Enhanced GPR Inversion Method in Complex Underground Environment

Ground penetrating radar (GPR) data is difficult to interpret due to the multiple complex features exhibited by complex media in practical engineering applications. However, it is extremely arduous to extract key features of different targets in B-scan data due to the intersection of hyperbolas representing multiple media targets in a complex environment, increasing the difficulty of interpreting the echo signals. Additionally, deep learning methods for end-to-end supervised learning are faced with challenges in addressing the matching relationship between the dimensional space of B-scan data and the spatial scale of actual target media, which leads to confusion in determining the spatial positions of underground targets during inversion. In this paper, an inversion method based on Kirchhoff migration and dynamic snake convolution is proposed. First, the Kirchhoff migration algorithm is introduced to preprocess the B-Scan data as a data dimension transformation unit, transforming forward space to imaging space. This preprocessing process achieves data dimension compression, alleviating the computational burden and costs and improving the accuracy of the target position for imaging. Then, the dynamic snake convolution is utilized to enhance the U-net architecture, improving the network's perceptibility to different targets by emphasizing local features with hyperbolic slender and curved shapes. Finally, to further assess the disparity between the imaging outcomes and the ground truth while considering variables such as target morphology, location, and dielectric properties, the combined MSELoos and SSIMLoss are employed for updating network parameters through backpropagation, evaluating the differences between imaging results and the ground truth in terms of pixel-level errors and structural similarity. The results show that the proposed method provides better generalization performance on the target inversion in a complex underground background compared with other state-of-the-art data-driven inversion algorithms.