Applications of neural networks to ocean acoustic tomography

Abstract

Ocean acoustic tomography differs from medical ultrasound tomography and seismic tomography in that one must first understand the forward problem, that is, how the sound channel and the mesoscale feature refracts sound in three dimensions and how such refraction alters the pulse-arrival sequence. The parabolic equation (PE) model is used in the forward problem. A neural network is used to perform the inversion of tomography data. The author uses the feedforward neural network to implement the filtered back projection algorithm. The advantages are that one does not need to assume weak scattering and the instability problem of the frequency domain interpolation algorithm does not exist.<>