Triplanar Imaging of 3-D GPR Data for Deep-Learning-Based Underground Object Detection

Abstract

This article proposes a deep-learning-based underground object classification technique that uses triplanar ground-penetrating radar images consisting of B-, C-, and D-scan images. Although multichannel ground-penetrating radar (GPR) systems provide three-dimensional (3-D) information about underground objects, there is currently no suitable technique available for processing 3-D data as opposed to 2-D images. In this article, a triplanar deep convolutional neural network technique is proposed for use in processing 3-D GPR data for use in automatized underground object classification. The proposed method was validated experimentally using 3-D GPR road scanning data obtained from urban roads in Seoul, South Korea. In addition, the classification performance of the method was compared to that of a conventional method that uses only B-scan-images. The results of the validation and comparison tests reveal that the classification performance of the proposed technique is notably better than that of the conventional B-scan-image-based method and that its use results in decrease misclassification ratios.