Geoacoustic inversion with artificial neural networks

A geoacoustic inversion technique involving artificial neural networks (ANNs) is proposed to estimate ocean bottom properties and source information from experimental data. The method is applied to data from the TRIAL SABLE experiment that was carried out in shallow water off Canada´s east coast. The inversion is designed to incorporate the a priori information available for the site in order to improve the estimation accuracy. The inversion scheme involves training feedforward ANNs to estimate the geoacoustic and geometric parameters using simulated input/output training pairs generated with a forward model. The inputs to the ANNs are the spectral components of the received pressure at each sensor for the two lowest frequencies used, 35 and 55 Hz. The output is the set of environmental parameters corresponding to the received field. In this way the ANNs effectively simulate an inverse model. In order to decrease the training time a separate network was trained for each parameter. The results for the parallel estimation error are 10% lower than that for the bulk estimates, and the training time is decreased by a factor of 6. When the experimental data are presented to the ANNs the geometric parameters such as source range and depth are estimated with high accuracy. The compressional speed in the sediment and the sediment thickness are found with moderate accuracy