Matched-field geoacoustic inversion using Gaussian Processes for field prediction

Abstract

Matched-field geoacoustic inversion is the process of maximizing a correlation measure between received data at deployed hydrophones and replica fields calculated via an acoustic model for a set of values for environmental parameters, frequently along with source range and depth. The values that produce the maximum correlation form the geoacoustic parameter and source location estimates. Matched-field inversion performance improves as the number of receiving hydrophones increases and the water column is sampled in a dense manner. In order to achieve dense water column sampling when only a few phones are available, we predict the field at a large number of virtual depths within a given aperture using Gaussian Processes. Similarly to classical matched-field inversion, the predicted field, computed after the choice of a suitable kernel and its hyperparameters, is correlated with replicas calculated at the same phone depths. It is demonstrated with both synthetic and real data that the Gaussian Process matched-field inversion approach outperforms conventional processing. The effect is more pronounced as the Signal-to-Noise Ratio attains lower values. [Work supported by ONR.]