Novel deep learning based approaches for clutter removal and target detection in GPR

Abstract

Ground Penetrating Radar (GPR) is a non-destructive geophysical tool that uses electromagnetic waves to create detailed images of shallow underground structures and buried objects. The basic operating principle of this system is based on the interaction of waves emitted by the transmitter antenna with materials in the subsurface that have different dielectric properties (permittivity) and conductivity. Signals returning from different layers or objects are recorded by the receiver antenna and processed in the time-depth plane. Particularly in B-scan images obtained by moving the antenna along a line, point targets (e.g., pipes, mines, or archaeological remains) appear as characteristic hyperbolic (inverted U) signatures resulting from the change in distance as the antenna approaches and moves away from the target. However, the raw GPR data obtained in practice is quite different from ideal theoretical models. Direct interaction between antennas (crosstalk), strong surface reflections at the air-soil boundary, scattering caused by inhomogeneous ground structure, and multiple reflections caused by the signal echoing between layers create intense noise components referred to in the literature as "clutter." These noises generally mask weak signals returning from the target, reducing image quality and making target detection a serious signal processing and engineering problem. To solve this problem, statistical subspace methods such as Singular Value Decomposition (SVD) and Principal Component Analysis (PCA) have been used in the literature for many years. However, these linear methods are insufficient for separating the target signal from noise in complex and variable ground conditions and often lead to the loss of target data. Moreover, although advanced techniques such as Low-Rank Sparse Decomposition (LRSD) offer improvements, they suffer from a critical dependency on the manual tuning of regularization parameters. Although Convolutional Neural Networks (CNNs) have become popular with the advent of deep learning, the "limited receptive field" problem of these architectures prevents the capture of long-range (global) context spread across the entire image, such as hyperbolic reflections in GPR radargrams. Furthermore, most existing studies treat the processes of clutter removal and target detection as separate, sequential blocks. In this approach, even the slightest data corruption during the cleaning phase can cause the target to be completely missed during the detection phase. This thesis contributes four original and innovative architectures to the literature to overcome the aforementioned methodological and structural limitations. First, an Autoencoder Guided Low-Rank Approximation (AE-LRA) approach is introduced to address the parameter dependency of classical LRSD methods by automatically learning the low-rank clutter distribution from the data. Second, DC-ViT (Densely Connected Vision Transformer), based on the Vision Transformer architecture, was developed to overcome the local focus limitation of CNNs. This model uses the "Self-Attention" mechanism to model the relationship between information at any point in the image and all other points, capturing the global context. At the same time, thanks to the densely connected connections and "Locally-enhanced Feed-Forward" (LeFF) blocks integrated into the structure, the preservation of local details, which Transformers are generally weak at, is ensured. Third, the DC-Mamba architecture is presented as a solution to the quadratic computational load problem that Transformer models create with high-resolution data. This model adapts the "State Space Models" (Mamba) architecture, which can model long-range dependencies in data with linear computational complexity, to the GPR domain. In particular, the Tensor-Train (TT) decomposition technology used in the model's V2 variant radically reduces the model's number of parameters and memory usage, enabling it to operate with high efficiency on embedded systems with limited processing power (FPGA, mobile processors, etc.). The fourth and most comprehensive solution, RAFDEC-YOLO, is an end-to-end "Joint Learning" architecture that combines clutter removal and target detection tasks in a single network. The "Residual Adapter Fusion" (RAF) mechanism developed in this model transfers noise-free feature maps obtained from cleaning layers to the detection neck without disrupting the gradient flow of the main detection network. Thus, the network learns not only to visually improve the image but also to highlight features that will maximize detection success. The performance of the developed models was validated using extensive synthetic hybrid datasets and sandbox experiments conducted under challenging laboratory conditions. Comparative experiments indicated that the proposed AE-LRA method achieved greater Signal-to-Clutter Ratio (SCR) improvement than RPCA and NMF methods, without the need for manual parameter adjustment. The DC-Mamba architecture also delivered superior image quality and processing efficiency, producing the cleanest B-scan images, especially in conditions with intense signal attenuation, such as beneath asphalt, and high surface complexity. On the target detection side, RAFDEC-YOLO detected weak targets with high reliability even in data where sequential methods completely failed, and targets were lost in noise. Furthermore, thanks to the integrated training strategy, it was shown that the model's bounding boxes encircle targets much more tightly and accurately. This study shows that joint models considering the global context in GPR are efficient and can share information between tasks. As a result, they deliver results that are much better, more reliable, and more practical than traditional methods.