Deep-Neural-Network-Based Subsurface Defect Identification Method for Multipolarimetric Ground-Penetrating Radar Data

A deep-neural-network-based identification method for multipolarimetric ground-penetrating radar (GPR) data was proposed to address the challenges of unbalanced polarimetric data and take advantage of multipolarimetric GPR data for subsurface defect identification. The proposed method comprised two subnetworks—namely, an unbalanced polarimetric data augmentation module based on a cycle generative adversarial network (CycleGAN) to generate a large number of paired multipolarimetric GPR data, and an identification module based on multipolarimetric GPR data to accurately identify internal defects with irregular shapes and occlusions. Specifically, the proposed physics-informed convolution (PIC) in the identification module could enhance the scattering features of different polarimetric data and extract location-sensitive physical properties from multipolarimetric GPR data; thus, defects with occlusions could be effectively identified. Additionally, the identification module adopted a criss-cross attention (CCA) module and a global attention module (GAM) channel, which could reduce information reduction and magnify global interactive representations to accurately identify the shape of irregular targets. Validation experiments were conducted at two levels—that is, first, using synthetic multipolarimetric data based on the subsurface defect model, and second, using a sandbox model test in a realistic scenario. The results demonstrated that the proposed method was capable of effectively identifying internal defects with irregular shapes and occlusions from multipolarimetric GPR data.