Abstract:

Objective: To address the problems of complex detection images with ground penetrating radar (GPR), difficult interpretation, and low inversion accuracy in tree root zone, an improved fully convolutional networks (MWFCNet) based on migration weight guidance is proposed to invert the relative dielectric constant of tree root zone, achieving high-precision inversion and reconstruction of the underground relative dielectric constant environment in tree root zone, providing an efficient and reliable technical means for non-destructive testing of tree roots and detection of root zone soil environment, aiming to provide new tools and methods for in-depth research on the interaction mechanism between trees and soil dielectric environment. Method: The mature triploid Populus tomentosa root zone environment was taken as the research object. The open source software gprMax was used to generate GPR B-scan simulation samples, and combining CycleGAN to achieve sample style transfer, and construct 3 000 pairs of training samples for GPR B-scan and corresponding two-dimensional relative dielectric constant models of measurement line profiles. To solve the problem of poor inversion performance of background media by inversion networks, GPR migration image sequences and their corresponding migration weight sequences were introduced into the input module, and a network architecture with encoder decoder as the backbone was constructed. Two different convolution sizes were used for parallel processing, and multi-scale feature extraction of feature images was achieved through skip connections. The image features were further integrated with fully connected layers to enhance feature expression ability, and output a two-dimensional relative dielectric constant model of the measured root zone underground. Structural similarity index (SSIM) was selected, peak signal to noise ratio (PSNR), and mean squared error (MSE) were used as evaluation indicators for GPR inversion performance, and background variance was used as an evaluation indicator for the degree of background medium restoration. Result: Compared with existing methods such as Enc-Dec, U-net, PInet, etc., the MWFCNet method improved SSIM by 0.11%?3.23%, MSE by 0.11?0.73, and PSNR by 0.31?5.83 dB in the inversion of the same test set. In the restoration of the background medium, the background variance of the MWFCNet method decreased by 0.035?0.15. Conclusion: The MWFCNet based method for inverting the relative dielectric constant of tree roots can accurately identify the position of thick roots of trees, and also achieve two-dimensional reconstruction and restoration of the underground relative dielectric constant map of GPR survey lines. Combined with GPR sampling method, it can reconstruct and restore the three-dimensional relative dielectric environment of the root zone underground.

Key words: improved fully convolutional neural network based on migration weight (MWFCNet), relative dielectric constant, tree root zone, ground penetrating radar, B-scan image inversion