Geoacoustic inversion in shallow water - stochastic and machine learning approaches

Accurate knowledge of the ocean environment is a crucial factor in establishing national security, being intimately tied to accurate threat detection, localization, and identification. The goal of this work is to improve our knowledge of a changing ocean medium by solving inverse problems that use physics of sound propagation and statistical signal processing. This will be achieved by performing geoacoustic inversion, using novel methods that address challenges causing uncertainties and errors in the estimation process. Two of the challenges in such endeavors are (i) the presence of noise in the propagation environment and (ii) possible undersampling of the acoustic field because of availability of a few hydrophones within a receiving array. The impact of these challenges will be alleviated by using Gaussian processes, which quantify the spatial coherence of the field. Gaussian processes allow the denoising of acoustic signals as well as their spatial interpolation so that dense, virtual arrays are generated. Another direction to be pursued is the improvement of existing algorithms for arrival time and amplitude estimation of multiple paths from time-series recorded at vertical line arrays; real data from several experiments will be used. By developing a forward-backward filter, which includes a smoothing step, accurate estimates will be obtained and will be employed along with a forward model in sediment sound speed and attenuation extraction; frequency dependence will be studied. Machine learning approaches and nonlinear regression will also be pursued, with the goal of sediment classification using distinct features of received signals. For example, moments that can be extracted from time-series in different environments are tightly related to sediment geoacoustic properties. Classifiers of particular interest are decision trees, which make classification decisions sequentially by minimizing uncertainty. The methods to be developed within this project will allow the fast identification of environmental properties and can be integrated by the US Navy in already existing techniques for accurate source detection and localization.