Kernel methods for the detection and classification of fish schools in single-beam and multibeam acoustic data

A kernel method for clustering acoustic data from single-beam echosounder and multibeam sonar is presented. The algorithm is used to detect fish schools and to classify acoustic data into clusters of similar acoustic properties. In a preprocessing routine, data from single-beam echosounder and multibeam sonar are transformed into an abstracted representation by multidimensional nodes, which are datapoints with spatial, temporal, and acoustic features as components. Kernel methods combine these components to determine clusters based on joint spatial, temporal, and acoustic similarities. These clusters yield a classification of the data in groups of similar nodes. Including the spatial components results in clusters for each school and effectively detects fish schools. Ignoring the spatial components yields a classification according to acoustic similarities, corresponding to classes of different species or age groups. The method is described and two case studies are presented.