Deep Learning for Ground Penetrating Radar Velocity Inversion

Abstract:

Ground Penetrating Radar (GPR) is a robust geophysical imaging technology that is able to image the shallow subsurface. GPR velocity models are prerequisite for accurate migration. While conventional velocity analysis methods are designed for multi-offset GPR data, to our knowledge, the velocity analysis method for zero-offset GPR has not been explored. Inspired by recent deep learning seismic impedance inversion, we propose a deep learning guided technique that is based on convolutional neural networks (CNNs) to directly learn the intrinsic relationship between GPR data and electromagnetic (EM) velocity. The neural network takes in GPR data as input and outputs the corresponding EM velocity. First, we simulate numerous zero-offset GPR data from a range of pseudo-random velocity models and feed the datasets into the neural network for training. Each training dataset comprises of a pairing of zero-offset GPR data and its corresponding 1D EM velocity model. During the training phase, the neural network’s weights and biases are updated iteratively until convergence. This process is analogous to full-waveform inversion in which the best model is found by iterative optimization that is when the simulated data matches observed data. Tests of the method on synthetic GPR datasets show that the predicted velocity models are accurate and comparable to the ground truth. We further tested our method on a field zero-offset data. We qualitatively evaluated our predicted EM velocity model by comparing it to other EM velocity models generated from multi-channel GPR acquisition in previous studies. The predicted EM velocity model show consistent geological interpretations with that of previous studies of the field area. It also reveals hidden geological structures that were otherwise not discovered in previous studies.