Multiaspect Classification of Sidescan Sonar Images: Four Different Approaches to Fusing Single-Aspect Information

Abstract

In this paper, the classification of an object on the seabed from two sidescan sonar looks is considered. Four different approaches to fusing the information from the individual looks are described. The first method uses a kernel regression classification method with the combined feature (CF) vectors from the two looks. The other three approaches are based upon the Dempster-Shafer (DS) fusion of the outputs from a single-look kernel-based classifier. They differ with respect to the manner in which the DS masses for each look are derived from the single-look classifier. The four approaches are evaluated with Klein 5500 sidescan sonar data collected from four classes: rock and three dummy mine types: manta, rockan, and cylinder. The sonar data were collected in October 2005 when the Defence Research and Development Canada Atlantic (DRDC-Atlantic, Dartmouth, NS, Canada), participated with the NATO Undersea Research Centre (NURC, La Spezia, Italy) and Groupe D'Etudes Sous-Marines de l'Atlantique (GESMA, Brest, France) in a joint trial (CITADEL) with the DRDC remote, semisubmersible vehicle DORADO. Using these data, it is shown that two of the approaches for DS mass assignments yield two-look classification performances very similar to each other and to the CF method. The DS approaches have the advantage that they do not require explicit training with a CF set.