Efficient localization and spectral estimation of an unknown number of ocean acoustic sources using a graphics...

This paper develops a matched-field approach to localization and spectral estimation of an unknown number of ocean acoustic sources employing massively parallel implementation on a graphics processing unit (GPU) for real-time efficiency. A Bayesian formulation is developed in which the locations and complex spectra of multiple sources and noise variances are considered unknown random variables, and the Bayesian information criterion is minimized to estimate these parameters, as well as the number of sources present. Optimization is carried out using simulated annealing and includes steps that attempt to add/delete sources to/from the model. Closed-form maximum-likelihood (ML) solutions for source spectra and noise variances in terms of the source locations allow these parameters to be sampled implicitly, substantially reducing the dimensionality of the inversion. Source sampling, addition, and deletion are based on joint conditional probability distributions for source range and depth, which incorporate the ML spectral estimates. Computing these conditionals requires solving a very large number of systems of equations, which is carried out in parallel on a GPU, improving efficiency by 2 orders of magnitude. Simulated examples illustrate localizations and spectral estimation for a large number of sources (up to eight), and investigate mitigation of environmental mismatch via efficient multiple-frequency inversion.