Semisupervised Deep Neural Network-Based Cross-Frequency Ground-Penetrating Radar Data Inversion

Ground-penetrating radar (GPR) with different center frequencies can detect defects at different depths with a range of resolutions enabling it to be used for subsurface defect inspection. However, the existing deep learning methods cannot accurately invert the permittivity from GPR data of different frequencies, due to the limited number of labeled GPR images for every center frequency. To tackle this challenge, a semisupervised deep neural network (DNN)-based cross-frequency GPR data inversion method was proposed, which enables the generalized model to be trained on the GPR data with one frequency (source domain) for migration to other frequencies (target domain). The method was trained in a semisupervised manner using a small number of paired GPR data with permittivity labels and a large amount of unlabeled GPR data without corresponding permittivity maps. An adversarial learning mechanism together with a novel random perturbation strategy was designed to improve the global inversion performance for a large-scale structure and avoid a discontinuity in the reconstructed shapes. Furthermore, a mean-teacher architecture is introduced to improve the inversion accuracy of detailed information from the unlabeled GPR data under different perturbation conditions. The ablation and comparative experiments results indicated that the proposed method outperforms other methods and can be effectively generalized to GPR B-Scan data with different frequencies and signal-to-noise ratios. In addition, sandbox model testing was conducted and the results indicate that this method can transfer the knowledge from the synthetic data domain to the real data domain with satisfactory results.