A machine learning algorithm for GPR sub-surface prospection

The paper presents a novel approach for the (semi-) automatic extraction of sub-surface layers´ properties from GPR data. The methodology solves the inverse scattering problem by means of artificial neural networks which are able to map proper features derived from the electromagnetic signal onto the dielectric permittivity and thickness of the layer which has backscattered the radiation. The whole procedure is first described and then tested over a set of simulated scenarios and their corresponding GPR traces, showing high reconstruction accuracies and denoting the opportunity of a wide range of applicability.