A comparison of Bayes', Dempster-Shafter and endorsement theories for managing knowledge uncertainty in the context of land cover monitoring.

Three commonly used techniques for combining uncertain evidence are explored with reference to different three types of land cover knowledge: numerical distributions, relative spectral distances, and human expert “rules of thumb”. In attempting to combine such evidence Bayes’, Dempster-Shafer and Endorsement theories answer different questions depending on nature of the land cover evidence (completeness and format). The approaches therefore have different utilities in the development of automated approaches to land cover monitoring. Whilst Bayes’ and Dempster-Shafer theories may be more useful in situations where evidence is expressed numerically, Bayes’ theorem requires a complete probability model. The advantage of Endorsement theory derives from its ability to represent different kinds of evidence in a natural form. It is a fundamentally symbolic approach that represents and reasons with knowledge of real-world problems and allows inferences to be drawn from partial knowledge. Suchan approach is advantageous in knowledge acquisition and expert system development.