Summary

Seabed classification, the inference of physical properties of the sea bottom from multi-beam bathymetric sonar, is generally attempted by analysis of backscatter amplitudes, a concept which is successfully applied to side-scan sonar data. Backscatter from multi-beam bathymetric sonars, however, involves several different physical scattering processes that depend on the angle of incidence of the sonar beam, and which have to be accurately modeled in order to obtain a normalized brightness image of the scattering surface. This is a classical inverse problem in which some a-priori knowledge of the physical properties of the surface and its orientation relative to the instrument platform is required in order to obtain a first order model.

Commercial systems are available that use a different approach: Multivariate statistical analysis of sonar returns, is taken in order to classify single-beam echo-sounder data

Such a system can perform what has been termed un-supervised classification by cluster analysis: Signals are grouped into classes according to their statistical properties. For very much the same reasons, the simple statistical approach fails with multi-beam data; returns from different angles of incidence, and thus different scattering domains, exhibit very different statistics.

The method presented in this report may be rightfully claimed by both the artificial neural-network (ANN) research community and those investigating generalized sub-space methods. Accordingly, the terminology varies with the perspective one wants to adopt. In ANN terminology the classification system proposed in this interim report employs a series of associative memories (linear NNs) which vote independently on a beams membership in one particular class. Votes are then subjected to a fuzzy decision rule which determines to which class, if any, the multi-beam return belongs.

With the limited test data available we have confirmed that the method :

• gathers returns from different scattering domains into one class.
• classifies beams, independent of their incidence angle.
• preserves the spatial resolution of the multi-beam dataset.
• is computationally feasible on a small workstation requiring minutes to load the memories with training-sets and fractions of a second to classify a beam.
• degrades in a gradual controlled way with increasing noise levels.

We now need to show that this method actually classifies acoustic diversity related to the physical properties of the sea-bottom. A dataset collected over a region with known sea-bottom geology and distinct geological features is needed to draw further conclusions.

Download this document in

 PDF-format