Geoacoustic model inversion using artificial neuralnetworks

Abstract

An inversion technique using artificial neural networks (ANNs) is described for estimating geoacoustic model parameters of theocean bottom and information about the sound source from acoustic field data. The method is applied to transmission lossdata from the TRIAL SABLE experiment that was carried out inshallow water off Nova Scotia. The inversion is designed toincorporate the a priori information available for thesite in order to improve the estimation accuracy. The inversionscheme involves training feedforward ANNs to estimate thegeoacoustic and geometric parameters using simulatedinput/output training pairs generated with a forward acousticpropagation model. The inputs to the ANNs are the spectralcomponents of the transmission loss at each sensor of avertical hydrophone array for the two lowest frequencies thatwere transmitted in the experiment, 35 and 55 Hz. The output isthe set of environmental model parameters, both geometric and geoacoustic, corresponding to the received field. In order todecrease the training time, a separate network was trained foreach parameter. The errors for the parallel estimation are 10%lower than for those obtained using a single network toestimate all the parameters simultaneously, and the trainingtime is decreased by a factor of six. When the experimentaldata are presented to the ANNs the geometric parameters, such assource range and depth, are estimated with a high accuracy. Geoacoustic parameters, such as the compressional speed in thesediment and the sediment thickness, are found with a moderateaccuracy.