💻 computer science

TriView-YOLO: Early Multi-View Fusion for Ground Penetrating Radar Cavity Detection in Soft, High-Water-Content Soils

This paper introduces TriView-YOLO, a multi-view YOLOv12-based detector that fuses longitudinal, horizontal, and cross-section GPR views to achieve automated cavity detection in challenging soft, high-water-content soils where traditional methods struggle due to signal attenuation.

Original authors: Suphawut Thawinutchokaudom, Sompote Youwai, Warat Kongkitkul, Mitsumasa Yamashina, Jose M. D. S. Rodrigues Neto, Jun Shinohara

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Suphawut Thawinutchokaudom, Sompote Youwai, Warat Kongkitkul, Mitsumasa Yamashina, Jose M. D. S. Rodrigues Neto, Jun Shinohara

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to find a hidden treasure buried under a busy city street. Usually, you might use a metal detector that beeps when it senses something shiny. But what if the ground is a giant, soggy sponge, soaked with water? In this wet, muddy environment, your metal detector gets confused. The water acts like a heavy blanket, soaking up the signal and making the "beep" of the treasure incredibly faint, or even silent. This is the daily nightmare for engineers trying to find empty holes, or "cavities," under roads. These holes are dangerous because they can swallow a car without warning, but finding them in soft, water-logged soil is like trying to hear a whisper in a hurricane.

To solve this, scientists use a special tool called Ground Penetrating Radar (GPR). Think of GPR as a high-tech flashlight that shoots invisible waves into the ground and records the echoes. When the waves hit a hole, they bounce back, creating a picture. However, in wet soil, these pictures are often blurry, broken, and full of static noise. For years, experts have had to stare at these messy images by hand, squinting to guess where the holes are. This is slow, tiring, and prone to human error. The big question for the scientific world has been: Can we teach a computer to be a better detective than a tired human, even when the ground is a soggy mess?

This is exactly what the researchers behind the new paper, TriView-YOLO, set out to do. They didn't just build a smarter computer; they built a computer with "super-vision." Instead of looking at the ground from just one angle, like a person peering through a single keyhole, their new system looks at the same spot from three different perspectives simultaneously.

Here is how their invention works, using a simple analogy: Imagine you are trying to spot a hidden object in a dark room. If you only have one flashlight, you might miss it if it's hidden behind a shadow. But what if you had three flashlights shining from the front, the side, and the top all at once? Even if the front light is blocked by a shadow, the side or top light might catch a glint of the object.

The researchers took this idea and applied it to road safety. They created a system called TriView-YOLO. It takes three different "views" of the same patch of road:

The Long View (B-scan): A side-profile slice, like looking at a loaf of bread cut in half.
The Top View (C-scan): A flat map looking down from above, like a map of a city.
The Cross View (B-scan): Another side slice, but from a different angle, like looking at the bread from the other side.

Normally, a computer might try to look at just one of these slices. But in the wet, muddy soil of Bangkok (where the study was tested), a single slice is often too blurry to see the hole. The researchers' system, however, fuses all three slices together into one giant, 9-channel image. It's like giving the computer a 3D glasses set that lets it see the "ghost" of the hole even when the signal is weak. They trained this system on 1,600 real-world images taken from actual road surveys in Bangkok, where the soil is incredibly soft and holds a massive amount of water (up to 140% of its own weight!).

The results were promising, but with a very important reality check. The new system found the hidden holes significantly better than any single-view computer model. It managed to spot about 63% of the hidden cavities that were marked by human experts, compared to only about 47% for the single-view models. In the world of detecting invisible holes in muddy ground, this is a big step forward. It means the computer is catching more of the dangerous "ghosts" that would have otherwise been missed.

However, the paper also rules out some popular shortcuts that other scientists often try. The researchers tested if they could just add in "cleaner" pictures from computer simulations or public databases to help the system learn. They found that this actually made the system worse. It's like trying to teach a student who needs to pass a test in a stormy forest by showing them pictures of a sunny beach; the student gets confused when they actually enter the forest. The system learned that the "clean" pictures didn't match the messy reality of the wet soil, so it got less accurate. Similarly, using pre-trained models (computers that learned to recognize cats and dogs) didn't help much either. The lesson here is that for this specific, muddy problem, you cannot use generic data to bypass the need for data that looks exactly like the messy reality you are trying to solve.

The final takeaway is that while the system isn't perfect yet—it still misses some holes and sometimes raises false alarms—it is a powerful new tool. It proves that by looking at the problem from three angles at once, we can see through the "fog" of wet soil better than ever before. This doesn't mean roads will be safe tomorrow, but it gives city planners a much sharper pair of eyes to find the hidden dangers before they swallow a car. The system is fast enough to scan roads in real-time, acting as a helpful assistant that flags suspicious spots for human experts to double-check, turning a slow, manual search into a rapid, automated screening process.

