Multisource data and knowledge fusion for intelligent SAR sea ice classification

The authors describe the fusion of various data and knowledge sources for intelligent SAR sea ice classification, thereby addressing the weaknesses of each information source while improving the overall reasoning power of the classifier. They equip their ice classification system, ARKTOS, with the capability of analyzing and classifying images unsupervised by emulating how a human geophysicist or photo-interpreter classifies SAR images. To imitate human visual inspection of raw images, they have designed and implemented a data mining application that first categorizes pixels into regions, and then extracts for each region a complex feature set of more than 30 attributes. In addition, they have incorporated other sea ice data and knowledge products such as ice concentration maps, operational ice charts, and land masks. Finally, they solicited human sea ice expertise as classification rules through interviews, and collaborative refinements during the early-stage evaluations. Using a Dempster-Shafer belief system, they are able to perform multisource data and knowledge fusion in ARKTOS´ rule-based classification. ARKTOS has been installed at the National Ice Center and Canadian Ice Service