Deep Learning for Applications to Ground Penetrating Radar and Electromagnetic Diagnostic

In this paper, a Machine Learning (ML), and more specifically, a Deep Learning (DL) approach, is applied to the resolution of a typical electromagnetic problem such as the analysis and classification of Ground Penetrating Radar (GPR) radargrams. In particular, the study employs a DL architecture, known as DenseNet, to classify a set of radargrams, generated through the gprMax simulation software, and representing the scattering from perfect electric conductor (PEC) cylinders of infinite length, buried in various media, at different depths, and with different radius amplitudes. The network was trained to be able to extract the information hidden in the images through a multi-label approach. The objective of this study is both the generation of a dataset that can be employed to train DL algorithms and the exploration of the suitability of the DenseNet architecture to the classification of radargrams. The study has shown interesting results in terms of the ability of the DenseNet in extracting multi-label informations from radargrams given a small set of images. Limits of the approach considered are also highlighted and addressed.