der invest-
mlets and
inploughed
IS)
ling to the
| liklihood
assification
the classes
distributed
settlements
etation etc.
distributed
class with
ally unique
noice as in
are rather
unation. of
any cases
s.
re slightly
nage pixel:
y (PI) and
y (P2). The
this two
able is the
A slight
ead to the
ignificance
ility of the
y of P1 and
the PI and
pplied. We
* could be
Ids* and
" uncertain
r mixed up
above the
i| fields are
field areas
nly ones of
sibility to
n the one
an be used
ultispectral
ass, that is
> processed
to its high
earance. In
our project the panchromatic images have also been used
for the geometrically exact determination of the forest
boundaries.
3.1 Texture Analysis
Classification of settlements by texture analysis has
already been used in the past (e.g. Steinnocher, 1997).
One of the standard algorithms of texture analysis has
been published by Haralick (1973). In our project we
apply the Forstner operator (Fórstner, 1991), a well-
known feature extraction algorithm in digital photogram-
metric applications. The operator delivers for each pixel a
measure that indicates firstly whether a directional or
non-directional features has been detected (q-values), and
secondly whether the feature is more or less prominent
(w-values). By appropriate threshold settings for q- and
w-values one is able to distinguish between point features
such as corners of buildings, individual houses and fields)
and linear features, i.e. salient edges (such as roads,
tracks, river banks, field boundaries). Urban areas are
densely mixed up by point features and linear features. In
open areas linear features like field boundaries can
clearly be separated from point features like field corners.
These peculiarities open the chance to separate
settlements from fields although both types of features
can be found in both classes.
The procedure commences with region growing for point
features with the condition that they are allowed to
expand over linear features but not into homogeneous
areas. In settlements this process causes a significant
growth of the points to patches, while in agricultural
regions the expansion is less important. The second step
Is to remove remaining edge pixels. Eventually a expand-
shrink operation is applied that causes clumping of areas
of densely distributed point patches. In that way the
textural class ,, Settlement has been found.
3.2 Threshold Analysis
Forested regions appear rather dark in panchromatic
images compared to the surrounding areas. Therefore, it
seemed likely that by thresholding or one-dimensional
classification forest could be detected or at least separated
form neighbouring classes. A global threshold does not
deliver expected results as forests do not form uniformly
grey areas. They may be dark at the one boundary and a
bit brighter at the opposite boundary. Therefore focal
thresholding was a more suitable approach. The process
begins by selecting appropriate training areas in order to
find suitable intervals of arithmetic means and standard
deviation that are typical for the forested areas. The focal
analysis checks within the focal window whether the
threshold boundaries are fulfilled thus obtaining the
thresholded class ,,Forest*.
We should bear in mind, that this procedure is still a one-
band classification process and we must not expect that
the segmentation delivers the correct class assignment in
all cases. One knows for instance, that also water bodies
appear very dark in panchromatic images. In fact, the
above explained threshold algorithm will also classify
water bodies as forest. This does not cause any real
problem as water can be clearly and reliably separated by
the multispectral classification algorithm. As we shall see
later, none of the classification steps is used alone for the
final class decision and therefore a contradictory classi-
fication result will cause either the class assignment by
the method that is most reliable for the respective class or
most likely a class assignment after involvement of
several of the input data sets or the class may be marked
as not reliable with a ,,to be checked" attribute.
4 COMBINING ALL DATA IN A GIS
The following step is the connection of all data in a geo-
information system (GIS) that allows the application of a
great variety of decision rules. Input to the GIS (see Tab.
I) are the results of the multispectral classification, the
result of the texture analysis, the result of the focal
thresholding and several sets of the existing DLM: the
forest layer mask and the so-called situation layer mask.
Both masks are raster images and have been generated by
scanning the respective separations of the topographic
map OeK50 with a pixel size of 2.5 m x 2.5 m ground
resolution. There exists also a vector layer with the most
important elements of the transportion network, such as
freeways, high order roads, railway lines.
The following table (Tab.1) lists all the imported layers
that will be available for GIS analysis.
GIS Layer (abbreviation) Type / Resol.
Max.Likelih. - highest probability
(maxlike 1)
Raster / 25 m
Max.Likelihood - second highest
probability
(maxlike 2)
Raster / 25 m
Thresholded forest
(forest-pan)
Raster / 10 m
Texture analysed settlement Raster / 10 m
(texture-pan)
DLM forest
(forest-DLM)
DLM situation
(situ-DLM)
Raster / 2.5 m
Raster / 2.5 m
DLM transportation network Vector
(transp-DLM)
Tab. 1: Layers in GIS
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 275
