Repository hosted by TU Delft Library

Home · Contact · About · Disclaimer ·
 

Bayesian reverberation inversion incorporating grain-size dependent regression relations as a priori information

Publication files not online:

Author: Abrahamsson, L. · Andersson, B.L. · Ivansson, S. · Pihl, J. · Chalindar, B. · Cristol, J.X. · Juhel, B. · Eidem, E.J. · Dybedal, J. · Olsen, G.K. · Ainslie, M.A. · Benders, F.P.A. · Colin, M.E.G.D. · Vossen, R. van
Type:article
Date:2012
Source:Proceedings of the 11th European Conference on Underwater Acoustics - ECUA 2012, 2-6 July 2012, Edinburgh, UK
Identifier: 462563
Keywords: Defence · Low-Frequency Active Sonar · Sea Bed Parameters · Backscattering Strenght · Acoustic Reverberation · Inversion · Mathematical models · Safety and Security · Defence, Safety and Security · Physics & Electronics · AS - Acoustics & Sonar · TS - Technical Sciences

Abstract

An operational low-frequency active sonar system has been used in the EDA (European Defence Agency) project Rumble-2 to collect reverberation data at sea trials in the North Sea. Using a fast ray model for the forward computations, the reverberation data are inverted to determine bottom parameters: Lambert back-scattering parameter μ and sound speed c, density ρ, absorption α, and sediment thickness h. Hamilton-Bachman regression relations for c, ρ, and α, are incorporated with uncertainties and with the mean grain size Mz as a common descriptive parameter. This fits nicely into a Bayesian inversion framework, but with a non-uniform prior pdf (probability density function). Markov-chain Monte-Carlo techniques are applied to produce an ensemble of models distributed in model space according to the PPD (posterior probability density), which is thereby assessed. Laplace distributions are chosen to model the data residuals, or data errors, and particular attention is given to estimation of the corresponding covariance matrix in the non-stationary case. There appears to be a risk of finding a point in model space for which the reverberation trace matches the noisy data trace better than the “true“ model. This could lead to too small data error estimates and a too narrow estimate of the PPD. Application to the actual Rumble-2 reverberation data pings indicates that the mean grain size Mz and the Lambert parameter μ are reasonably well determined, whereas there are uncertainties concerning the sediment thickness. The pings involve bottom areas mainly covered by sand and silt, respectively, and the inversion results reflect this difference in the anticipated way