A classification approach to the characterization of seabed geoacoustic profiles via deep learning

Acoustic source localization algorithms in the shallow ocean require information about the ocean seabed. While past approaches have characterized seabed geoacoustic profiles via continuous physical parameters, a classification approach approximates the distribution of seabed types as a discrete set of seabed classes. To exclude acoustically indistinguishable seabeds from the discrete classes, the acoustic separation between two seabeds may be evaluated by the Pearson correlation between their characteristic transmission loss curves. This process yields a subset of discrete seabed classes. To distinguish these classes, we train a convolutional neural network classifier on the pressure time-series of SUS charges, which are simulated using the geoacoustic profiles of the seabed classes. This network successfully predicts seabed class, when tested on a holdout set where the sound speed profiles of the ocean are adequately similar to those in the training set. An acoustic separation measure between sound speed profiles, analogous to the one for seabeds, directly relates to the generalization error of the network. [Work supported by the Office of Naval Research Contract No. N00014-16-C-3065 and by the NSF REU program, Grant No. 1757998.]