Ground-penetrating radar (GPR) images buried pipes as crescent-shaped (hyperbolic) reflections in B-scans, but the reflected amplitude decays exponentially with delay time until receiver quantization noise dominates. Kobayashi and Nakano (2012; English translation 2013) built a statistical emphasis pipeline for these reflections — an Akaike-Information-Criterion-derived 'analytical limit' (AL) bounding the usable delay range, Kullback-Leibler normality screening of scans, Gabor-wavelet time-frequency peak detection, Markov-chain-Monte-Carlo sampling with infinite-Gaussian-mixture clustering — and left two unresolved issues: detection sensitivity degrades near the AL (weak crescents that should be detected are missed), and no emphasis method exists for the region beyond the AL, which the paper postponed on the way to its final goal of accurately estimating buried-pipe positions. The problem therefore remains open: develop and quantitatively validate emphasis/detection methodology that works near and beyond the analytical limit, in service of accurate buried-pipe localization.

Background

Ground-penetrating radar (GPR) is a non-destructive subsurface imaging technique that transmits short electromagnetic pulses into the ground and records the echoes. A B-scan is the two-dimensional image formed by stacking one-dimensional echo traces (A-scans, one per antenna position) as the radar cart moves along a survey line; an A-scan at fixed delay time across positions is called a horizontal scan. Because echoes from off-center positions arrive later than the echo from directly above a buried pipe, a pipe appears in a B-scan as a crescent-shaped (hyperbolic) reflection pattern, called a crescent signal or reflected wave in this literature. A practical motivation is roadwork safety: official drawings of water, gas, and sewer pipe layouts are frequently inaccurate, so pipes must be located before excavation.

Kobayashi and Nakano (IEEJ Transactions on Industry Applications 132(4):487-500, 2012; English translation in Electronics and Communications in Japan 96(10):22-36, 2013) proposed a statistical preprocessing pipeline for automatically 'emphasizing' (marking) reflected waves in measured B-scans, building on their earlier denoising method based on the 2D Gabor wavelet transform (a time-frequency analysis correlating the signal with Gaussian-windowed complex exponentials at multiple scales and positions; the denoising companion paper appeared in the same journals). Their setting is a hand-pushed GPR with delay-time sampling of 0.5 ns, horizontal sampling of 0.01 m, and signed 16-bit fixed-point amplitude recording. Since the amplitude decays exponentially with delay time, deeper samples are progressively corrupted by quantization noise (the rounding error of the analog-to-digital converter), so the signal-to-noise ratio is not uniform over the B-scan. They therefore define an analytical limit (AL): for each delay index ntn_t, compute V(nt)=log⁡10(std⁡{B(nt,nx)}nx)V(n_t)=\log_{10}\big(\operatorname{std}\{B(n_t,n_x)\}_{n_x}\big), the logarithm of the standard deviation of the horizontal scan; find the split index nn that minimizes the sum of two Akaike Information Criterion values (AIC, a model-selection score trading goodness of fit against parameter count) from fitting a straight line to VV on the early interval and a constant on the late interval; the AL is the delay time Δtn0\Delta_t n_0 at the minimizing split. The emphasis pipeline is applied only at delay times at or before the AL; the measured ALs were 128.0 ns (single-layer site), 66.0, 81.5, and 98.5 ns (three multilayer road sites).

Within the usable range, the pipeline screens horizontal scans by a normality test: the B-scan is re-quantized to 8-bit brightness, a 256-level histogram is formed per scan, a Gaussian is fitted to it, and the Kullback-Leibler divergence (KLD, an asymmetric distance between probability distributions) between histogram and fitted Gaussian is computed; scans whose KLD exceeds the scan average are selected as likely crescent-bearing, because background-only scans are assumed to have Gaussian brightness. Peaks are then located on the Gabor-wavelet power map of each selected scan, either by a simple -3 dB threshold (connected 'islands' around the global maximum) or by treating the powered map (∣W∣2)λ\big(|W|^{2}\big)^{\lambda} (with λ=−9/(2ln⁡γ)\lambda=-9/(2\ln\gamma), γ≈0.5\gamma\approx0.5) as a probability distribution and drawing samples with a Metropolis-Hastings Markov chain Monte Carlo sampler on a linearly interpolated surface (M-HI; Markov chain Monte Carlo is a random-walk sampling scheme whose stationary distribution is the target). The samples are clustered with an Infinite Gaussian Mixture Model (IGMM, a Bayesian nonparametric Gaussian mixture whose number of components is inferred rather than fixed), and candidate clusters are filtered by Fisher's exact test (an exact statistical test of equality of proportions) on peak counts and by a Thompson outlier-rejection test (a significance test that flags a value whose deviation from the sample mean exceeds a critical multiple of the sample standard deviation) on cluster-covariance eigenvalues. On four real sites, the cluster means estimated lateral pipe positions with errors of roughly 0.02-0.3 m; all pipes were detected at one multilayer site, and the 2D-MCMC variant (one of the paper's three detector variants, alongside the -3 dB threshold and 1D-MCMC) was the most accurate in the hardest environment.

Two limitations were left explicit. First, near the AL the KLD screening loses discriminability: quantization noise concentrates the amplitude histogram into a few brightness levels, making its divergence from the narrow fitted Gaussian small, so genuine reflections near the AL were not detected (pipes at 1.4 m depth in one multilayer site and pipes at 2 m and deeper in another were missed by all detector variants); the authors state that improving detection performance, including the choice of the detection function, is a high-priority open issue. Second, the emphasis procedure is simply not applied beyond the AL, and the paper explicitly postpones developing detection methods for that region; closing this gap is stated as necessary for the final objective of the research, accurate estimation of buried-pipe positions.

Later literature took other routes. The same authors' follow-up (IECON 2014) replaced the IGMM with a Dirichlet-process crescent-signal mixture model (a Bayesian nonparametric mixture built from a Dirichlet process prior, which infers the component count) and studied estimation accuracy, reporting that depth estimates were biased by the assumption of constant relative permittivity — its own new future work — without addressing the analytical-limit issues. A separate line (Zhou, Chen, and Li, IEEE TGRS 2018; Zhou, Chen, and Hao, IEEE TGRS 2019; Wu, Zhou, and Chen, IEEE Access 2020) developed fully automatic B-scan interpretation and pipeline mapping using gradient-based thresholding, open-scan clustering, hyperbola fitting, and Dirichlet-process mixtures of survey points; it cites and compares with the 2014 model but does not adopt or extend the analytical-limit concept. A 2025 comprehensive review of GPR pipeline B-scan data processing and target recognition describes a field now dominated by machine-learning and deep-learning detectors, still lists the detection of deeply buried weak targets under attenuation as an open challenge, and does not mention the analytical-limit framework at all.

Problem Statement

In the ground-penetrating-radar reflection-wave emphasis framework of Kobayashi and Nakano (2012/2013), processing is restricted to delay times at or before an AIC-derived analytical limit (AL), because beyond it attenuation-driven quantization noise dominates the recorded amplitudes and the emphasis procedure is not applied. Two unresolved issues stand between this framework and its stated final objective of accurately estimating buried-pipe positions:

(1) Near the analytical limit, reflections that should be detected are missed: the Kullback-Leibler normality test used to select target scans loses discriminability there, because quantization noise concentrates the amplitude histogram into a few brightness levels and its divergence from the narrow fitted Gaussian becomes small; the detection function (including the choice of the test function) and overall detection performance need improvement so that weak, deeply attenuated crescent signals are detected.

(2) No emphasis or detection method exists for the region beyond the analytical limit, which the source paper explicitly postponed: a new method is required that can emphasize and detect reflection waves in delay regions where the present procedure is not applicable.

The problem is to develop and quantitatively validate methodology that extends reliable crescent-signal emphasis and detection to the near-AL and beyond-AL regions on multi-environment B-scans with known true pipe positions (the source's measured sites where obtainable, or realistically simulated B-scans reproducing the source's acquisition and 16-bit quantization regime) and to assess how far this advances the final objective of accurate estimation of buried-pipe positions (lateral position and depth), by quantitative comparison against true pipe positions, extending the source evaluation (lateral errors of roughly 0.02-0.3 m within the AL, with near-AL and deeper pipes missed) to the previously excluded regions.

