2019_programme: WAVEFIELD-ACCURATE SEABED CLASSIFICATION USING LOCALIZED FORWARD MODELING AND DEEP LEARNING



  • Session: 13. Machine learning, compressive sensing and signal processing
    Organiser(s): Xenaki Angeliki, Gerstoft Peter
  • Lecture: WAVEFIELD-ACCURATE SEABED CLASSIFICATION USING LOCALIZED FORWARD MODELING AND DEEP LEARNING [invited]
    Paper ID: 877
    Author(s): Frederick Christina, Michalopoulou Zoi-Heleni, Villar Soledad
    Presenter: Frederick Christina
    Presentation type: oral
    Abstract: The key to a successful recovery of seafloor characteristics from measured backscatter data generated from SONAR systems is finding a balance between the computational cost of forward modeling and the desired resolution. In the high frequency regime, many propagation models are often limited by costly simulations or unrealistic environmental assumptions, so a detailed recovery is not always possible. There is a growing need for sophisticated mathematical and computational tools that accommodate complicated scenarios, i.e., multiple seafloor layers or uncharted seafloor landscapes. To enable a rapid, remote, and accurate seafloor parameter recovery, we propose a combination of localized forward modeling and machine learning. The idea is to partition environments into much smaller “template” domains, a few meters in width, where the sediment layer can be described using a few parameters. Machine learning is used to train a classifier using a reference library of simulations of Helmholtz equations on the domains. Our results shed light on the potential of deep learning for classification of seafloor properties, such as sediment type, roughness, and layer thickness using the reference library first in a high-frequency regime using backscatter. These methods are also applied to synthetic data generated by a normal mode propagation model, that efficiently and accurately approximates the acoustic field at lower frequencies. We expect to classify features that smoothly vary on a scale comparable to the acoustic wavelength.
  • Corresponding author: Dr Frederick Christina
    Affiliation: NJIT
    Country: United States
    e-mail: