International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
134 
Figure 3. Sensor Fusion demonstrated on the aerial view of a purification plant: Rejected (thin lines) and accepted (wide lines) road features 
from (a) optical and (b) infrared image with (c) fusion result. 
integration and simultaneous interpretation of images from 
multiple sensors. New sensor types can be introduced by simply 
defining another specialization of the Sensor node with the 
corresponding geometrical and radiometrical transformations. 
According to the images to be interpreted, the different sensor 
layers (SAR, IR, optical) are activated. 
For the application of road extraction, the advantages of a 
multisensor image analysis are illustrated in Fig. 3. Using only 
the aerial image (a) or the infrared image (b) yields fragmented 
results. If both images are analyzed simultaneously, the gaps can 
be closed. In those areas where both images provide a hint for a 
road segment the reliability of the interpretation is increased. In 
other areas, the information from the images complement each 
other. Other examples for the fusion of multisensor images are 
given in (Tonjes, 1998). 
5. INTERPRETATION OF MULTITEMPORAL IMAGES 
Currently, the system is being extended for the interpretation of 
multitemporal images. Applications like change detection and 
environmental monitoring require the analysis of images from 
different acquisition times. By comparing the current image with 
the latest interpretation derived from the preceding image, land 
use changes and new constructions can be detected. In the 
following, the necessary extensions to a multitemporal analysis 
with the system AIDA are described. Preliminary results are 
shown for the extraction of an industrial fairground. 
5.1. Extension of the Knowledge Based System 
The easiest way to generate a prediction for the current image 
from an existing scene interpretation is to assume that nothing 
has changed during the elapsed time. The latest scene 
interpretation represented in an instantiated semantic net is 
therefore regarded as a kind of GIS and it is used to guide the 
analysis of the current image. The objects found in the last image 
are verified in the current one but changes are difficult to detect 
and to explain. However, in many cases humans have knowledge 
about possible or at least probable temporal changes. Hence, the 
knowledge about possible state transitions between two time 
steps should be exploited in order to increase the reliability of the 
scene interpretation. 
Temporal changes can be formulated in a so called state 
transition graph where the nodes represent the temporal states 
and the edges model the state transitions. To integrate the 
transition graph in a semantic net the states are represented by 
concept nodes which are connected by a new relation: the 
temporal relation (see Fig. 4). For each temporal relation a 
priority can be defined in order to sort the possible successor 
states by decreasing probability. As states can either be stable or 
transient, the corresponding state transitions differ in then- 
transition time which can be also specified in the temporal 
relation. As normally no knowledge about the temporal changes 
of geometrical objects or materials is available, the state 
transition diagram is part of the semantic layer (compare Fig. 2). 
In contrast to hierarchical relations like part-of or con-of, the 
start and end node of temporal relations may be identical - 
forming a loop - to represent that the state stays unchanged over 
time. Figure 4 shows a simple example of a state transition graph 
consisting of three states. For the different transitions priorities 
and transition times are defined. 
To exploit the temporal knowledge, a time stamp is attached to 
each instance of the semantic net which documents the time of its 
instantiation. Thus, it will be possible to filter time slices out of 
the semantic net. The possible time stamps are given by the 
Figure 4. State transition graph represented by concepts of a 
semantic net. To each temporal relation a priority and a 
transition time can be assigned.