Three-dimensional ground-penetrating radar data inversion based on attention mechanism

Three-dimensional inversion of ground-penetrating radar is of great significance for geological surveys, non-destructive testing, and other fields. We proposed a two-stage three-dimensional (3D) inversion network co-attention inversion network based on coordinate attention and attention gate (AG). The network infers the underground 3D structure and the physical characteristics and position shape of the detection target by analyzing the received electromagnetic wave reflection signal. The network first removes underground environmental noise through 3D convolution and coordinate attention mechanisms and then combines group normalization and AGs to design a U-shaped 3D inversion network to establish the best inverse mapping of denoised C-Scan and 3D dielectric constant maps. Compared with the latest two-stage 3D inversion network 3DInvNet based on multi-receptive fields, the experiment results from the simulation data and the actual measurement data show that this method can efficiently invert the distribution of underground dielectric constant, and it has good versatility and robustness with small mean square error, mean absolute error, mean absolute percentage error, and high structural similarity index.