Considerations in large-scale acoustic seabed

No abstracts are to be cited without prior reference to the author.

In 1999, NMFS Alaska and QTC collected about 18,000 line miles of seabed acoustic data at 38 and 120 kHz from the eastern Bering Sea. With four million echoes at each frequency, this data set permitted thorough explorations of some practical considerations that influence every acoustic seabed classification. Our unsupervised classification involved an objective determination of the optimal number of classes for each of the pre-classification methods we explored, allowing useful comparisons among methods. Stacking, one of the pre-classification steps, is the process of averaging sequential echoes to allow sediment information to express itself in spite of ping-to-ping variability. With stacks of fifty pings, feature spaces had more detail and better defined clusters, thus more classes in unsupervised classification, compared to stacks of five pings. Classification by echo shape requires changing the sampling rate to compensate for depth changes. While effective, this changes the apparent roughness and the amount of detail submitted to the feature-generating algorithms. Depth and stack size affect spatial resolution; the scale of the survey and the sharpness of sediment boundaries guide the surveyor’s choice of spatial resolution. Even such a huge data set is a sampling of the sea bottom, and further sub-sampling simplified feature spaces further, reducing the optimal number of classes. The two frequencies differed in beam width and sediment penetration, and thus gave complimentary information. The influences of each of these pre-classification processes and other considerations are presented, with representative maps of acoustic diversity and with statistical comparisons, accompanied by preliminary correlations with fish census data.