International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
• Production systems (Stilla, 1995; Stilla and Michaelsen, 
1997) 
• Blackboard systems (Nagao and Matsuyama, 1980) 
• Fuzzy logic (Zadeh, 1987) 
• Possibility theory (Dubois and Prade, 1985) 
• Semantic nets (Quint, 1997a; Growe, 1999; Kunz, 1999) 
The most common approaches are ruled-based systems and 
semantic or Bayesian nets. During the last period, possibility 
theory and fuzzy logic have been increasingly used. Semantic 
networks in particular are attractive as they are flexible and 
extendable and allow a complete modelling. Growe, 1999 gives 
a nice example of their use, including extensions towards 
multitemporal analysis. However, criticism has been expressed 
that to built-up the network for each specific application is very 
time-consuming and complicated, and thus such structures, 
although theoretically and conceptually attractive, are not 
suitable for implementation in commercial systems and use in 
production practice. 
Independently of the approach used, the above systems provide 
very limited intelligence with respect to object recognition as 
compared to humans. The major problems, apart from the 
difficulty to transfer knowledge to a computer and encode it in a 
computer-understandable form, are the following. The human 
decision process is highly nonlinear. A single indicator may 
lead us to change opinion radically and vote for an option that 
had a small weight up to this point. Computers lack this ability 
and their decisions are based on hard numbers (fuzzy logic 
although it helps in combining information more "softly", still 
does not solve this problem). Lack of learning and sensing 
capabilities is another computer weakness. A final point, which 
in our opinion is very decisive, is the lack of memory with 
computer processing. For each individual task and application, 
computers start with very little or no accumulated knowledge 
and "experience". During the many years of development of 
image analysis algorithms, the connection and relation to data 
and knowledge bases has been almost totally ignored. The 
integration of image analysis and GIS can thus also provide 
benefits in this respect, as data storage and management systems 
are integral GIS components. 6 
6. SOME IMPORTANT POINTS 
Before concluding, we would like to summarise here some 
major points: 
• Use of knowledge, rules, models 
Although the last period the above information has been 
increasingly used, it still is very weakly integrated in image 
analysis of aerial and satellite imagery. 
• Quality indicators 
Reliable quality indicators of partial results are a must; they 
should be provided as locally (fine-grained) as possible. 
• Need of redundancy 
Data fusion is often mentioned as a process to exploit 
complementarities of the data. This is a correct statement. 
However, data should also be combined just to provide 
redundancy. This improves accuracy and especially reliability 
and is one of the few counter-measures to account for the lack 
of algorithm intelligence mentioned above. 
• Reliability 
The lack of reliability in the results is the single, most crucial 
factor, which hinders an increased use and efficiency of 
automated procedures. 
• Need of extensive tests 
Published results, especially from academia, refer to very 
limited datasets, which are used for years, and after careful 
tuning of the algorithm parameters in order to get good 
results. Use of extensive, comprehensive datasets in 
cooperation with the industry is necessary. This will provide 
not only a framework for an objective comparison of methods 
but also provide useful feedback and possibilities of 
improvement of the used algorithms, strategies and input data. 
• System architecture and strategy 
To account for various scenes, objects, object types and 
applications, without having to redevelop for each case 
everything from scratch, the whole strategy, architecture and 
control of the image analysis system should be modular, 
flexible and adjustable. 
• Quality of algorithms 
The quality of the individual processing algorithms (e.g. for 
edge extraction, 3D matching etc.), although probably not the 
most important factor in achieving high quality results, can 
make a difference and should be chosen accordingly. 
• Selection of input data and preprocessing 
This aspect can make a huge difference in performance. While 
development of sophisticated algorithms with poor input data 
might cost a lot and bring little, more suitable data, permitting 
an easier detection of the target object and its discrimination 
from other ones, might lead to significant improvement of the 
results, even when using simpler processing algorithms. The 
fact that an increasing number of sensors become available 
should be exploited in this direction. The criteria of data 
selection should serve the better separability of the target 
object, provision of data complementarity and redundancy, 
while correlations between the data should also be modelled 
and taken into account. 
• Importance of data structures and use of databases 
As mentioned above, the system should have an active and 
continuously increasing memory through accumulation of 
knowledge and data. To store and manage this information, 
appropriate data structures are needed, which are 
unfortunately not provided yet by commercial systems. Such 
data structures should be able to include heterogeneous and 
multivariate data (multi-valued, raster, vector, attribute etc.), 
support flexible data types, include 3D and temporal 
modelling. An object-oriented paradigm with inheritance, 
encapsulation, methods associated with the objects etc. could 
be useful in building up such data structures. 
• Quality control of final results and feedback 
Quality control is naturally essential but apart from that it 
should be used for the analysis of the size of the occurring 
errors, where they occur and why. The results should also be 
used to derive statistical data on the extracted objects and 
their attributes, which can be accumulated in a "ground truth" 
database and used in future similar processing tasks, e.g. for