DMRF-UNet: A Two-Stage Deep Learning Scheme for GPR Data Inversion Under Heterogeneous Soil Conditions

Traditional ground-penetrating radar (GPR) data inversion leverages iterative\nalgorithms which suffer from high computation costs and low accuracy when\napplied to complex subsurface scenarios. Existing deep learning-based methods\nfocus on the ideal homogeneous subsurface environments and ignore the\ninterference due to clutters and noise in real-world heterogeneous\nenvironments. To address these issues, a two-stage deep neural network (DNN),\ncalled DMRF-UNet, is proposed to reconstruct the permittivity distributions of\nsubsurface objects from GPR B-scans under heterogeneous soil conditions. In the\nfirst stage, a U-shape DNN with multi-receptive-field convolutions (MRF-UNet1)\nis built to remove the clutters due to inhomogeneity of the heterogeneous soil.\nThen the denoised B-scan from the MRF-UNet1 is combined with the noisy B-scan\nto be inputted to the DNN in the second stage (MRF-UNet2). The MRF-UNet2 learns\nthe inverse mapping relationship and reconstructs the permittivity distribution\nof subsurface objects. To avoid information loss, an end-to-end training method\ncombining the loss functions of two stages is introduced. A wide range of\nsubsurface heterogeneous scenarios and B-scans are generated to evaluate the\ninversion performance. The test results in the numerical experiment and the\nreal measurement show that the proposed network reconstructs the\npermittivities, shapes, sizes, and locations of subsurface objects with high\naccuracy. The comparison with existing methods demonstrates the superiority of\nthe proposed methodology for the inversion under heterogeneous soil conditions.\n