GPR data interpretation by the deep learning with coloring data

Abstract

The experts perform interpretation of GPR (ground penetrating radar) data by extracting characteristic reflection patterns from GPR time sections due to the contrast of the relative permittivity of the underground structure. An amount of acquired GPR data has been massive and keeps increasing due to an improvement of the data acquisition system and aging of the infrastructure to be maintained, therefore, automation and labor saving of interpretation works are demanded. Object recognition ability by deep learning which is one of machine learning methods has greatly improved, and many architectures of deep learning model have been advocated and studied. In particular, AlexNet is a sort of fundamental deep learning model which opened the leap-improving accuracy by using the method such as CNN (Convolution Neural Network) imitating the real visual cortex and it has been applied to a variety of object recognition applications. Today machine learning for interpretation of GPR data is often applied to grayscale images. However, like experts using GPR color images in visual inspection, it is considered that the information is less lost and more useful interpretation result is obtained in color images than gray images. We did the experiments of AlexNet for GPR data interpretation to design a learning model suitable for GPR data interpretation. The experiment result shows the colorized images makes a better performance, 0.9819 in the F value and 0.9875 in the accuracy than gray-scale images input. We also analyzed the internal output of the several steps in the AlexNet to discuss the appropriate learning model design for GPR data.