Permittivity Inversion of Ground Penetrating Radar by Attention-Based Deep Learning

A deep learning network, GPR1DNet-depth, consisting of self-attention mechanism layers and convolutional layers, has been innovatively proposed to obtain 1-D depth domain permittivity from 1-D time domain ground penetrating radar (GPR) data. Convolutional methods can successfully obtain the time domain permittivity model from time domain GPR data. However, it is not easy to obtain the depth permittivity model directly from time domain data due to the spatial misalignment between the two domains. The proposed network adopts an encoder-decoder structure overall. The proposed GPR1DNet-depth network can accurately extract global features using self-attention mechanism layers and obtain local features using convolutional layers. The architecture designed is helpful in representing the local spatial misalignment between time domain data and depth domain model. The synthetic and field data experiments verified that GPR1DNet-depth is more accurate in inverting the depth information of subsurface permittivity. The proposed network has great potential for development in the inversion of permittivity from GPR data and solving the problem of transformations between time and depth domains.