International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
131
interpretation can be increased by using the information from
preceding images. Hence, it becomes possible to distinguish for
example between the construction and the dismantling of
buildings or between the regeneration and degeneration of
moorland areas [Pakzad, 1999]. To realize such a multi temporal
analysis, an interpretation system must be able to administrate
images from different time instances and to represent and exploit
information about possible or at least probable temporal changes.
The leading idea of this work is to automate the evaluation of
aerial images of complex scenes using prior knowledge about the
object structure, GIS, sensor type, and temporal changes. To ease
the adaptation of the analysis system to new requirements and the
extension to future tasks, the knowledge is represented explicitly
and is separated from system control. Such a so-called
knowledge based approach constitutes the focal point of this
work.
In the literature various approaches to image interpretation and
sensor fusion have been presented. Only a few authors try to
formalize the representation of the objects and sensors, and the
control of the information integration. Most interpretation
systems like SPAM (McKeown, 1985) and SIGMA (Matsuyama,
1990) use a hierarchic control and construct the objects
incrementally using multiple levels of detail. The system
MESSIE (Clement, 1993) models the objects explicitly
distinguishing four views: geometry, radiometry, spatial context,
and functionality. It employs frames and production rules. In the
BPI system (Stilla, 1997) a net of production rules representing a
part-of-hierarchy describes the structural prior knowledge. A
blackboard realized by an associative memory is used for process
communication. Another blackboard-based architecture is
suggested by Mees (1998). He distinguishes between strategy
knowledge represented by an AND/OR-tree, global knowledge
described by sensor-independent fuzzy production rules, and
sensor-dependent local knowledge stored in attributed
prototypes and image processing operators called local detectors.
ERNEST (Kummert, 1993) uses semantic nets to exploit the
object structure for interpretation. The MOSES system extends
the ERNEST approach to extract man-made objects from aerial
images (Quint, 1997). The presented system AIDA (Liedtke,
1997) adopts the idea to formulate prior knowledge about the
scene objects with semantic nets. In addition, the control
knowledge is represented explicitly by rules which are selected
by an inference engine.
In the following, the system architecture of AIDA is described
and a common concept is presented to distinguish between the
semantics of objects and their visual appearance in the different
sensors considering the physical principle of the sensor and the
material and surface properties of the objects. The necessary
extensions to provide a multitemporal image analysis are
described and illustrated in chapter 5.
2. KNOWLEDGE BASED INTERPRETATION SYSTEM
For the automatic interpretation of remote sensing images, the
knowledge based system AIDA (Liedtke, 1997; Tonjes, 1999)
has been developed. The prior knowledge about the objects to be
extracted is represented explicitly in a knowledge base.
Additional domain specific knowledge like GIS data can be used
to strengthen the interpretation process. From the prior
knowledge, hypotheses about the appearance of the scene objects
are generated which are verified in the sensor data. An image
processing module extracts features that meet the constraints
given by the expectations. It returns the found primitives - like
line segments - to the interpretation module which assigns a
semantic meaning to them, e.g. road or river. The system finally
generates a symbolic description of the observed scene. In the
following, the knowledge representation and the control scheme
of AIDA is described.
2.1. Knowledge Representation
The knowledge representation is based on semantic nets.
Semantic nets are directed acyclic graphs and they consist of
nodes and edges in between. The nodes represent the objects
expected in the scene, while the edges or links of the semantic net
form the relations between these objects. Attributes define the
properties 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 object 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 built out of concepts. During
interpretation a symbolic scene description is generated
consisting of instances.
The object properties are described by attributes attached to the
nodes. They have a value measured in the data and a range
describing the expected attribute value. During instantiation the
attribute range of the instance is taken from the corresponding
concept and - if possible - is restricted further by the information
of instantiated parent nodes. For example, an already detected
street segment can constrain the position of the adjacent segment.
For both attribute value and attribute range a computation
method can be defined. A judgement function computes the
compatibility of the measured value with the expected range.
The relations between the objects are described by edges or links
forming the semantic net. The specialization of objects is
described by the is-a relation introducing the concept of
inheritance. Along the is-a link, all attributes, edges and
functions are inherited to the more special node which can be
overwritten locally. Objects are composed of parts represented
by the part-of link. Thus, the detection of an object can be
reduced 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 to structure the knowledge in different
conceptual layers like for example a scene layer and a sensor
