Computer classification of sea beds by sonar

Abstract

Direct sampling of the sea bed is not always feasible and can be unreliable because the samples obtained may not be representative. An alternative approach is to classify the superficial sea bed using the backscattered signal from side-scan sonars. The usual paper display contains less information than the signal creating it, because of its low dynamic range, and interpretation of sediment types is often difficult. There are also problems with statistical analysis of the signals¹. The usual methods of pattern recognition cannot resolve small-scale features in otherwise homogeneous regions² and, although the backscattering strength is a useful quantity, meaningful numerical values can be obtained only with a calibrated system³. To overcome these shortcomings we have now developed a spectral analysis method that is insensitive to system gain calibration. The method retrieves information from the signals that is well below the usual power spectrum point, thus effectively widening the system bandwidth, and can discriminate between the signals returned from six sea-bed types.