Automatic Hyperbola Detection and Apex Extraction Using Convolutional Neural Network on GPR Data

Abstract

Ground Penetrating Radar (GPR) is a non-destructive geophysical method used for subsurface mapping. It is frequently used for detecting buried cylindrical objects such as underground pipes and cables. Buried cylindrical objects show a hyperbolic signal pattern on a radargram. The typical shape of the hyperbolic reflections depends on the depth and material of the buried objects and the surrounding materials. In many cases, detecting buried cylindrical objects is quite a time-consuming task, thus limiting further interpretation procedures. In this paper, we propose a new method for automating hyperbola detection and apex extraction on radargram. Our work consists of two modules that take radargram as input in a form of images. In the first module, we used the Faster-RCNN to extract the hyperbola segments as a set of rectangular boundary boxes. The network was trained using synthetic radargram data simulated by the gprMax software. The second module is to estimate the coordinates of the hyperbola apex using an image processing technique. We had correctly detected all hyperbola on the simulated radargram from the test set. For the test on field radargram, the framework is capable of processing radargram that is similar to the simulated radargram data. The problem with the second module occurs on interference hyperbola as the searching window is disturbed by the noise. Apart from those problems, by using these two modules, the detection of buried cylindrical objects using GPR can be automated with a minimal amount of time.