⚡ electrical engineering

A Migration-Assisted Deep Learning Scheme for Imaging Defects Inside Cylindrical Structures via GPR: A Case Study for Tree Trunks

This paper proposes a three-stage migration-assisted deep learning scheme that integrates dual-permittivity estimation, modified Kirchhoff migration, and shape reconstruction to accurately image the shape and permittivity of defects inside cylindrical structures like tree trunks using ground-penetrating radar, demonstrating superior performance and in-field feasibility through synthetic and experimental validation.

Original authors: Jiwei Qian, Yee Hui Lee, Kaixuan Cheng, Qiqi Dai, Arda Yalcinkaya, Mohamed Lokman Mohd Yusof, James Wang, Abdulkadir C. Yucel

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Jiwei Qian, Yee Hui Lee, Kaixuan Cheng, Qiqi Dai, Arda Yalcinkaya, Mohamed Lokman Mohd Yusof, James Wang, Abdulkadir C. Yucel

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ 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 have a tree trunk, and you suspect there is a hollow, rotting cavity hidden deep inside. You can't just cut it open to check, because that might kill the tree or ruin a historic column. You need a way to "see" inside without touching it.

This paper presents a new, high-tech "X-ray vision" system for trees and cylindrical structures (like utility poles) using Ground-Penetrating Radar (GPR). However, looking inside a round tree is much harder than looking inside a flat wall. Here is how the authors solved this problem using a clever mix of physics and artificial intelligence (AI).

The Problem: The "Round Wall" Confusion

When you scan a flat wall with radar, the signal bounces back in a predictable way, like a ball hitting a flat floor. But when you scan a cylinder (like a tree trunk), the radar waves bounce off the curved surface and the internal layers in a messy, confusing pattern.

Think of it like trying to hear a whisper in a round, echoey bathroom versus a quiet hallway. In the bathroom (the tree), the sound bounces around the walls, creating echoes that make it hard to tell exactly where the whisper is coming from or what shape the object is. Traditional radar tools get confused by these "echoes" and often produce blurry, inaccurate pictures of the rot.

The Solution: A Three-Step "Detective" Team

The authors created a three-stage system that acts like a team of detectives working together to solve the mystery of the hidden defect.

Step 1: The "Material Detective" (DPE-Net)

First, the system needs to know what the tree is made of. Is it dry wood? Wet wood? Is the rot dry or soggy?

What it does: This AI network looks at the raw, messy radar data and guesses two things: the "density" (permittivity) of the healthy wood and the "density" of the rot.
The Trick: To teach this AI, the researchers invented a new way to check its work. Instead of just asking "is the image sharp?" (which can be misleading), they asked, "Does the shape of the blur match the shape of a ring?" They used a tool called SSIM (Structural Similarity Index) to ensure the AI learned to pick the right material properties that create the most accurate shape, not just the sharpest image.

Step 2: The "Time-Traveler" (Modified Migration)

Once the AI knows the material properties, it needs to fix the messy radar picture.

What it does: This step uses a physics-based algorithm called Kirchhoff Migration. Imagine the radar signals are like ripples in a pond. If you drop a stone, the ripples spread out. This algorithm works backward: it takes the spreading ripples and "rewinds" them to show exactly where the stone (the defect) was dropped.
The Innovation: Usually, this "rewinding" process has to be tried over and over with different guesses until the picture looks right. Because the first AI step already guessed the material properties, this step only has to run once. It's like having a GPS that knows the road conditions perfectly, so you don't have to drive around in circles to find the right route.

Step 3: The "Image Cleaner" (SR-Net)

The "rewound" image is better, but it's still a bit noisy and cluttered with leftover echoes.

What it does: A second AI network (a specialized version of a U-Net) takes this cleaned-up image and acts like a photo editor. It erases the background noise and sharpens the edges to reveal the exact shape and location of the rot.
The Secret Sauce: This network uses special "ResPaths" (shortcuts) to make sure it doesn't lose the fine details while cleaning up the image, ensuring the final picture of the rot is crisp and accurate.

