Variable data structures and customized deep learning surrogates for computationally efficient and reliable...

Area of Science:

• Geophysics
• Artificial Intelligence
• Electromagnetics

Background:

• Accurate characterization of buried objects using Ground Penetrating Radar (GPR) is crucial for various applications.
• Traditional methods often involve high computational costs and complex data processing.
• Deep learning offers a promising avenue for developing efficient surrogate models for GPR data analysis.

Purpose of the Study:

• To develop a computationally efficient deep-learning-based surrogate modeling approach for characterizing buried objects using GPR.
• To independently predict characteristic parameters (radius, depth, lateral position) of buried objects in various subsurface media.
• To analyze the trade-offs between computational cost and accuracy for different GPR data structures.

Main Methods:

• Utilized 3-D full-wave electromagnetic simulations of GPR models.
• Employed a deep regression network (DRN) for analyzing time-frequency spectrograms (TFS) of consecutive A-scans.
• Compared the performance of DRN with conventional network models using B-scan images (2D data) and state-of-the-art regression techniques.
• Evaluated the model's robustness using noisy data and validated it with physical measurement data.

Main Results:

• The proposed DRN model achieved approximately 13 times acceleration compared to conventional B-scan image-based models.
• Achieved mean absolute errors of 3.6 mm and relative errors of 4.7% for object characterization.
• Demonstrated competitive performance against state-of-the-art regression techniques.
• Showed suitability for handling noisy data and physical measurement data.

Conclusions:

• The deep regression network (DRN) provides a computationally efficient and accurate surrogate modeling approach for GPR-based buried object characterization.
• The TFS analysis with DRN is effective for predicting object parameters in diverse subsurface conditions.
• The validated approach shows significant potential for real-world applications involving physical GPR measurements.