🤖 AI

Automatic Road Subsurface Distress Recognition from Ground Penetrating Radar Images using Deep Learning-based Cross-verification

This paper proposes a novel cross-verification strategy using three YOLO-based models trained on different views of 3D Ground Penetrating Radar images to achieve over 98.6% recall in automatically detecting road subsurface distress, thereby reducing human inspection labor by approximately 90%.

Original authors: Chang Peng, Bao Yang, Meiqi Li, Ge Zhang, Hui Sun, Zhenyu Jiang

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chang Peng, Bao Yang, Meiqi Li, Ge Zhang, Hui Sun, Zhenyu Jiang

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 the road you drive on every day is like a giant, multi-layered cake. Sometimes, the layers underneath get soggy, crumbly, or even develop empty holes (voids) that you can't see from the top. If you don't find these weak spots, the whole cake could collapse under a car, causing a dangerous pothole or even a sinkhole.

For a long time, finding these hidden problems has been like trying to find a needle in a haystack while wearing blindfolded gloves. Experts use a machine called Ground Penetrating Radar (GPR) to "see" underground. It shoots radio waves into the ground and listens for echoes. But the data it spits out looks like a confusing mess of static and squiggly lines. Traditionally, a human expert has to stare at thousands of these images, trying to guess which squiggle means "danger" and which means "safe." It's slow, exhausting, and prone to human error.

This paper introduces a new, super-smart way to automate this process using Artificial Intelligence (AI). Here is how they did it, explained simply:

1. The Problem: One Angle Isn't Enough

Imagine trying to identify a mysterious object in a dark room.

If you only look at it from the front, it might look like a flat circle.
If you look from the side, it might look like a tall rectangle.
If you look from above, it might look like a square.

The researchers realized that GPR data comes in three different "views" (called B-scan, C-scan, and D-scan). A specific type of road damage (like a hollow void) looks very different depending on which angle you view it from. Some AI models are great at spotting damage from the "top view" but terrible at the "side view," and vice versa.

2. The Solution: The "Three-Headed Detective" Team

Instead of building one giant, super-complex AI that tries to do everything (which often gets confused), the researchers built a team of three specialized AI detectives, each trained on a different camera angle:

Detective C (The Big Picture): This AI looks at the horizontal "map" view. It's really good at spotting something is wrong (like a loose patch of road), but it sometimes mistakes a manhole cover for a hole.
Detective B (The Side-Profile): This AI looks at the vertical cross-section. It's excellent at distinguishing a manhole cover from actual road damage.
Detective D (The Transverse View): This AI looks at the cross-section from the other side. It's the expert at telling the difference between a hollow void (a big empty space) and loose soil (just crumbly dirt).

3. The "Cross-Verification" Strategy: The Filter System

Here is the clever part. They didn't just let the three detectives shout out their guesses. They created a three-step filter system:

Step 1 (The Sweep): Detective C scans the whole road. If it says, "This area looks healthy," we ignore it. If it says, "Something is weird here," we pass it to the next step.
Step 2 (The Manhole Check): The suspicious spots go to Detective B. If B says, "That's just a manhole cover," we throw it out. If B says, "No, that's definitely damage," we keep it.
Step 3 (The Diagnosis): The remaining "dangerous" spots go to Detective D. D decides: "Is this a hollow void (very dangerous) or just loose soil (needs fixing but less urgent)?"

By having them check each other's work, they cancel out their mistakes. If one detective is confused, the other two usually catch it.

4. The Results: Faster, Smarter, Safer

The team tested this on over 1,250 kilometers of real roads in China.

Accuracy: The system found 98.6% of all the road problems. In a real-world test on 15 new roads, it found 100% of the problems. It didn't miss a single one.
Speed: This is the biggest win.
- Old Way: A human expert takes about 1 hour to check just 1 kilometer of road.
- New Way: The computer does the heavy lifting in seconds. The human only needs to spend a few minutes double-checking the computer's "suspects."
- Result: This cuts the human workload by 90%. It's like going from manually sorting a pile of mail by hand to using a high-speed sorter that just needs a quick glance to catch the weird letters.

Why This Matters

The researchers admit the system is a bit "paranoid." It sometimes flags healthy spots as dangerous (false alarms) just to be absolutely sure it doesn't miss a real problem. But that's a good thing! It's better to check a few extra spots than to miss a sinkhole.

In a nutshell: This paper teaches us that instead of trying to build one "super-brain" AI to solve a complex problem, it's often better to build a team of specialized experts who check each other's work. This approach makes road inspections faster, cheaper, and much safer for everyone driving on the road.