Technical Summary: TriView-YOLO for GPR Cavity Detection in Soft, High-Water-Content Soils

Problem Statement
Automated detection of subsurface cavities using Ground Penetrating Radar (GPR) faces its most significant challenge in soft, high-water-content soils, such as the marine clay found in Bangkok, Thailand. In these environments, conductive, water-saturated soil strongly attenuates electromagnetic signals, degrading cavity reflections and rendering them weak, noisy, and difficult to distinguish from background clutter. While cavities form most readily under these hydrogeological conditions, existing deep learning detectors have largely been validated on firm or well-drained ground (e.g., in Seoul, Shenzhen, or airport runways), where signal-to-noise ratios are significantly higher. Consequently, methods validated on favorable ground cannot be assumed to transfer to high-attenuation deltaic environments. Furthermore, current automated screening often relies on single-view B-scan analysis or requires dedicated 3D volumetric arrays with custom fusion networks, limiting practical deployment using standard commercial GPR software outputs.

Methodology
The paper proposes TriView-YOLO, a multi-view object detection framework based on the YOLOv12 architecture, specifically designed for road cavity screening in adverse ground conditions.

Multi-View Input Fusion: The core innovation is the fusion of three co-registered GPR views into a single 9-channel input tensor:
1. Longitudinal B-scan (depth vs. distance).
2. Horizontal C-scan (plan view amplitude at a specific depth).
3. Cross-section B-scan (orthogonal depth vs. distance).
  These views are derived from standard vehicle-mounted multichannel 3D GPR systems.
TripleInputConv Stem: The standard YOLOv12 stem is replaced by a TripleInputConv layer. This layer splits the 9-channel input into three separate branches, each processing one view with independent convolutional filters (to account for the distinct geometric and statistical nature of B-scans vs. C-scans). The branch outputs are concatenated and projected back to the standard channel width via a $1 \times 1$ convolution. This design allows the network to fuse evidence at the earliest stage while preserving compatibility with pretrained weights and requiring bounding box annotations only on the primary longitudinal view.
Data Strategy: The model was trained and evaluated on a dataset of 1,600 expert-verified field samples, primarily from Bangkok surveys (soft marine clay, 80–140% water content). A smaller fraction of training data came from firmer Japanese subgrades to introduce low-attenuation characteristics, but the test set consisted exclusively of unaugmented Bangkok field data. The study explicitly tested and rejected the use of public datasets (Roboflow) and synthetic data (gprMax) for this specific environment, finding that they degraded performance due to domain shift.
Ablation Protocol: The study employed a rigorous ablation study using three random seeds to evaluate the impact of input views, training data sources (field-only vs. multi-source), model scale (nano, small, medium), initialization (random vs. COCO-pretrained), and feature enhancement (DINOv3 integration).

Key Results
Evaluated on the challenging Bangkok test set (soft marine clay), the proposed TriView-YOLO (small scale) achieved:

mAP50: $0.558 \pm 0.028$
Recall: $0.629 \pm 0.050$
Precision: $0.581 \pm 0.009$
Efficiency: 23.6 GFLOPs and 3.1 ms inference latency per image.

Comparative findings include:

Multi-View Advantage: Removing the auxiliary views (C-scan and cross-section) reduced the best single-view model's mAP50 to $0.497$ and recall to $0.469$, demonstrating that multi-view fusion is critical for recovering weak signatures in high-attenuation soil.
Data Source Sensitivity: Adding public and synthetic images to the training set lowered performance (mAP50 dropped to $0.442$), indicating that cleaner auxiliary data biases the decision boundary away from the noisy reality of wet-clay field data.
Transfer Learning Limitations: COCO pretraining and DINOv3 feature injection (at P4 or P5 levels) did not provide consistent gains and, in some cases, degraded performance, suggesting that natural image features do not transfer well to GPR radargrams.
Scale: Increasing model size from "small" to "medium" did not improve accuracy, suggesting the dataset size, rather than model capacity, is the limiting factor.

Significance and Claims
The paper positions TriView-YOLO not as a solution that achieves high absolute accuracy comparable to studies on firm ground, but as a realistic baseline for automated screening in the most adverse, high-risk environments where cavities are most likely to form.

Realistic Evaluation: The authors emphasize that the reported mAP50 of ~0.56 is significantly lower than the 0.87–0.96 range reported in previous literature, attributing this difference to the extreme signal attenuation in Bangkok's soft clay and the use of an unaugmented, field-only test set. They argue that procurement criteria based on favorable-ground studies would incorrectly reject systems capable of operating in these difficult conditions.
Operational Utility: The system is designed as a screening aid where recall is prioritized over precision. The authors note that at the optimal operating point, roughly 40% of flagged locations are false alarms, which is an acceptable trade-off given that a missed cavity represents a latent safety hazard.
Practical Deployment: The architecture is significant because it operates on standard image exports from commercial GPR software (requiring no raw volumetric data or custom 3D arrays) and requires annotations on only one view, making it adaptable for existing multichannel survey workflows.
Domain Specificity: The study concludes that for high-attenuation environments, "matched" auxiliary data (e.g., simulations of conductive, high-moisture soils) is more valuable than generic or synthetic data, and that future progress depends on domain-specific pretraining rather than natural image transfer.

Technical Summary: TriView-YOLO for GPR Cavity Detection in Soft, High-Water-Content Soils

More like this