Representing sea ice knowledge in a Dempster-Shafer belief system

Abstract

The authors describe the sea ice knowledge acquisition and representation effort involved in an intelligent system called the Advanced Reasoning using Knowledge for Typing Of Sea Ice (ARKTOS) for satellite sea ice classification. ARKTOS first uses a segmentation technique to extract contiguous sea ice regions (features), then computes attributes of each feature, and finally applies an expert system to classify the features. To classify the features using an expert system, the authors obtained transcripts regarding the sea ice applications and domain through interviews with sea ice experts. A qualification process ensued to identify relevant attributes and to also provide a consistent reference for all rules. To enable the utility of the rules in a Dempster-Shafer belief system, each rule is also associated with a number that indicates how confident one feels about the assertions made by the rule if it is actually fired. Once fired, the belief system accumulates the evidence and handles conflicting hypotheses systematically to arrive at a classification of the feature. The authors' current implementation of ARKTOS facilitates improvement of the reasoning process by allowing expert users to evaluate and edit numerous aspects of the rules or sea ice knowledge conveniently. Presently, they have included a graphical user interface package to enable visual evaluation for such knowledge refinement. ARKTOS is able to classify ERS and RADARSAT sea ice imagery satisfactorily based on the current rule bases.