Neural network for geoacoustic inversion of sub-bottom profiler data

Sub-bottom profilers are utilized to extract features pertaining to the sub-seafloor environment sediment stratification. Acquisition and analysis of sub-bottom profiles can provide insight into the sediment composition and acoustical properties. Typical analysis of profiles involves computationally expensive inversions, such as model-based or Bayesian techniques, which require large computational costs. Here, a neural network is developed to perform a geoacoustic inversion on simulated sub-bottom profiler data. The network is used to derive attenuation and acoustical impedance measurements corresponding to the layered media. Geoacoustic properties of the layered sediments are compared to values determined through a direct inversion of reflection coefficient, testing how well these techniques recover the ground truth values. The network, trained on simulated data, is tested on real sub bottom profiler data acquired over a well-studied area called the New England Mud Patch, roughly 80 km south of Nantucket. The simulated data-trained network is compared to a network trained on experimental data acquired by the R/V Tioga over the same test region.