References

  1. Makoto Kobayashi and Kazushi Nakano (2012). A Method for Emphasizing Reflection Waves from Buried Objects by Using Ground-penetrating Radar. IEEJ Transactions on Industry Applications 132(4):487-500 (Japanese original, open access on J-STAGE). DOI 10.1541/ieejias.132.487. URL https://doi.org/10.1541/ieejias.132.487
  2. Makoto Kobayashi and Kazushi Nakano (2013). A Method for Emphasizing Reflection Waves from Buried Objects by Using Ground-Penetrating Radar. Electronics and Communications in Japan 96(10):22-36 (English translation of the 2012 IEEJ paper). DOI 10.1002/ecj.11484. URL https://doi.org/10.1002/ecj.11484
  3. Makoto Kobayashi and Kazushi Nakano (2013). A Denoising Method for Detecting Reflected Waves from Buried Objects by Ground-Penetrating Radar. Electronics and Communications in Japan 96(9):1-13. DOI 10.1002/ecj.11517. URL https://doi.org/10.1002/ecj.11517
  4. Makoto Kobayashi and Kazushi Nakano (2014). Dirichlet process crescent-signal mixture model for ground-penetrating radar signals. IECON 2014 - 40th Annual Conference of the IEEE Industrial Electronics Society, Dallas, TX, pp. 3431-3437. DOI 10.1109/IECON.2014.7049007. URL https://doi.org/10.1109/IECON.2014.7049007
  5. Makoto Kobayashi, Tomohiro Uchikado, and Kazushi Nakano (2012). Wavelet-based position detection of buried pipes from GPR signals by use of angle information. 2012 International Conference on Wavelet Analysis and Pattern Recognition, pp. 390-396. DOI 10.1109/ICWAPR.2012.6294813. URL https://doi.org/10.1109/ICWAPR.2012.6294813
  6. Tomohiro Uchikado, Makoto Kobayashi, Kazushi Nakano, and Seiichi Shin (2012). Emphasis of GPR signals reflected from buried pipes using statistical clustering techniques. 2012 International Conference on Wavelet Analysis and Pattern Recognition, pp. 410-416. DOI 10.1109/ICWAPR.2012.6294816. URL https://doi.org/10.1109/ICWAPR.2012.6294816
  7. Yoshihiro Ogino, T. Uchikado, K. Nakano, Yasuyuki Nakamura, T. Ogawa, and Takashi Matsuyama (2013). Denoising and detection of reflected waves from buried pipes with ground-penetrating radar data. Proceedings of the IEEE Asia-Pacific Conference on Synthetic Aperture Radar (APSAR) 2013. No DOI registered. URL https://www.semanticscholar.org/paper/fd1f84163401ce830dea965595cb89833b12d7c6
  8. Xiren Zhou, Huanhuan Chen, and Jinlong Li (2018). An Automatic GPR B-Scan Image Interpreting Model. IEEE Transactions on Geoscience and Remote Sensing 56:3398-3412. DOI 10.1109/TGRS.2018.2799586. URL https://doi.org/10.1109/TGRS.2018.2799586
  9. Xiren Zhou, Huanhuan Chen, and Tong Hao (2019). Efficient Detection of Buried Plastic Pipes by Combining GPR and Electric Field Methods. IEEE Transactions on Geoscience and Remote Sensing 57:3967-3979 (June 2019 issue; published 2019). DOI 10.1109/TGRS.2018.2889248. URL https://doi.org/10.1109/TGRS.2018.2889248
  10. Qingyuan Wu, Xiren Zhou, and Huanhuan Chen (2020). Underground Pipeline Mapping Based on Dirichlet Process Mixture Model. IEEE Access 8:118114-118124. DOI 10.1109/ACCESS.2020.3005420. URL https://doi.org/10.1109/ACCESS.2020.3005420
  11. Chen Liu, Jue Li, Zhengnan Liu, Sirui Tao, and Mingxuan Li (2025). A comprehensive review of data processing and target recognition methods for ground penetrating radar underground pipeline B-scan data. Discover Applied Sciences 7:310. DOI 10.1007/s42452-025-06791-y. URL https://doi.org/10.1007/s42452-025-06791-y
  12. Man-Sung Kang and Yun-Kyu An (2020). Frequency-Wavenumber Analysis of Deep Learning-based Super Resolution 3D GPR Images. Remote Sensing 12:3056. DOI 10.3390/rs12183056. URL https://doi.org/10.3390/rs12183056
  13. Zhenghuan Xia, Qunying Zhang, Shengbo Ye, Youcheng Wang, Chao Chen, Hejun Yin, and Guangyou Fang (2015). A novel low-frequency coded ground penetrating radar for deep detection. IEICE Electronics Express 12:20150200. DOI 10.1587/elex.12.20150200. URL https://doi.org/10.1587/elex.12.20150200
  14. Soonbo Yang, Shinji Sassa, Kenta Kudai, and Yuki Wada (2024). Effects of Pavement Types and Thickness on the Accuracy of Ground-Penetrating Radar for Detecting Cavity under Pavement at the Waterfront. Japanese Journal of JSCE 80. DOI 10.2208/jscejj.24-17201. URL https://doi.org/10.2208/jscejj.24-17201
  15. G. P. Pochanin (2004). Problems and promising lines of development of UWB ground penetrating radiolocation. 2004 Second International Workshop Ultrawideband and Ultrashort Impulse Signals, pp. 61-66. DOI 10.1109/uwbus.2004.1388049. URL https://doi.org/10.1109/uwbus.2004.1388049