Deep learning-based cross-hole GPR travel time tomography

Abstract

As a non-destructive geophysical detection method that is not limited by the target depth, cross-hole radar can accurately identify the positions and properties of underground cavities. Traditional cross-hole radar methods face significant challenges in stability, accuracy, and computational efficiency. In recent years, with the rapid development of deep learning and its remarkable achievements in the field of imaging, it has been applied in various fields. As a deep learning network structure for image segmentation, the Unet model can realize “end-to-end” image output. Therefore, the Unet model of convolutional neural network is introduced into cross-hole radar travel time tomography in this paper. A deep-learning trave time tomography method based on the nonlinear relationship between travel time data and models is proposed. Firstly, the models with cavities of different permittivities are constructed. The 2D traveltime images of transmitting and receiving pairs are calculated based on the GPR travel time fast-picking algorithm. Then, a Unet model is built to learn the nonlinear mapping relationship between the 2D travel time images and models with cavities. Finally, the intelligent travel time tomography method based on the Unet model of underground cavities is realized. Experiment results of models demonstrate the effectiveness of the proposed method.