USE OF BAYESIAN NETWORKS AS JUDGEMENT CALCULUS IN A KNOWLEDGE BASED IMAGE INTERPRETATION SYSTEM

The increasing amount of remotely sensed imagery from multiple platforms requires efficient analysis techniques. The presented image interpretation system tries to automate the analysis of multisensor and multitemporal images by the use of structural, topological, and temporal knowledge about the objects expected in the scene. The knowledge base is formulated by a semantic net. Temporal knowledge about object states and their transitions is represented in a state transition graph which is integrated within the semantic net. The analysis of multitemporal images is improved by the prediction of possible object states derived from the knowledge base. During analysis the system has to deal with uncertainty and imprecision. Competing interpretations have to be judged to succeed with the most promising alternative. For this reason the measured object properties are compared to the expected ones. A probabilistic judgement calculus based on Bayesian networks is presented which uses the rules of belief updating and propagation. The approach integrates the probabilities of object states and their transitions within the judgement procedure. Hence it is well suited for a multitemporal image interpretation. For an example dealing with the detection of an industrial fairground from a set of aerial images the probabilistic judgement is compared with an existing possibilistic approach. It is shown, that the use of Bayesian networks increases the efficiency of the interpretation process.