Application of artificial neural networks to geoacoustic inversion
Date
1998
Authors
Benson, Jeremy
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Artificial neural networks (ANNs) were used to estimate ocean bottom parameters from experimental data collected at sea during the TRIAL SABLE experiment. Both backpropagation ANNs (BPNNs) and radial basis function ANNs (RBFNNs) were investigated. These networks use supervised training to learn the behavioural mapping from the data to the bottom parameters. The input/output examples of the training data were generated using the geoacoustic forward model ORCA, which predicts the experimental data from the geoacoustic properties of the ocean and ocean bottom. The network was trained to approximate the inverse of the ORCA model. The TRIAL SABLE experiment involved a vertical line array of hydrophones which sampled the pressure field generated by a towed multi-tone continuous wave source. From the a priori information available for the experiment site we were able to construct a set of nominal ocean parameters with an upper and lower bound for each parameter. Using ORCA we show that the field is strongly dependent on the ocean parameters which are not known precisely. We train two types of neural networks to estimate these parameters from the data. The BPNN achieves good estimation accuracy for several of the most important parameters, but other parameters are poorly estimated. The RBFNN estimates the single most important parameter very well but the other parameter approximations are inaccurate. We propose that this performance difference is due to the choice of training algorithm. The problem in estimating the less important parameters is overcome by simplifying the task to one of bottom classification. The bottom classification is done with the two neural networks using supervised training. The BPNN performs very well with only a 4% chance of misclassification. The RBFNN has a 20% probability of misclassification. The BPNN classifies the data as resulting from a gravel bottom, and the RBFNN chooses sand. The bottom type is known a priori to consist of a mixture of sand and gravel.