortion“-invariant. Al-
though distortion is an inexact terminology, it has been used
for any kind of non-linear distortion.
Shape descriptors such as area, perimeter, moments, elonga-
tion, eccentricity, skeletons, Euler number and Fourier de-
scriptors fulfil the requirements fairly well and have been cho-
sen as low-level feature descriptors. Many different shape de-
scriptors have been reported and applied in different research
areas [Ballard 1981, Marshall 1989, Pavlidis 1976, Rauber
66
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
1994, Reeves 1981, Zahn 1972].
Despite its wide use, template matching, either in the spatial
or frequency domain, has been shown to be an inappropriate
detection method [Ballard 1982, Stengele 1995]. Map sym-
bols are often distorted due to the limited resolution of the
scanning process and therefore the missing rotation and scale
invariance is a severe limitation. Tree symbols contain less
than 10-15 pixels and are highly sensitive to filtering by image
processing. Some experiments have been made with a
rotation-dependent set of templates but the computational re-
quirements are so large as to make the process impractical.
Deformable templates have been applied successfully using
weak models such as ,snakes“ [Henricsson 1994, Kass 1987,
Lee 1989], but gray-level data are needed to calculate poten-
tial surfaces.
Morphological operators such as erosion and dilatation are in-
adequate [Pitas 1992, Trahanias 1992] because of the limited
object dimension, and the resulting deviation from the ,origi-
nal“ object form makes even a weakened rule definition un-
predictable.
À À & À
a 400 006
Figure 2: distorted tree and triangulation point symbols
3. Discrimination
Shape descriptors enable classification according to the values
of the different descriptors. Contrary to classification methods
for remote sensing data, different shape features are often cor-
related among each other and build a non-orthogonal n-
dimensional feature space, which is not appropriate for the re-
quired statistical independence. Dimension reduction methods
such as the Karhunen-Loéve transformation are not applicable
due to the inconsistent feature measuring units.
Each symbol discriminator is defined by a set of rules based
on the basic shape descriptors. The definition process is the
only interactive part of the method, whereby the user iterative-
ly solves the complex discrimination problem.
[TT
Area 12
Eccentricity
Elongation
Spreadness
M20
|
Area 35
Eccentricity
Elongation
Spreadness
M20
ded
Area 47
Eccentricity
Elongation
Spreadness
M20
Figure 3: imay
Show (e
and
and
and
and
and
and
and
and
FPO-FP:
Absm!1 1
Table 1: 1
In an int
database
by defini
scriptors.
map syml
the syml
image Sy
character
fied shap
compose
values fo
In an int
database
by defini
scriptors
mapsym
the symb
