Classification of Buried Objects From Ground Penetrating Radar Images by Using Second-Order Deep Learning Models

Examples of preprocessed GPR images with the 200 MHz antenna in wet sand (after direct wave suppression and histogram correction thanks to GPRpys.

Novel AI System Achieves Breakthrough in Ground-Penetrating Radar Object Classification

Researchers at the University Savoie Mont-Blanc and Geolithe have developed a groundbreaking artificial intelligence system that significantly improves the ability to identify buried objects using ground-penetrating radar (GPR) technology. The new approach, which uses advanced "second-order deep learning models," demonstrates superior accuracy and reliability compared to conventional methods, particularly when working with limited training data.

The research team, led by Dr. Guillaume Ginolhac and including graduate student Douba Jafuno, created a system that can distinguish between different types of buried objects - such as metal objects, non-metallic objects, and wooden structures - by analyzing their radar signatures. Their method proved especially robust when dealing with challenging real-world conditions like varying soil types and weather conditions.

This advancement has significant implications across multiple fields:

• - In archaeology, the system could help researchers more accurately identify and classify buried artifacts and structures without the need for excavation, potentially revolutionizing site surveys and preservation efforts.
• - For civil engineering, the technology offers improved capabilities for inspecting underground infrastructure like pipes, cables, and structural foundations, helping to prevent accidents and reduce maintenance costs.
• - In military and humanitarian applications, the system could enhance the detection and classification of buried explosives, potentially making mine clearance operations safer and more efficient.

The research, published in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, demonstrates particularly strong performance in situations where traditional AI systems struggle, such as when dealing with limited training data or when radar conditions vary between training and deployment environments.

"What makes this system particularly valuable is its ability to maintain high accuracy even with limited training data," explains Dr. Ginolhac. "This means it can be deployed more quickly and in new environments where extensive training data might not be available."

The project, supported by the University Savoie Mont-Blanc and CNRS/IN2P3 MUST Computing Center, represents a significant step forward in making GPR technology more reliable and practical for real-world applications. The researchers have made their implementation available through an open-source repository, enabling further development and application of the technology.

Performance and Training

Here are the key quantitative performance metrics and training requirements:

Dataset & Training Requirements:

• - Total dataset: 1,584 radar image thumbnails across 4 categories (metallic, non-metallic, wooden shelters, and empty)
• - Training set: 1,108 images (70%)
• - Validation set: 238 images (15%)
• - Test set: 238 images (15%)

Performance Metrics:

• 1. Baseline Training Data Requirements:
• - The new system (RCNet) achieved ~80% accuracy with just 60% of the full training dataset
• - Conventional CNN approaches required the full dataset to achieve similar performance
• - The system maintained >70% accuracy even with only 40% of training data
• 2. Robustness to Data Quality Issues:
• - Maintained >75% accuracy with up to 10% mislabeled training data
• - Performance only dropped to ~70% accuracy even with 20% mislabeled data
• - Traditional CNN approaches dropped to ~40% accuracy with 20% mislabeled data
• 3. Environmental Adaptability:
• Scenario testing showed varying performance across different conditions:
• - Elevation changes (50cm vs 75-100cm height): ~80% accuracy
• - Frequency changes (200MHz vs 350MHz): ~60% accuracy
• - Soil type variations (dry vs wet sand): ~70-75% accuracy
• - Gravel vs dry gravel: ~75-80% accuracy
• 4. Dataset Distribution:

• - Tested across multiple radar frequencies (200MHz and 350MHz)
• - Multiple elevation heights (0, 25, 50, 75, 100, 150 cm)
• - Various soil conditions (wet sand, dry sand, gravel, dry gravel)
• - Total of 699 medium-sized radar scans processed

• 5. Computational Efficiency:
• - Processing time scaled better than traditional deep learning approaches
• - Required less computational resources than full ResNet implementations
• - Maintained real-time processing capabilities for field deployment

These metrics demonstrate significant improvements over existing systems, particularly in scenarios with limited training data or variable field conditions. However, the paper notes that performance can still be impacted by significant changes in radar frequency or soil conditions, suggesting areas for future improvement. 

Artifacts

The researchers have made two code repositories publicly available:

1. For the main SPD (Symmetric Positive Definite) matrix network implementation:
```
https://github.com/ammarmian/anotherspdnet
```

2. For GPR data preprocessing:
```
https://github.com/NSGeophysics/GPRPy
```

The GPRPy repository appears to be an open-source Ground Penetrating Radar processing and visualization software that was used for preprocessing tasks like direct wave suppression and histogram correction.