1. Problem Statement

Road Subsurface Distress (RSD), including voids, loose structures, and interlayer debonding, poses significant risks to road safety and infrastructure integrity. While Ground Penetrating Radar (GPR) is the standard non-destructive tool for detecting these issues, current practices face critical limitations:

Manual Bottleneck: Interpreting GPR images (A-scan, B-scan, C-scan, D-scan) is labor-intensive, time-consuming (approx. 1 hour per km), and relies heavily on expert intuition, leading to subjectivity.
Limitations of Existing AI: Traditional rule-based methods fail in complex real-world environments due to signal distortion. While deep learning (DL) has improved automation, existing models often suffer from:
- Insufficient Recall: Many models fail to detect all defects, missing critical safety hazards.
- Data Scarcity & Diversity: Training datasets are often small or lack diversity in soil types and distress categories.
- Single-View Bias: Most approaches rely on a single view (usually B-scan) or simple feature concatenation, failing to fully exploit the complementary information available in 3D GPR data.

2. Methodology

The authors propose a Multi-View Cross-Verification Strategy utilizing three independent YOLO-based object detection models trained on different 2D slices of 3D GPR data.

A. Dataset Construction

Source: Field data collected from 105 urban road sections (1,250 km total) in Chengdu and Guangzhou, China.
Diversity: Covers four landforms (alluvial plains, fluvial terraces, hills, mountains) and five soil types with varying geotechnical properties.
Annotations: A rigorously validated dataset of 2,134 samples (553 healthy, 539 voids, 536 loose structures, 506 manholes) annotated by experts and verified via core sampling and Falling Weight Deflectometer (FWD) tests.
Views: The 3D GPR data was sliced into three orthogonal planes:
- B-scan: Longitudinal section.
- C-scan: Horizontal section.
- D-scan: Transverse section.

B. Model Architecture

Three separate YOLO-based (anchor-free) models were trained independently:

Model-C: Trained on C-scan images.
Model-B: Trained on B-scan images.
Model-D: Trained on D-scan images.

Architecture: Utilizes a Backbone (Focus, CBS, CSP, SPP), Neck (feature fusion), and Head (decoupled heads for regression, objectness, and classification).
Training: Conducted with data augmentation (Mosaic, MixUp, flipping, rotation) over 400 epochs using AdamW optimizer.

C. The Cross-Verification Strategy

Instead of fusing features early, the strategy employs a sequential, three-step filtering process to leverage the unique sensitivity of each view:

Step 1 (Sifting Healthy Areas): Model-C (trained on C-scan) is used first. Since C-scans clearly show the shape of voids and loose structures as fragmented speckles or irregular patches, this model acts as the primary filter to identify potential distress regions and discard healthy segments.
Step 2 (Filtering Manholes): Model-B (trained on B-scan) processes the remaining candidates. B-scans provide reliable patterns to distinguish manholes (regular ripples) from actual distress, effectively removing false positives caused by manholes.
Step 3 (Classifying Distress Type): Model-D (trained on D-scan) classifies the remaining defects. D-scans help differentiate between voids and loose structures, which often look similar in other views.

3. Key Contributions

Novel Cross-Verification Framework: Moves beyond simple multi-view concatenation to a sequential logic that exploits the specific strengths of B, C, and D scans for different classification tasks (detection $\to$ filtering $\to$ classification).
Large-Scale, High-Quality Dataset: Creation of a diverse, rigorously validated dataset of over 2,000 field samples, addressing the common issue of small or synthetic datasets in GPR research.
High-Recall Focus: The system is explicitly designed to prioritize Recall (minimizing false negatives) to ensure no potential road collapse risks are missed, accepting a higher rate of false positives that can be easily filtered by humans.

4. Results

A. Performance on Testing Subset

The cross-verification strategy significantly outperformed individual models and existing literature:

Recall: Achieved 98.6% for RSD detection (vs. ~58–92% in recent state-of-the-art studies).
Precision: Achieved 95.9%.
Specific Breakdown (Step 3):
- Void Recall: 84.1%
- Loose Structure Recall: 86.0%
- Manhole filtering: 100% recall (no manholes missed as distress).

B. Field Verification (Real-World Application)

Tests were conducted on 15 new roads (69.4 km) in Zhuhai, Guangzhou, and Yunfu.

Detection Rate: The method achieved 100% recall, successfully identifying all 75 manually detected distress cases without missing any.
False Positives: 127 false positives were generated (mostly healthy areas or manholes misidentified as distress). However, the authors argue this is a feature, not a bug, as it flags high-risk sections for closer inspection.
Efficiency:
- Manual Processing: 1,128 minutes (~19 hours).
- Automatic Processing + Manual Review: 101 minutes (~1.7 hours).
- Labor Reduction: Approximately 90% reduction in human labor.

5. Significance

Safety & Reliability: By prioritizing recall, the system ensures that no subsurface defects are overlooked, directly addressing the risk of catastrophic road collapses.
Operational Efficiency: The 90% reduction in inspection time makes large-scale, long-term road monitoring economically feasible for municipal authorities.
Scalability: The framework is model-agnostic (independent of the specific object detection backbone), allowing it to serve as a benchmark for future GPR detection algorithms.
Practical Deployment: The method bridges the gap between academic deep learning research and engineering practice by using a workflow that integrates automated detection with efficient human verification, rather than attempting fully autonomous "black box" decision-making.