ral 
ng 
ch 
he 
se 
or 
in 
or 
of 
OT 
1€ 
of 
1S 
n 
--— A) 
iv vy 
ISPRS Commission III, Vol.34, Part 3A Photogrammetric Computer Vision“, Graz, 2002 
To fulfil condition 3 feature analysis operators were created, 
which are described in section 4. 
The aim of this work was to detect changes over the time. 
Therefore another condition is the possibility to model and use 
possible temporal landscape changes in the system, which is 
described in section 5. 
2. KNOWLEDGE BASED SYSTEM 
The system used for the presented approach is the knowledge 
based system AIDA (Liedtke et al., 1997, Tonjes, 1999) which 
was developed for automatic interpretation of remote sensing 
images. The system strictly separates the control of the image 
analysis process from the semantics of the scene. The 
knowledge representation is based on semantic nets (Niemann 
et al., 1990). Semantic nets are directed acyclic graphs. They 
consist of nodes and edges in between the nodes. The nodes 
represent the objects expected in the scene while the edges or 
links of the semantic net model the relations between these 
objects. Attributes define the properties and methods of nodes 
and edges. 
The nodes of the semantic net model the objects of the scene 
and their representation in the image. Two classes of nodes are 
distinguished: the concepts are generic models of the objects 
and the instances are realizations of their corresponding 
concepts in the observed scene. Thus, the knowledge base, 
which is defined prior to the image analysis is composed of 
concepts. During the interpretation a symbolic scene description 
is generated consisting of instances. 
The relations between the objects are described by edges or 
links of the semantic net. The specialization of objects is 
described by the is-a relation introducing the property of 
inheritance. Along the is-a link the description of the parent 
concept is inherited to the more special node which can be 
locally overwritten. Objects are composed of parts represented 
by the part-of link. Thus, the detection of an object can be 
simplified to the detection of its parts. The transformation of an 
abstract description into its more concrete representation in the 
data is modelled by the concrete-of relation, abbreviated con-of. 
This relation allows for structuring the knowledge in different 
conceptual layers, for example a scene layer and an image 
layer. 
To make use of the knowledge represented in the semantic net 
control knowledge is required, which states how and in which 
order the image interpretation has to proceed. The control 
knowledge is represented explicitly by a set of rules. The rule 
for instantiation for example changes the state of an instance 
from hypothesis to complete instance, if all subnodes, which are 
defined as obligatory in the concept net have been completely 
instantiated. If an obligatory subnode could not be detected, the 
parent node becomes a missing instance. The control of 
interpretation is also performed by an A*-Algorithm. For 
further details see (Tônjes, 1999). 
3. EXTRACTION OF SEGMENT BORDERS 
As described above a separation between the extraction of 
segment borders and the interpretation of segments was carried 
out. Extraction of the segment borders was the first step for 
every epoch of the multitemporal interpretation, then the 
interpretation was done. The border extraction for the first 
epoch, in the following called initial segmentation, differs from 
the others, called resegmentation. 
Investigating the position of segment borders in manually 
created vegetation and biotope mappings shows that landscape 
objects were often used as borders. This applies also for 
industrially used moorland, which was investigated in this 
work. Suited as a border are streets and paths, waters and 
ditches. These objects can be extracted by existing methods (an 
overview of different approaches is given in (Fórstner, 1999, 
Mayer, 1998)) and used directly as borders. As researches in 
this area were not part of this work, existing Geo-Data were 
used directly as segment borders. Thus, in a first step the first 
segments were created. This step is only carried out for the first 
epoch. 
In a second step, which is used for all epochs, these segments 
were processed. In each segment parts are selected during 
processing. The selection criteria depend on the kind of input 
images: Greyscale or CIR(Colorinfrared)-Images. For CIR- 
Images all parts without vegetation (low NDVI-Value) are 
selected, and for greyscale images all parts without textures 
with parallel lines. Then the selected parts are processed by 
morphological operations. Finally it is decided if the segment 
splits into the processed parts or not. This depends on several 
criteria: 
e They must have a particular relative and absolute 
minimum size. 
e À part of the new segment border must match the old 
border. 
e The shape has to fulfil specific conditions. 
e Further criteria. 
Greyscale and colour-images are handled differently. The 
higher uncertainty of greyscale images (because of the missing 
colour information) leads to stricter criteria for the acceptance 
of new segments. 
The described resegmentation is carried out for every segment 
found during the initial segmentation. Resegmentation for the 
next epochs will only be performed considering particular 
conditions and only for segments, which were assigned to 
special classes. This means, that resegmentation is only 
necessary for particular classes. This reduces the error of 
resegmentation. 
4. RECOGNITION OF CHARACTERISTIC 
STRUCTURES 
The interpretation keys, as explained in section 1, describe 
features and structures, which have to be recognized. To 
implement this concept into an automatic interpretation system 
the interpretation keys must be implemented as image 
processing operators. Automatic recognition of features and 
structures is done by automatic feature analysis operators. 
The input of a feature analysis operator is a segment, which has 
to be examined for special features and structures. Also, 
additional parameters of the kind of the structures and of the 
resolution of the images are given to the operators. After 
verification the operator’s result is a value between 0 and 1, 
which describes the number of examined features in the given 
segment. 
Transformation of indistinct descriptions, which are given in 
standard language, as for example “dismembered structure”, 
into image processing operators is difficult. For example, the 
operator for “dismembered structure” has to fulfil the following 
conditions: 
1. The edges of the image structures have to be curved. 
2. Higher density of curved structures should lead to 
better results. 
3. The spatial distribution of the curved structures 
should be equal all over the segment. 
4.  Uncurved structures should not influence the result of 
the operator. 
A - 235