The Results: From Simulation to Real Trees

The researchers tested this system in three ways:

Computer Simulations: They created thousands of fake tree trunks with fake rot. The system was much better at finding the rot than existing AI methods.
Lab Models: They built a fake tree trunk out of a bucket of sand and buried soil samples with different moisture levels inside. The system successfully identified the shape and moisture content of the "rot."
Real Trees: They took the system to actual trees in Singapore. They compared their results to other methods like sonic tomography (which uses sound waves) and drilling (which pokes holes in the tree).
- The Verdict: The new system found the rot more accurately and with a clearer shape than the sonic tomography, which often produced blurry, oversized blobs. It was also non-invasive, unlike drilling.

Why This Matters

This paper doesn't claim to cure trees or predict exactly when a tree will fall. Instead, it claims to provide a faster, more accurate, and non-invasive way to see inside cylindrical objects.

By combining a physics-based "rewinding" technique with smart AI, the authors created a tool that can:

Tell you where a defect is.
Show you the shape of the defect.
Estimate the material properties (like moisture content) of the defect.

This is a significant step forward for arborists and engineers who need to inspect trees, utility poles, and historical columns without damaging them. The code and data are even made available for others to use and improve upon.

Technical Summary: A Migration-Assisted Deep Learning Scheme for Imaging Defects Inside Cylindrical Structures via GPR

Problem Statement
Ground-penetrating radar (GPR) is a critical tool for detecting internal defects in cylindrical structures such as tree trunks, utility poles, and cultural heritage columns. However, imaging these defects presents significant challenges that render traditional signal processing and existing deep learning (DL) techniques ineffective.

Geometric Mismatch: Conventional DL models (e.g., MRF-UNet, PiNet) are trained on flat-surface B-scans where reflection hyperbolas align spatially with target locations. In circumferential scanning of cylindrical structures, reflections form sinusoidal patterns with no direct spatial correspondence to the defect's location, causing severe performance degradation in standard models.
Clutter and Multiple Reflections: Internal layers and bark surfaces generate complex multiple reflections and environmental noise, creating clutter that obscures defect signatures.
Permittivity and Shape Reconstruction: Existing methods struggle to simultaneously reconstruct the precise shape and dielectric permittivity of defects. Migration-based algorithms often require iterative trials over a range of wave velocities to find the optimal focusing, which is computationally expensive. Furthermore, conventional entropy-based autofocusing techniques (AFTs) often yield geometrically distorted images, failing to preserve the true defect shape.
Limitations of Current DL Approaches: Hybrid approaches that use migrated images as auxiliary inputs suffer from the same spatial misalignment issues and rely on suboptimal migration results generated by entropy-based AFTs.

Methodology
The authors propose a three-stage, migration-assisted deep learning framework designed specifically for circumferential GPR scanning of cylindrical objects. The scheme integrates physics-based signal processing with specialized neural networks to achieve near-real-time reconstruction of defect shape and permittivity.

Stage 1: Dual-Permittivity Estimation (DPE-Net)
- Objective: To simultaneously estimate the equivalent permittivity of the host cylindrical medium ( $\varepsilon_{r,medium}$ ) and the permittivity of the internal defect ( $\varepsilon_{r,defect}$ ) directly from the raw B-scan.
- Architecture: The network utilizes a Feature Extraction Module (FEM) based on Residual Blocks, a Feature Aggregation Module (FAM) to combine multi-level features, and a Convolutional Block Attention Module (CBAM) to refine feature representation.
- Training Strategy: To generate ground truth labels for $\varepsilon_{r,medium}$ during training, a novel Structural Similarity Index Measure (SSIM)-based Autofocusing Technique (AFT) is employed. Unlike entropy-based AFTs that favor energy concentration (often leading to geometric distortion), the SSIM-based AFT selects the migration velocity that maximizes the structural similarity between the migrated image and a ring mask representing the defect boundary. This ensures the selected permittivity preserves the defect's geometric fidelity.
Stage 2: Modified Kirchhoff Migration
- Objective: To map the reflected signals from the B-scan domain to the spatial imaging domain.
- Process: Using the $\varepsilon_{r,medium}$ predicted by DPE-Net, a modified Kirchhoff migration algorithm is executed once (one-shot migration). This maps the signals to the defect boundaries, significantly reducing computational cost compared to iterative velocity searches. The algorithm utilizes an exploding source model (ESM) to approximate the two-way propagation as a one-way radiation problem.
Stage 3: Shape Reconstruction (SR-Net)
- Objective: To process the cluttered migrated image and reconstruct the precise geometry of the defect.
- Architecture: An enhanced U-Net architecture featuring ResPath connections between the encoder and decoder to bridge the semantic gap between low-level and high-level features. It also employs a CBAM-based decoder to dynamically weigh feature contributions.
- Output: The network suppresses noise and clutter, outputting a clean binary mask of the defect's shape.

