Acoustic seabed segmentation for echosounders through direct statistical clustering of seabed echoes

A new method is presented for inferring seabed type from the properties of seabed echoes stimulated by echosounders. The methodology currently used classifies echoes indirectly through feature extraction, usually in conjunction with dimensional reduction techniques such as Principal Components Analysis. The features or principal components derived from them are classified by statistical clustering or other means into groups with similar sets of mathematical properties. However, a simpler technique is to directly cluster the echoes themselves. A priori modelling or curve fitting, feature extraction, and dimensional reduction are not required, simplifying the processing and analysis chain, and eliminating data distortions. In effect the echoes are treated as geometrical entities, which are classified by their shapes and positions. Direct clustering places the analysis focus on the actual echoes, not on proxy parameters or mathematical techniques. This allows simple and direct evaluations of results, without the need to work in abstract mathematical spaces of unknown relation to echo properties. The direct clustering method for seabed echoes is demonstrated with echosounder data obtained in Balls Head Bay, Sydney Harbour, Australia, an area with mud, sand, and shell beds.

► A simple new method of echosounder based acoustic seabed classification is presented. ► Current indirect methods rely on feature extraction and/or dimensional reduction. ► Direct statistical clustering of the actual echoes is used instead. ► Data are not distorted and mathematical complexity and abstraction are removed. ► Direct clustering of curves can be used for many applications in the geosciences.