Shape-Aware Topological Representation for Pipeline Hyperbola Detection in GPR Data

TL;DR

The study tackles reliable underground-pipeline detection in noisy GPR data by introducing TE-S2R, a framework that embeds shape-aware topological information from persistent homology into CNN inputs and couples it with Sim2Real transfer learning. By focusing on $H_1$ loop structures and encoding their lifetimes into a spatial topological map, the method preserves global shape cues while leveraging pixel-level detail. The two-stage training—pretraining on diverse synthetic GPR data and subsequent fine-tuning on real field data—paired with topological augmentation, yields improved robustness and higher detection accuracy (notably $mAP@0.5$ and $mAP@0.5:0.95$) under domain shifts. This approach demonstrates how integrating TDA-derived shape priors with cross-domain learning can deliver reliable, real-time subsurface object detection and holds potential for extension to other grid-based imaging modalities.

Abstract

Ground Penetrating Radar (GPR) is a widely used Non-Destructive Testing (NDT) technique for subsurface exploration, particularly in infrastructure inspection and maintenance. However, conventional interpretation methods are often limited by noise sensitivity and a lack of structural awareness. This study presents a novel framework that enhances the detection of underground utilities, especially pipelines, by integrating shape-aware topological features derived from B-scan GPR images using Topological Data Analysis (TDA), with the spatial detection capabilities of the YOLOv5 deep neural network (DNN). We propose a novel shape-aware topological representation that amplifies structural features in the input data, thereby improving the model's responsiveness to the geometrical features of buried objects. To address the scarcity of annotated real-world data, we employ a Sim2Real strategy that generates diverse and realistic synthetic datasets, effectively bridging the gap between simulated and real-world domains. Experimental results demonstrate significant improvements in mean Average Precision (mAP), validating the robustness and efficacy of our approach. This approach underscores the potential of TDA-enhanced learning in achieving reliable, real-time subsurface object detection, with broad applications in urban planning, safety inspection, and infrastructure management.

Figures (7)

• Figure 1: Persistent homology applied to grayscale imagery. (a) A typical grayscale image used for topological analysis. (b) Filtration process for constructing a filtered cubical complex through intensity-based thresholding. As the threshold increases from 0 to 255, pixels are progressively included in the complex.
• Figure 2: Topological feature construction from a GPR image. (a) Original GPR B-scan. (b) Persistent homological features extracted via $H_1$ generators. (c) Shape-aware representation formed by stacking the original image and persistence-weighted features along the channel dimension.
• Figure 3: Sim2Real training process for GPR detection. The model is first pre-trained on simulated data and then fine-tuned on field data. In both stages, inputs are transformed into shape-aware representations by combining the original GPR images with persistent homology-derived topological features.
• Figure 4: Overview of the experimental setup and representative data samples. (a) Schematic diagram of the simulation model showing the pipe geometry, grid configuration, and boundary conditions, with relevant dimensions labeled (pipe diameter = 0.5 m, burial depths = 0.3--0.4 m, grid resolution = 6 mm). (b) On-site photograph of the field data collection setup showing the GPR scanning apparatus and buried pipe layout. (c) Example of a simulated B-scan image generated from the synthetic model corresponding to the setup in (a). (d) Example of a real B-scan image collected from Site A, corresponding to a 2 m survey line and a buried pipe at a depth of approximately 0.5 m.
• Figure 5: Line chart comparing the test-set performance of YOLOv5- and YOLOv11-based models across six configurations.