Unsupervised Detection of Underground Tunnels in Ground-Penetrating Radar Using Depth-Restricted Reconstruction Scoring

Clandestine tunneling poses a significant threat to critical infrastructure like oil and gas pipelines, yet existing monitoring systems typically only detect damage after a breach has occurred. While ground-penetrating radar (GPR) can image these tunnels, manual interpretation is slow and subjective, and supervised machine learning models are often impractical due to the scarcity of labeled tunnel data. The authors propose a fully unsupervised detection pipeline trained exclusively on normal, tunnel-free subsurface radargrams.

The core of the system is a denoising convolutional autoencoder that learns to reconstruct normal ground structures. Because the model is never trained on tunnel signatures, it fails to reconstruct them accurately at inference time. The authors introduce a novel depth-restricted top-k anomaly score, which calculates the reconstruction error only within the lower half of the radargram—the depth band where tunnels are physically expected to exist. This approach avoids the dilution of anomaly signals that occurs when averaging errors across the entire image.

The study demonstrates that incorporating geometric prior knowledge—specifically the depth restriction—is more effective than complex post-processing. Applying this restriction improved the area under the curve (AUC) from 0.986 to 0.994 and reduced missed detections by 77%. The authors also discovered an interaction between the pooling fraction and the depth restriction: while a 1% pooling fraction is optimal for full-image analysis, a 5% fraction is superior when the scoring is restricted to the relevant depth band.

Furthermore, the research characterizes the utility of spatial aggregation (voting across overlapping windows). While such post-processing helps improve weak detectors, it offers no benefit to the final, high-performing system, as the remaining errors are systematic rather than isolated. The final model achieved an F1 score of 0.975 and a recall of 0.973 on a held-out test set, demonstrating that a well-tuned unsupervised detector can effectively identify rare anomalies without any labeled training data.