Geomorphic parameter estimations with neural networks using bathymetry and backscatter data.

Abstract

A bistatic scattering strength model (BISSM) for low-frequency acoustics, which uses high-resolution geomorphology as input parameters, has been proposed [Caruthers etal., NOARL, SP023:200:90 (1990)]. Included in these geomorphic parameters are (1) deterministic bathymetry and local bottom-facet mean slope and azimuth, (2) the stochastic parameters, rms slopes in orthogonal directions and roughness, and (3) the empirical acoustic parameter, Lambert/Mackenzie scattering coefficient. The issue of how to obtain these parameters to support wide-area applications of the model led to a consideration of the inverse problem using the model itself at higher frequencies and in the backscatter direction. Swath sonar systems, simultaneously providing high-resolution bathymetry and backscattering strength as a function of grazing angle, would appear tractable as survey tools for these parameters, if the inverse problem is solvable. Presented here is a sensitivity analysis using neural networks to determine the potential validity of this approach. The model is presumed to be valid and is used to simulate noise-free backscatter data for different sets of parameters. Presented is the ability of neural networks to be trained to provide estimates of the desired parameters under these ideal conditions. [Work is supported by CNOC.]