Efficient analysis of ocean-acoustic big data from 3D seabed surveys with Bayesian machine learning and
Abstract
" Shallow-water continental-shelf environments represent acoustic wave guides, where underwater sound waves interact with the sea surface and seabed as they propagate. Therefore, geophysical seabed properties profoundly influence the propagation of underwater sound in shallow-water continental-shelf environments. Consequently, any technology that applies underwater sound in these environments, for example sonar, relies on knowledge about the seabed. Most currently employed methods that aim to provide such knowledge make simplifying assumptions about the physical processes governing sound propagation. Importantly, seabed structures are often treated as simple with little or no stratified sediment layers and sediments are assumed to be purely elastic or viscoelastic. The inference process and reporting of knowledge are also often simplified by ignoring parameter uncertainty, instead focusing on only optimal parameter estimates. This work employs recent research advances in signal processing, mathematicalgeophysics, statistics, computational seismology, and acoustics as a basis to develop new probabilistic inference methods to analyze acoustic data from recent ONR experiments. We propose to use these recent data sets to develop tools that can be applied at scales relevant to Navy applications on continental shelves. We will employ several key technologies to develop these new methods: (1) Bayesian statistics are applied as a for uncertainty quantification. (2) We reduce requirements for simplifying assumptions by using advanced numerical methods (Markov-chain Monte Carlo and convolutional neural networks) for parameter estimation. (3)The physical processes of sound interacting with the seabed will be described by a poro-elastic seabed model (viscous grain shearing, VGS, theory). (4) To control the number of parameters required for VGS models, we will employ Bayesian model selection. This approach is capable to infer the complexity of a seabed model (the number of sediment layers and the required complexity of physical theory) from the acoustic data. (5) To address the resulting computational costs for this general treatment, we employ massively parallel computer technology that combines central processing units and graphics processing units. We propose to employ 2Dtrans-dimensional parametrizations to characterize large areas of seabed more efficiently and more rigorously than is possible under current assumptions. For this work, we will consider both data that are simulated for realistic shallow-water environments and data recorded by Pennsylvania State University (PSU) at the New England Mud Patch using a ship-towed source and receiver array. Similar data were collected autonomously using an AUV for a site on the Malta Plateau. These data were previously analyzed and may also be considered here to benchmark our new development. This proposal is a collaboration between University of Calgary, PSU, and University of Victoria. The description of this effort can be located in R2 Sub-Activity Ocean Sciences in PE 0601153N."
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 17, 2020
- Source ID
- N000142012631
Entities
People
- Jan Dettmer
Organizations
- Office of Naval Research
- United States Navy