Inverting tomographic data with neural nets

Ocean acoustic tomography is a synoptic observation method of the internal structure of the ocean. The underlying idea is to estimate the sound velocity environment (as well as temperature, salinity and currents) by inverting travels times issued from the propagation of sound between sources and receivers. Most of the inversion methods are deriving from a linearization around a prior reference sound velocity profile. The goal of this paper is to present the results of a new nonlinear method for OAT inversions. This method relies upon the ability of neural nets to learn and generalize from examples. A set of sound velocity environments C(z) is build and the arrival time patterns T in given experimental conditions are computed from a ray tracing model. This provides a set of examples of behavior of the forward problem. The inverse mapping is learned from this set of examples by a standard multilayered perceptrons. Then the generalization property of the net is used to estimate unknown sound speed environments from their associated time patterns. The results shows a remarkable efficiency of the method compared to a standard linear method. Examples of inversion are shown in a Mediterranean environment