Investigating the Capabilities of Deep Learning for Processing and Interpreting One-Shot Multi-offset GPR Data: A Numerical Case Study for Lunar and Martian Environments

TL;DR

The potential of deep learning for interpreting and processing GPR data is investigated via a coherent numerical case study, showcasing the potential of deep learning for reconstructing the dielectric distribution of the the near surface of Terrestrial planets, and filling missing or bad-quality traces.

Abstract

Ground-penetrating radar (GPR) is a mature geophysical method that has gained increasing popularity in planetary science over the past decade. GPR has been utilised both for Lunar and Martian missions providing pivotal information regarding the near surface geology of Terrestrial planets. Within that context, numerous processing pipelines have been suggested to address the unique challenges present in planetary setups. These processing pipelines often require manual tuning resulting to ambiguous outputs open to non-unique interpretations. These pitfalls combined with the large number of planetary GPR data (kilometers in magnitude), highlight the necessity for automatic, objective and advanced processing and interpretation schemes. The current paper investigates the potential of deep learning for interpreting and processing GPR data. The one-shot multi-offset configuration is investigated via a coherent numerical case study, showcasing the potential of deep learning for A) reconstructing the dielectric distribution of the the near surface of Terrestrial planets, and B) filling missing or bad-quality traces. Special care was taken for the numerical data to be both realistic and challenging. Moreover, the generated synthetic data are properly labelled and made publicly available for training future data-driven pipelines and contributing towards developing pre-trained foundation models for GPR.

Figures (6)

• Figure 1: A) A generic one-shot multi-offset numerical case study. The discretisation of the model is 2 cm, and the central frequency of the source is 80 MHz. B) Processed B-Scan. C) NMO for picking velocities. D) The mean permittivity profile (black) and the reconstructed one (red) using Dix conversion Giannakis:2022_AGU_Abstract.
• Figure 2: A set of training data from the 4400 samples used in this study. The colorbar illustrates the relative electric permittivity. Left figures depict the numerical models, and right figures are their corresponding B-Scans. A time-varying gain and signal saturation is applied to the B-Scans for illustration purposes.
• Figure 3: The architecture of the ensemble U-nets used for FWI. The input B-Scans have 230x230 dimensions each, and the output models have 224x224.
• Figure 4: A set of examples comparing the ground truth permittivity models to the reconstructed ones using the suggested ensemble U-nets. These are unknown models that were not included in the training process. Colourbars depict the relative permittivity values. It is apparent that the proposed ensemble approach has the capability to reconstruct a smooth representation of the subsurface.
• Figure 5: The architecture of the U-net used for filling missing data. The inputs are the corrupted B-Scans each having 230x230 dimensions, and the outputs are the initial B-Scans.