Range and depth estimates from array invariant and model inversion, compared with machine learning approach

Abstract

The array invariant is derived from the waveguide invariant, and can be used to predict the shape of multipath arrival patterns on a 2D plane with axes of time and angle of arrival. The shape observed on a vertical line array is an ellipse that is a function of the waveguide invariant and the range. Once the range is obtained, ray tracing can be used to invert the multipath pattern for a depth estimate. We will compare results of using the physics-based range and depth estimator with results obtained by training a neural network on synthetic data produced by a ray tracer. We will produce two sets of exemplars to train our estimator. In one, we will provide the range as an input, and train the network to produce a depth. In the second, we will train the network to produce both range and depth. The method derived from the array invariant is expected to be brittle with respect to range-dependent bathymetry. We will compare how well the machine learning approach deals with such variability, compared to the array invariant approach.