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

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

Problem Statement
Ground Penetrating Radar (GPR) is a critical non-destructive testing (NDT) technique for inspecting subsurface infrastructure, particularly pipelines. However, traditional interpretation methods are often labor-intensive, time-consuming, and highly susceptible to noise and environmental variability. While deep learning models, such as YOLO and Faster R-CNN, have improved automation, they face significant challenges:

1. Data Scarcity: High-quality, annotated real-world GPR datasets are difficult and expensive to collect due to logistical constraints and the complexity of subsurface environments.
2. Domain Gap: Models trained on synthetic data often fail to generalize to real-world scenarios due to structural inconsistencies and noise differences (the "Sim2Real" gap).
3. Feature Limitations: Standard Convolutional Neural Networks (CNNs) excel at capturing local textures but struggle to perceive global shape structures, such as the hyperbolic reflection patterns characteristic of buried cylindrical objects.

Methodology
The authors propose TE-S2R (TDA-Enhanced Sim2Real), a framework that integrates Topological Data Analysis (TDA) with deep learning to address these limitations. The methodology consists of two primary stages:

1. Shape-Aware Topological Representation (TDA Integration):

   • Instead of converting topological features into abstract statistical vectors (e.g., persistence images), the authors encode topological information directly into the image domain.
   • Using Persistent Homology on 2D B-scan images, the method extracts $\beta_1$ features (loops), which correspond to the closed, crescent-shaped patterns of pipeline hyperbolas.
   • A lifetime-weighted feature map is generated where the intensity of the loop structures is modulated by their topological "lifetime" (death time minus birth time). Features with longer lifetimes, representing significant structural elements, are rendered with higher intensity, while short-lived features (likely noise) are suppressed.
   • This topological map is stacked with the original grayscale GPR B-scan along the channel dimension, creating a multi-channel input that preserves spatial coherence and allows the CNN to learn both local textures and global geometric continuity simultaneously.

2. Sim2Real Training Strategy:

   • Synthetic Pre-training: A dataset of 300 synthetic B-scan images is generated using gprMax (FDTD method) with diverse soil conditions (homogeneous and heterogeneous via fractal generation) and varying pipe parameters. These images are processed through the TDA pipeline to create topologically enhanced synthetic data.
   • Real-World Fine-tuning: The model is pre-trained on the synthetic data and then fine-tuned on a limited set of real-world GPR data collected from three sites in South Korea (Sites A, B, and C).
   • This two-stage approach aims to bridge the distributional gap by leveraging the structural stability of topological features, which are less sensitive to domain-specific noise than raw pixel values.

Key Contributions

• Novel Topological Encoding: The paper introduces a method to directly embed persistent homology-derived features as spatial intensity maps within CNN inputs, rather than using statistical embeddings. This allows the network to access global shape information without architectural modifications.
• Synergistic Framework: The integration of TDA with Sim2Real transfer learning creates a structure-aware transfer mechanism. TDA mitigates the domain gap by emphasizing shape and connectivity over raw pixel intensity, enhancing the robustness of the Sim2Real process.
• Robustness to Noise: By focusing on topological invariants (loops with long lifetimes), the method reduces sensitivity to the noise and environmental variability inherent in field GPR data.

Experimental Results
The study evaluated six model configurations using YOLOv5 and YOLOv11 architectures, employing 5-fold cross-validation on real data (Sites A and B) and testing on an independent site (Site C).

• Performance Metrics: The proposed TE-S2R model achieved the highest performance across all metrics (Precision, Recall, mAP@0.5, and mAP@0.5:0.95) on the independent test set.
  • On the YOLOv5 test set, TE-S2R achieved an mAP@0.5 of 0.643, significantly outperforming the Base model (0.357) and the Sim-only model (0.031).
  • Similar trends were observed with YOLOv11, where TE-S2R achieved an mAP@0.5 of 0.806.
• Ablation Studies: Comparisons confirmed that both the Sim2Real strategy and the TDA augmentation independently improved performance. However, their combination yielded the most robust results, particularly in handling domain shifts between training and testing sites.
• Statistical Significance: Mann–Whitney U tests confirmed that the improvements of TE-S2R over baseline models were statistically significant ($p < 0.01$).

Significance and Claims
The authors claim that this work represents the first study to directly embed homological information as shape descriptors within a deep learning pipeline for GPR analysis. The significance of the work lies in:

• Overcoming Data Scarcity: Demonstrating that synthetic data, when augmented with topological features, can effectively pre-train models for real-world detection tasks where labeled data is scarce.
• Enhancing Geometric Salience: Proving that topological representations improve the detection of specific geometric structures (hyperbolas) that standard CNNs often miss or misinterpret due to noise.
• General Applicability: The framework is presented as a scalable solution for subsurface sensing that can be extended to other grid-based imaging modalities where structural information is obscured by noise.

The paper concludes that the integration of TDA and Sim2Real is mutually reinforcing, providing a reliable solution for subsurface object detection with broad implications for urban planning and infrastructure management.