culture areas will produce wrong assignments. As fig. 5 (com-
pared to fig. 1) has shown, there are some confusions caused
by harvested or uncovered fields.
A second approach for texture parameter is based on the mo-
dified variance/standard deviation of the reflectance values
inside an object area. Using the common standard deviation
disturbances (e.g. by mixing with other classes, digitizing
errors etc.) will increase the standard deviation enormously.
Therefore, a histogramm analysis of the reflectance values
is applied, where the main maximum is extracted. Now the
related statistical distribution and its 'cleaned' standard de-
viation can be determined. The results for the same object
classes and training areas (as used for Haralick parameters)
are shown in fig. 10.
A better separability can be achieved (compared to the Har-
alick parameters) — especially between the object classes sett-
lement and agriculture. This may be caused by the mostly
quite inhomogenous texture in satellite images, where the
modified standard deviation offers a good robustness.
Markov random fields are just under investigation and deliver
(up to now) in the first trials no satisfying results applied to
satellite images.
3.3 Shape / Size Feature
Shape and size of objects can be used as non-spectral, geo-
metrical features for satellite image analysis. Based on the
results of segmentation (where object contours were achieved
directly in vector format) parameters for shape and size can
be derived.
One of the simplest shape parameters may be 'roundness'
defined as area to perimeter ratio. After reducing redundancy
in the contour lines, e.g. by Peuker method, other more
complex parameters like straightness or parallelism of edges
etc. can be determined.
It is very easy to calculate the size of an object (or a seg-
ment). Using, for instance, the Gauss algorithm the size can
be computed by means of the image- or absolute coordinates,
respectively.
It has to be pointed out again, that both feature contains
only fuzzy or uncertain information in the decision process.
3.4 Relation Feature
Beside the features of the objects itself also the interrelation
between the objects carry useful information for image analy-
sis. A part of these relations can be defined in rules, e.g. the
relation between buildings and streets, between settlements
and agricultural areas. Therefore the relations between all
adjacent objects/segments of our working area is modelled
explicitely in the decision structure (semantic network).
4 SEMANTIC MODELLING
After creation of segments (incl. their extented features) as
candidates for semantic objects, a systematic structuring of
the extracted knowledge is necessary. This information has
to be formalized synthetically. It is used as input for the
processing and analysis of image data.
4.1 Analysis System
The first step in semantic modelling of the image contents
will be the verification of the DLM200 objects in the satellite
image by means of the extracted candidates (segments). If
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
objects don’t have changed and therefore have (with a certain
probability) no significant difference, they can be verified. If
not, a general classification procedure is applied to these non-
verified objects. To do this methods on higher, symbolic level
are necessary. For this special type of data processing general
knowledge / rules and specific knowledge about the topo-
graphic objects is used. Therefore, the analysis system must
be able to represent and process this knowledge. Different
formalisms for knowledge representation are known like pred-
icate logic, rule-based systems, formal gammatics or semantic
networks (SN). SN belong to the most useable schemes con-
cerning knowledge representation. For using and modelling
of knowledge about the database (DLM200) and the actual
image — as well as central control unit — ERNEST (Erlanger
Semantisches Netzwerksystem) is applied in this work ((NIE-
MANN ET AL. 1990); (KUMMERT ET AL. 1993)).
A SN contains two different types of knowledge: declarative
and procedural knowledge. Declarative knowledge consists
of concepts and links, while procedural knowledge contains
methods for determination of attributes of concepts as well
as for valuation of concepts and relations.
The analysis process is controlled by a special algorithm,
which leads to a multi-level analysis concept (fig. 11).
generative
model scene area
image area
generic
model
map area
instantiation
map scene
description
pronounced
model
image scene
description
classification
change detection
update
Figure 11: Structure of image analysis
A basic SN contains general knowledge about the scene to
be analysed (generative model). In the next step two SN
are created containing specific knowledge about the database
and about the image objects, respectively (generic models
‘database’ and 'image’). The analysis is based on a compar-
ison of this two models. In this level procedural knowledge
(feature extraction) is added to the knowledge still present.
An actual description of the scene in the database domain
is built up by means of the generic database model. This
description of the 'database scene' will be transformed to the
756
T --Q4---———-———--------------2-.2-------------
aut
crez
ciali
desc
spec
will
tion
but
(dat
4.2
The
clas:
cess
of d
SN.
the
this
conc
(adc
the :
base
This
valu
for :
dete
well
and
acts
for 1
obje
Fig
Fig.
conn
