Training data generation for machine learning using GPR images

Abstract

Ground Penetrating Radar (GPR) systems allow the acquisition of images displaying the contents of the underground. Hence, GPR is used everywhere, where structures underneath the visible surface have to be investigated. Consequently, typical application fields are archeology and civil engineering, especially the detection of cables, pipes or other manmade objects. GPR sensors can consist of one channel or of multiple channels, placed side by side. In the latter case, it is possible to acquire a two-dimensional image for each measurement, where the number of channels represents the number of columns in the image matrix. Since a typical track of measurements often contains multiple of thousands GPR images, a visual analysis with focus on the detection of buried objects might be uneconomically. Moreover, due to its noisy characteristic in relation to the specific underground, it is often not easy to interpret GPR images immediately. In this study, an unsupervised approach is presented, that provides both help for the visual analysis of GPR images and for the detection of potential buried objects. Therefore, it is usable to quickly generate or enlarge training datasets for machine learning approaches aiming at the analysis of GPR data. As test data, several measuring tracks acquired by the multi-channel Stream C system at the site of Frankfurt University (GER) are available. The workflow consists of two central processing steps: Change detection and data augmentation.