Automatic seabed classification by the analysis of sidescan sonar and bathymetric imagery

The authors present a technique for making use of both sidescan amplitude and bathymetric data provided from sidescan bathymetric sonars for the classification of underwater seabeds. Sidescan amplitude is corrected for physical factors and used to plot ´processed´ sidescan images. Both amplitude and textural features are derived from these images. Textural features are obtained using 2-D discrete wavelet transforms. Bathymetric images are used to derive features indicating seafloor variability. These features are combined together and the most relevant ones are selected by feature selection algorithms. If grab samples are available, the areas around them are used as training data in a supervised approach. The backpropagation elimination algorithm is used on the training dataset to select relevant features. If training data are not available, an unsupervised approach can be used. The dimensions of the whole dataset are reduced using principal component analysis in this case, and the principal components are used as features. In both cases, clustering techniques are used to cluster the data into sediment classes. The classified points are then plotted against their GIS position in the survey. Classification results correlate with grab sample types from the areas considered (in the supervised case) and with expert observation of sidescan images, where training data is not available.