bul 2004
tr Verlag,
TOM]
al Images
hives of
. pp. 462-
)3. Graph
y. In: Shi,
ernational
raction by
ngineering
. External
'rammetrie
is for GIS
urban road
ournal of
base using
metry and
ds for road
ind Pattern
1ternational
32. Part 3-
e matching
: P (Eds),
and Space
y contextual
O. (Eds.),
and Space
Automatic
2P imagery.
nsing, Vol.
jc by image
metry, ETH
automated
nd geodata.
8(3-4), 166-
analysis for
Archives of
rt B3/2, pp:
raphic road
rchives of
3A, pp. 400-
TRANSFERABILITY OF KNOWLEDGE-BASED CLASSIFICATION RULES
K. Leukert, A. Darwish. W. Reinhardt
GIS lab (AGIS), University of the Bundeswehr Munich. 85577 Neubiberg, Germany
kristin.leukert@unibw-muenchen.de
Theme Session 11: Automatic Image Interpretation in the GIS Environment
KEY WORDS: Remote Sensing, Land Cover, GIS, Segmentation, Classification, Knowlede
ABSTRACT:
ce Base
e
Because of the rising complexity of high resolution imagery, object-based classification methods are becoming more suitable for
land cover classifications and acquisition of GIS data than traditional pixel-based classification methods. Object-based classification
requires a preliminary segmentation of the image and the application of knowledge-based rules to classify the image into the desired
output classes. Currently the rules for classification are defined again for each image. This is quite time intensive and a major
obstacle to the automation of the classification process. One goal o
f the research reported in this paper is to offer one set of
knowledge-based classification rules which is suitable for a specific geographic region and reuse/transfer it to similar images. This
paper presents an extensive transferability test of knowledge-based classification rules and results from a base rule set which should
be valid in several images. Radiometrically uncorrected IRS pan-sharpened images with a pixel size of 5m of eight test sites in
Northern Algeria are used. The results show that the transfer of rule bases is feasible and that it is possible to define a base rule set
for a specific geographic region for radiometrically uncorrected images. The achieved accuracies are lower than those of individual
classifications but the advantages are the reduction in analysis time and the possibility to semi-automate the data acquisition process.
I. INTRODUCTION
With the availability of high resolution image data, new
classification methods are necessary to handle the rising
complexity of the images as traditional pixel-based
classification methods produced unacceptable results. One of
those new methods is object-based classification where the
processing units are no longer single pixels but image objects.
As a first step the complete image has to be segmented into
meaningful areas/segments. The next step is to define a set of
knowledge-based classification rules to describe each target
class. Those rules can contain spectral, spatial, contextual, and
textural information. The final step is to assign each segment to
the class that fulfils the rules best. The software eCognition is
used for this research.
One of the main problems associated with such a classification
technique is the necessity to develop a new set of rules for each
new image, which is time consuming and is an obstacle to
automation. Therefore, the main objective of this paper is to test
the prospects of transferring classification rules to different test
sites. Transferability means the reuse of segmentation
parameters and knowledge-based classification rules that were
defined for one image in other images. This research focuses on
the transferability of classification rules.
Few attempts (e.g. Esch et al., 2003; Kressler et al., 2003, Mitri
et al., 2002) have been made in that direction. It is reported, that
classification rule bases were successfully transferred to other
images but that adaptation of some rules was necessary. But it is
not clear what kind of rules are used in the rule base and which
properties are stable in different images. Stable means that an
image object property characterises a f(eature class in several
images ideally with the same value range. Quantitative
evaluation is also missing in most publications.
In the publication of de Kok and Wever (2002) it is stated that
texture is a stable feature and suited to extract built-up areas in
very large data sets.
This paper starts with a short overview of the object-based
classification concept and a relatively new segmentation
algorithm (Multiresolution Segmentation). After specification
of the test sites, the practical investigations about transferability
of rule sets are presented and results are discussed. At the end
of this paper a summary and a conclusion including an outlook
are given.
2. OBJECT-BASED CLASSIFICATION
Object-based classification starts by segmenting the image into
meaningful objects. The resulting image objects “know” their
neighbours and they are subject to the succeeding classification.
The classification process is controlled by a knowledge-base
that describes the properties of the desired feature classes as
fuzzy membership functions. Object-based classification allows
the user to take not only spectral properties into account during
classification, but also shape, texture, and context information.
Arbitrary data like existing GIS layers or digital surface models
(DSM) can easily be integrated and used as supportive a priori
information in the classification process. For example, Hofmann
(2001) used high-resolution IKONOS data in combination with
additional elevation information to detect buildings and roads.
Object-based classification is also suitable for radar images. For
example, Corr et al. (2003) used object-based classification for
the production of urban mapping data from interferometric
polarimetric synthetic aperture radar (SAR) data.
2.1 Multiresolution Segmentation
The object-based image analysis software eCognition offers a
relatively new segmentation technique called Multiresolution
1059
