Detecting Subsurface Voids From GPR Images by 3-D Convolutional Neural Network Using 2-D Finite Difference Time Domain Method

Abstract

In this article, an algorithm for detecting subsurface voids under the road from ground penetrating radar images is proposed. A multichannel radar system mounted on vehicle enables dense and highspeed monitoring. The novelty of the algorithm is a unique ElectroMagnetic simulation method and state-of-the-art deep learning technique to consider three-dimensional (3-D) reflection patterns of voids. To train deep learning models, 3-D reflection patterns were reproduced by 2-D finite difference time domain method to drastically reduce the calculation cost. Hyperboloid reflection patterns of voids were extracted by 3-D convolutional neural network (3D-CNN). The classification accuracy of 3D-CNN was up to $90$%, about 10% improvement compared to previous 2D-CNN to demonstrate the effectiveness of 3-D subsurface sensing and detection. The results were validated by real void measurement data. After applying trained 3D-CNN to radar data, regions of voids were plotted in a 3-D map, offering clear visualization of areas of voids.

Keywords

• 3-D convolutional neural network (3D-CNN)
• deep learning
• finite difference time domain (FDTD) method
• ground penetrating radar (GPR)
• subsurface voids