Acoustic seabed segmentation from direct statistical clustering of entire multibeam sonar backscatter curves.

Abstract

A fast, simple method is presented to obtain acoustic seabed segmentation from multibeam sonar backscatter data, for situations where processed backscatter curves are already available. Unsupervised statistical clustering is used to classify multibeam sonar backscatter curves in their entirety, with the curves essentially treated as geometrical entities. High variability in the backscatter curves is removed by along-track averaging prior to clustering, and no further preprocessing is required. The statistical clustering method is demonstrated with RESON 8125 multibeam sonar data obtained in two bathymetrically complex environments. These are a sandwave field in Keppel Bay, Queensland, and an area of inter-island sand, reef, seagrass, and rhodolith beds in Esperance Bay, Western Australia. The resulting acoustic charts are visually compelling. They exhibit high spatial coherence, are largely artifact free, and provide spatial context to comparatively sparse grab samples with relatively little effort. Since the backscatter curve is an intrinsic property of the seafloor, the mappings form standalone charts of seafloor acoustic properties. In themselves they do not need ground truthing. Conceptually, use of the full angular backscatter curve should form the primary means of obtaining acoustic seabed segmentation. However, this is dependent on the scale and configuration of seabed backscatter features compared to the dimensions of the averaged swathe used to obtain reliable realisations of the backscatter curve.