Interpreting Electrical-Resistivity Tomography measurements using Neural Network

Itay Naeh, Yitzhak Peleg, Alex Furman, Shie MannorarXiv 2018 · DOI · arXiv

Abstract

Electrical Resistivity Tomography (ERT) has been extensively used for imaging the subsurface resistivity distribution and structure. Over the years, many algorithms have been developed in order to solve the subsurface resistivity distribution from the ERT measurements. In this paper a new method for interpreting the ERT measurements is presented. Using supervised learning to train a neural network, we are able to interpret the ERT measurement into a 2D image of the underground resistivity up to depths of 50 meters while using a simple Wenner-Schlumberger survey of 96 electrodes with 1 meter spacing. The neural network is trained and tested using simulative data and it is shown to have superior results over a well established inversion method.

Introduction

Mapping subsurface structures without excavation is critical for fields ranging from civil engineering to archaeology. Electrical Resistivity Tomography (ERT) is a primary tool for this purpose, where electrodes inject current and measure voltage to infer soil resistivity. However, converting these measurements into a 2D map—the "inverse problem"—is notoriously difficult. It is mathematically ill-posed, as the number of unknown resistivity values far exceeds the available measurements. Traditional iterative inversion methods are computationally intensive, sensitive to initial guesses, and often struggle to resolve deep features.

Methodology

This study proposes a supervised deep learning approach to bypass the limitations of iterative physics-based optimization.

Data Acquisition and Simulation

Due to the difficulty of obtaining large-scale real-world ground truth data, the researchers utilized 15,000 simulated ground profiles. Each profile consisted of a 120m x 50m cross-section with a linear resistivity gradient perturbed by 100 random 2D Gaussian distributions to mimic natural soil inhomogeneity. Forward modeling was performed using the R2 software to generate 1,488 ERT measurements based on a 96-electrode Wenner-Schlumberger survey.

Neural Network Architecture

The proposed model is a fully connected neural network featuring a single hidden layer of 250 neurons with ReLU activation. The input is the 1,488-element measurement vector, and the output is a 6,000-element vector reshaped into a 120x50 pixel resistivity map. The network was trained using the Adam optimizer and $L_1$ loss (Mean Absolute Error), which proved more effective than $L_2$ loss at preserving feature boundaries and handling outliers.

Results and Discussion

Quantitative Performance

The neural network was evaluated against the standard R2 inversion method on 1,500 unseen test cases. The results demonstrated a significant performance advantage for the deep learning approach.

Distribution of L1 loss and L1 distance comparing the Neural Network (NN) to the traditional R2 inversion method.

As shown in the histograms, the NN consistently achieves lower $L_1$ loss and distance across the test set. Even the "worst-case" scenarios for the NN often outperformed the average performance of the traditional iterative solver.

Qualitative Comparison

Beyond numerical metrics, the visual fidelity of the reconstructed images highlights the NN's ability to resolve deep features that traditional methods often miss or smear.

Visual comparison of the NN prediction, ground truth, and R2 prediction for a sample subsurface profile.

The NN accurately captures the location and geometry of resistivity anomalies up to 50 meters deep. In contrast, the classical method frequently produces artifacts or fails to define the boundaries of deep-seated features. Furthermore, while traditional inversion can take minutes, the NN generates these predictions in milliseconds, enabling real-time field interpretation.

Conclusion

This research demonstrates that a relatively simple neural network can outperform complex iterative solvers for ERT interpretation. By learning the non-linear mapping between surface measurements and subsurface structure, the model provides faster, more accurate, and deeper imaging. Future work will focus on bridging the "simulation gap" by incorporating realistic noise and adapting the architecture to variable electrode configurations.

Summary updated

Abstract

Electrical Resistivity Tomography (ERT) has been extensively used for imaging the subsurface resistivity distribution and structure. Over the years, many algorithms have been developed in order to solve the subsurface resistivity distribution from the ERT measurements. In this paper a new method for interpreting the ERT measurements is presented. Using supervised learning to train a neural network, we are able to interpret the ERT measurement into a 2D image of the underground resistivity up to depths of 50 meters while using a simple Wenner-Schlumberger survey of 96 electrodes with 1 meter spacing. The neural network is trained and tested using simulative data and it is shown to have superior results over a well established inversion method.
BibTeX
@misc{naeh2018interpreting,
  author = {Itay Naeh and Yitzhak Peleg and Alex Furman and Shie Mannor},
  title = {{Interpreting Electrical-Resistivity Tomography measurements using Neural Network}},
  year = {2018},
  eprint = {1805.03698},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1805.03698},
}
Figure