A 50 kHz echo sounder and a digital waveform acquisition and seabed classification system were used to record and process echoes from embayments near southeast Vancouver Island, Canada. Each echo was preprocessed and several statistical and spectral algorithms were used to extract its characteristic patterns. The patterns were grouped into four acoustic classes. Ground truth was available from each site allowing each cluster to be associated with a seabed type. The sediment properties and acoustic classification were compared using multivariate correlation analysis. The frequency of occurrence of fine sand was found to have large negative correlation with the penetration of a free-fall penetrometer. Fig. 1 Location of study sites, indicated by circles. Inset charts show Royal Roads and Esquimalt Harbour of southern Saanich Peninsula, and Wanoose Harbour, about 15 km northeast of Nanaims. 0-7803-4108-2/97/$10.00 0 1 997 IEEE I . INTRODUCTIONThe amplitude and shape of an acoustic signal reflected from the sea floor are determined mainly by the sea bottom roughness and by the contrast in acoustic impedance between water and bottom. The remote classification of the sea bottom requires an acoustic data acquisition system and a set of algorithms that analyze the data, determine the seabed type and relate the results of the acoustic classification to the physical properties of the marine sediments. Several studies have been conducted recently in the area of experimental recognition of the seabed using a wide-band chirp sonar [ 1,2] or a parametric array [3]. The data processing relies usually on the extraction of characteristic features from the echo. Classification implies some kind of ordination technique to group echoes with similar features. An inversion technique, similar to seismic inversion, could also be used for the seabed discrimination. The modeling of acoustic backscatter from sandy and muddy sediments has been developed recently (see for example, [4] and references therein). The forward models developed can be used as a basis for an iterative inversion algorithm such as a singular value decomposition inversion or various conjugate gradient algorithms. In this study, a more robust statistical approach was used. Compared to inversion techniques, the current approach does not require a detailed model of sound propagation. A 50 kHz echo sounder and a digital waveform acquisition system (ISAH-S) were used to record the echo envelope, and a QTC VIEW Seabed Classification System [5] was used to process the data. Preprocessing of the echo involved sea floor identification and filtering of the records. Several algorithms were used to extract characteristic patterns from the preprocessed individual echoes. Statistical analysis of characteristic patterns yielded a small number of linear combinations of features accounting for most of the variation in the data set. Plotted in the space of three primary indices, the echoes grouped into four clusters through unsupervised classification. Extensive ground truthing was avai...