Classifying GPR Images Using Convolutional Neural Networks

This paper focused on classifying ground penetrating radar (GPR) images of subsurface cylinders according to their depth, size, material, and the dielectric constant of the underlying medium using four different architectures of convolutional neural networks. Two CNNs were newly proposed in this study and then compared to two others that were used by other authors. These CNNs were trained by using a couple of adjusted training options including initial learning rate, learning rate drop factor, and learning rate drop period; which had a positive impact on some of the considered models, while the option maximum number of epochs worked well with all of the considered models. Results showed that the first proposed CNN showed superior performance due to the use of a deep network with a large number of small filters. It was also found that the first proposed CNN could obtain the best results when GPR B-scans were classified according to the cylinders' materials.