Key limitations regarding reproducibility:

1. While the code is available, the full training dataset is not publicly accessible, likely due to its sensitive/proprietary nature.

2. The paper does not specify hyperparameter optimization procedures in detail.

3. The exact environmental conditions and GPR hardware configurations would be difficult to replicate precisely.

4. The test scenarios (different soils, elevations) would need to be recreated for true verification.

For independent verification, researchers would need to:

• - Implement the provided code
• - Collect their own GPR dataset using similar hardware
• - Follow the preprocessing pipeline using GPRPy
• - Train and validate the model using their dataset

The available code allows verification of the methodology, but exact performance reproduction would require similar data collection capabilities and environmental conditions.

Reference Citation

 D. Jafuno, A. Mian, G. Ginolhac and N. Stelzenmuller, "Classification of Buried Objects From Ground Penetrating Radar Images by Using Second-Order Deep Learning Models," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 3185-3197, 2025, doi: 10.1109/JSTARS.2024.3524424.

Abstract: In this article, a new classification model based on covariance matrices is built in order to classify buried objects. The inputs of the proposed models are the hyperbola thumbnails obtained with a classical ground penetrating radar (GPR) system. These thumbnails are then inputs to the first layers of a classical CNN, which then produces a covariance matrix using the outputs of the convolutional filters. Next, the covariance matrix is given to a network composed of specific layers to classify symmetric positive definite matrices. We show in a large database that our approach outperform shallow networks designed for GPR data and conventional CNNs typically used in computer vision applications, particularly when the number of training data decreases and in the presence of mislabeled data. We also illustrate the interest of our models when training data and test sets are obtained from different weather modes or considerations.

keywords: {Buried object detection;Covariance matrices;Shape;Ground penetrating radar;Soil;Robustness;Radar imaging;Radar antennas;Permittivity;Symmetric matrices;Buried objects classification;covariance matrices;ground penetrating radar;symmetric positive definite matrix networks},

URL: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10836936&isnumber=10766875

Background of the study:
The researchers are studying how to classify buried objects using ground penetrating radar (GPR) images. GPR images can be very noisy and the shape of the hyperbolas (the visual patterns in the images) depends on various factors beyond just the buried object, making classification challenging.

Research objectives and hypotheses:
The researchers aim to develop a high-performance and robust classification model that can work well even with limited training data and in the presence of mislabeled data or differences between the training and test sets. They propose using second-order deep learning models, which capture the correlations between the outputs of the convolutional layers, as a way to achieve this goal.

Methodology:
The researchers first extract thumbnails of the hyperbolas from the GPR images. They then use these thumbnails as inputs to two proposed models:
1. SRCNet: Stacks the outputs from the first few layers of a ResNet-34 model to create a tensor, and then uses covariance pooling and specific layers for symmetric positive definite matrices to classify.
2. RCNet: Uses only the outputs from the last layer of ResNet-34 to create the tensor, and then applies the same covariance pooling and SPD matrix layers.

These models are compared to shallow CNN models, deep ResNet-34 models (both fine-tuned and retrained), as well as traditional machine learning algorithms like support vector machines and random forests.

Results and findings:
The proposed SRCNet and RCNet models outperform the other approaches, especially when the amount of training data is limited or when there is mislabeled data in the training set. RCNet in particular shows good performance while being more computationally efficient than SRCNet.

Discussion and interpretation:
The researchers attribute the good performance of the second-order models to their ability to capture the correlations between the convolutional features, which helps deal with the noisy and variable nature of the GPR images. The covariance pooling and SPD matrix layers also seem to provide robustness.

Contributions to the field:
The study demonstrates the benefits of using second-order deep learning models for classifying GPR images, which is a novel approach in this domain. The proposed architectures could be useful for other applications dealing with noisy sensor data and limited training samples.

Achievements and significance:
The researchers were able to develop classification models that outperform standard approaches, particularly in challenging scenarios with limited data or mislabeled samples. This represents an important advance in the field of GPR-based object classification.

Limitations and future work:
The study is limited to a specific dataset provided by Geolithe. Future work could involve testing the models on a wider range of GPR data and exploring ways to further improve the computational efficiency of the approaches.