Final Reconstruction: The final defect map is generated by combining the shape output from SR-Net with the permittivity value predicted by DPE-Net.

Key Contributions

Migration-Assisted Framework: The paper introduces a novel scheme that resolves the domain-mismatch bottleneck in cylindrical scanning by decoupling the tasks: permittivity estimation is performed in the B-scan domain (where material-dependent reflections are retained), while shape recovery is performed in the migrated-image domain (where spatial geometry is correctly aligned).
SSIM-Based Autofocusing: A new AFT is developed that prioritizes geometric fidelity over energy concentration. This enables a reliable "one-shot" migration strategy, eliminating the need for repetitive velocity trials and providing accurate supervisory labels for training the permittivity estimation network.
Domain-Specialized Networks: The design of DPE-Net (with FAM and CBAM) and SR-Net (with ResPath and attention mechanisms) specifically addresses the challenges of feature aggregation and semantic gaps in cylindrical defect imaging.
Comprehensive Validation: The framework is validated across three distinct datasets: synthetic layered cylinders, a laboratory trunk model with soil-based decay, and real tree trunk samples (both fresh-cut and live trees).

Results

Synthetic Data: The proposed scheme outperformed existing DL methods (PiNet, MRF-UNet, MultiPath-Net) in all metrics. It achieved an Intersection over Union (IoU) of 0.8906 and an SSIM of 0.9634, compared to significantly lower scores for baselines. The SSIM-based AFT reduced the Mean Absolute Error (MAE) of permittivity estimation to 0.24, compared to 2.72 for entropy-based AFT.
Laboratory Model: On measurement data from a sand-filled bucket model with artificial decay, the proposed method maintained superior performance with an IoU of 0.7947, accurately reconstructing irregular shapes (triangles, trapezoids) that other methods blurred or distorted.
Real Tree Trunks: In tests on 21 real tree trunk samples with simulated defects, the proposed scheme achieved the highest average IoU (54.74%) and demonstrated robust generalization.
Generalization (Zero-Shot & Live Trees):
- Zero-Shot: Models trained solely on synthetic data successfully generalized to real measurement data without fine-tuning, achieving an IoU of 38.72%, significantly outperforming baselines.
- Live Trees: Generalization tests on live trees with naturally developed defects (validated via sonic tomography and drilling resistance) confirmed the method's ability to localize defects accurately. While minor intensity variations occurred due to biological variability, the method consistently provided interpretable reconstructions where traditional migration and sonic tomography failed to resolve defect boundaries.
Efficiency: The total CPU time for the proposed framework was approximately 20 seconds per sample, offering a favorable trade-off between accuracy and speed compared to iterative migration methods (which took ~89 seconds).

Significance and Claims
The paper claims to present the first deployable, physics-consistent deep learning inversion framework tailored for GPR-based internal defect imaging in cylindrical objects. Its significance lies in:

Overcoming Domain Mismatch: Successfully addressing the spatial misalignment inherent in circumferential scanning, a limitation that has hindered the application of standard DL models to tree trunks and columns.
Physics-Guided Automation: Integrating a physics-informed migration step with deep learning to automate the selection of migration parameters without prior structural knowledge, enabling fully automated imaging.
Practical Deployability: Demonstrating feasibility for in-field deployment through successful testing on live trees, offering a non-invasive tool for assessing structural integrity and preventing tree falls.
Broad Applicability: While focused on tree trunks, the underlying principle is applicable to other circumferential GPR imaging scenarios, such as utility poles and cultural heritage columns.

The authors acknowledge limitations regarding the complexity of natural decay, moisture variability, and non-convex geometries, noting that future work will focus on expanding datasets and refining signal processing for irregular shapes. However, the current results establish a robust foundation for real-time, accurate defect imaging in cylindrical structures.