The probabilistic model is constructed by using the information theory and the maximum entropy principle. This probabilistic model exhibits a minimum number of parameters, is a non-Gaussian tensorvalued random field for which its mean function is not constant in the thickness direction, which is then non-homogeneous in space, and for which the random fluctuations are defined by a spatial correlation length in the thickness direction and by a space dependent dispersion parameter controlling the level of the random fluctuations. The parameters which have to be identified (by solving an inverse stochastic problem related to the wave propagation prediction model) are the mean function, the correlation length and the dispersion function of the random field. The purpose is to present a method and an application for this identification using experimental measurements in ultrasonic range. The stochastic inverse problem is carried out by solving an optimization problem based on a new adapted cost functions and on in vivo observations.