Automatic detection of single-hyperbola pipeline signatures in urban roads using data-augmented ground penetrating radar profiles

• A GAN-IET-GPR-SCPVD method is proposed for automatic detection and localization of urban underground pipelines. • GPR-GAN is employed to generate high-quality samples, addressing data scarcity and enhancing model performance. • A lightweight IET-GPR model is constructed, integrating IDWC, ELA-HSFPN, and TADDH modules to improve detection capability. • The SCPVD algorithm achieves a vertex localization error of approximately 3.27%, with an average runtime of 0.1169 seconds per instance. Currently, urban underground pipelines face challenges such as incomplete or inaccurate information records and unexpected leakages. Ground Penetrating Radar (GPR), with its high resolution and non-destructive detection capabilities, can generate a vast amount of subsurface pipeline imagery. However, the interpretation of GPR data still largely relies on manual analysis, which is inefficient and highly subjective.To address these issues, this paper proposes an automatic pipeline detection and localization method based on profile analysis, termed GAN-IET-GPR-SCPVD. First, a GPR-GAN is employed to augment the training dataset. Then, automatic detection is performed using a YOLOv11-based framework(IET-GPR), in which an Inception Depthwise Convolution (IDWC) multi-branch structure is integrated into the backbone to expand the receptive field.The neck incorporates an Efficient Local Attention–High-level Screening Feature Fusion Pyramid Network (ELA-HSFPN) to accurately capture key spatial regions. Furthermore, a Task Dynamic Align Detection Head (TADDH) is adopted to enhance the interaction between classification and localization tasks. Finally, a Symmetry-Constrained Parabolic Vertex Detection (SCPVD) method is proposed to perform vertex localization of hyperbolic reflections induced by pipelines. Experimental results demonstrate that the data augmentation network effectively alleviates the issue of limited training samples. The improved YOLOv11 architecture reduces model parameters by 17.4% while maintaining a high precision of 96.7%, ensuring both model efficiency and accuracy. The SCPVD method achieves a localization error of 3.27%, significantly improving the system’s positioning capability and practical applicability. Overall, the proposed method provides a complete pipeline—from data augmentation to automatic detection and vertex localization—enabling accurate and real-time interpretation of underground pipelines.