The deep learning techniques applied to electromagnetic imaging via ground-penetrating radar

• Main Author: Dai, Qiqi
• Other Authors: Abdulkadir C. Yucel
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/171847
• Institution: Nanyang Technological University

Ground-penetrating radar (GPR) has been widely used for non-destructive inspection of subsurface structures. In GPR inverse problem, the recorded GPR data is used to reconstruct subsurface permittivity maps, of great significance for mapping the subsurface utilities as well as detecting, recognizing, and monitoring the subsurface objects and living entities (such as tree roots). Traditional iterative algorithms developed for reconstructing permittivity maps from GPR data suffer from high computational costs and low accuracy when applied to complex scenarios. To tackle these issues, deep learning-based techniques have recently been proposed to characterize the mapping from the GPR data to the subsurface permittivity maps. However, these deep learning techniques suffer from three limitations. First, these techniques only take into account the ideal homogeneous soil environments; the noise and clutter due to heterogeneity of the subsurface environment are neglected, which can severely interfere with the objects’ signatures and lower the accuracy of reconstructed permittivity maps. Second, since these techniques only reconstruct the 2D (sectional) permittivity maps from B-scans, they cannot provide 3D geometrical information about the subsurface objects, such as their complete shapes and orientations. Third, these techniques use traditional physics-based GPR forward solvers for dataset generation, which require excessive computational time, especially when a large dataset is required for training and testing stages. To address these limitations of deep learning algorithms for GPR data inversion and imaging, three deep learning-based approaches for reconstructing 2D/3D subsurface permittivity maps and generating B-scans for given subsurface scenarios are proposed in this thesis. First, a two-stage deep neural network (DNN), called DMRF-UNet, is proposed to reconstruct the permittivity distributions of subsurface objects from GPR B-scans under heterogeneous soil conditions. In the first stage, a U-shape DNN with multi-receptive-field convolutions (MRF-UNet1) is built to remove the clutter due to the inhomogeneity of the heterogeneous soil. Then the denoised B-scan from the MRF-UNet1 is combined with the noisy B-scan to be inputted to the DNN in the second stage (MRF-UNet2). The MRF-UNet2 learns the inverse mapping relationship and reconstructs the permittivity distribution of subsurface objects. To avoid information loss, an end-to-end training method combining the loss functions of two stages is introduced. A wide range of subsurface heterogeneous scenarios and B-scans are generated to evaluate the inversion performance. The tests performed on numerical and real measurement data show that the proposed network reconstructs the permittivities, shapes, sizes, and locations of subsurface objects with high accuracy. The comparison of the proposed DMRF-UNet with existing methods demonstrates its superiority for the inversion under heterogeneous soil conditions. Next, a 3D deep learning scheme, called 3DInvNet, is proposed to reconstruct 3D permittivity maps from GPR C-scans. The proposed scheme leverages a prior 3D CNN with a feature attention mechanism to suppress the noise in the C-scans due to subsurface heterogeneous soil environments. Then a 3D U-shaped encoder-decoder network with multi-scale feature aggregation modules is designed to establish the optimal inverse mapping from the denoised C-scans to 3D permittivity maps. Furthermore, a three-step separate learning strategy is employed to pre-train and fine-tune the networks. The proposed scheme is applied to numerical simulation as well as real measurement data. The quantitative and qualitative results show the networks’ capability, generalizability, and robustness in denoising GPR C-scans and reconstructing 3D permittivity maps of subsurface objects. Finally, to alleviate the computational burden of the training stage of the deep learning-based inversion and to rapidly generate a large and diverse dataset, a fast deep learning-based GPR forward solver is proposed to predict the GPR B-scans of subsurface objects buried in the heterogeneous soil. The proposed solver is constructed as a bimodal encoder-decoder neural network. Two encoders followed by an adaptive feature fusion module are designed to extract informative features from the subsurface permittivity and conductivity maps. The decoder subsequently constructs the B-scans from the fused feature representations. To enhance the network’s generalization capability, transfer learning is employed to fine-tune the network for new scenarios vastly different from those in the training set. Numerical results show that the proposed solver achieves a mean relative error of 1.28%. For predicting the B-scan of one subsurface object, the proposed solver requires 12 milliseconds, which is 22,500x less than the time required by a classical physics-based solver. With its excellent efficiency and generalization capability, this solver serves as a promising forward solver for data augmentation and training.