Seabed classification using localized forward modeling and deep learning

Abstract

We use full wave solutions to generate synthetic high-frequency data for training seafloor classifiers with acoustic signals in the ocean. We expect to recover high level details that cannot be achieved by standard forward models that do not include seafloor roughness. By creating acoustic templates using micro local analysis of full-wave simulations on smaller domains, we avoid the high cost of solving the equations on large domains, the main obstacle in full wave inversion. To demonstrate the usefulness of this modeling, we generate a training library of templates and show how well these can be used in classifying seafloor parameters. We consider two-layer seafloors with varying material types and thicknesses. We address the geoacoustic inversion problem with standard machine learning techniques and more sophisticated deep learning methods. The standard classifiers provide results with worse accuracy that do not generalize well to other test environments; the deep learning classifiers are more costly to train but have higher accuracy, generalizing better. We compare the performance with simulations performed at lower frequencies using a normal mode approach with no roughness effects, where the classification problem becomes simpler. [Work supported by NSF, EOARD, and ONR.]