Symposium Abstract: Quantitative Measures of Acoustic Diversity to Support Benthic Habitat Characterization
J. M. Preston, A. C. Christney, W. T. Collins, and R. A. McConnaughey
The fundamental dataset produced by an acoustic classification system is a representation of the acoustic diversity of the sediments in the survey area. Each acoustic record, from a ping, a stack of pings, or a section of a sonar image, is transformed to a feature vector, typically in two or three dimensions. Features may be from spectral analysis or from integration of parts of an echo envelope. Rather than classifying sediments with just these few features, it is often more useful and adaptable to generate many features and use multivariate statistical techniques to select the linear combinations that capture most of the variance in the dataset. Classification can then be done by dividing the records into groups based on the values of the most important, typically three, principal components. A difficult step in this classification process is estimation of the appropriate number of clusters. Motivated by the need for an automated seabed classification process that is both objective and adaptable to a wide variety of survey applications, this paper describes objective methods for choosing the number of clusters, based on information theory. Actual classifications provide insights into acoustic diversity, which can be used as a proxy for change in sediment characteristics including the influence of benthos. QTC IMPACT™ calculated 166 features from each stack of a very large set of echoes from the Bering Sea. An optimum classification scheme, using the three most important principal components, was identified, based on K-means clustering guided by finding minima using information theory techniques.