Elsevier

Journal of Applied Geophysics

Volume 103, April 2014, Pages 172-185
Journal of Applied Geophysics

A modified DOI-based method to statistically estimate the depth of investigation of dc resistivity surveys

https://doi.org/10.1016/j.jappgeo.2014.01.018Get rights and content

Highlights

  • We developed a new method to identify the depth of investigation of ERT surveys.

  • This method is based on a modified DOI approach and is therefore empirical.

  • The mixture of 2 Gaussian distributions is fitted to the distribution of scaled DOI.

  • An interpretation index is defined to identify well-constrained cells of the model.

  • The efficiency of the method is assessed on synthetic and field datasets.

Abstract

Several techniques are available to estimate the depth of investigation or to identify possible artifacts in dc resistivity surveys. Commonly, the depth of investigation (DOI) is mainly estimated by using an arbitrarily chosen cut-off value on a selected indicator (resolution, sensitivity or DOI index). Ranges of cut-off values are recommended in the literature for the different indicators. However, small changes in threshold values may induce strong variations in the estimated depths of investigation. To overcome this problem, we developed a new statistical method to estimate the DOI of dc resistivity surveys based on a modified DOI index approach. This method is composed of 5 successive steps. First, two inversions are performed by using different resistivity reference models for the inversion (0.1 and 10 times the arithmetic mean of the logarithm of the observed apparent resistivity values). Inversion models are extended to the edges of the survey line and to a depth range of three times the pseudodepth of investigation of the largest array spacing used. In step 2, we compute the histogram of a newly defined scaled DOI index. Step 3 consists of the fitting of the mixture of two Gaussian distributions (G1 and G2) to the cumulative distribution function of the scaled DOI index values. Based on this fitting, step 4 focuses on the computation of an interpretation index (II) defined for every cell j of the model as the relative probability density that the cell j belongs to G1, which describes the Gaussian distribution of the cells with a scaled DOI index close to 0.0. In step 5, a new inversion is performed by using a third resistivity reference model (the arithmetic mean of the logarithm of the observed apparent resistivity values). The final electrical resistivity image is produced by using II as alpha blending values allowing the visual discrimination between well-constrained areas and poorly-constrained cells.

The efficiency of the proposed methodology is assessed on synthetic and field data. By using synthetic benchmark analysis, we demonstrate that the selected well-constrained cells are well-reconstructed in size and shape as well as in resistivity contrasts. Compared to the existing image appraisal tools, the proposed statistical method allows the identification of the statistically well-constrained cells of the model without using any arbitrary cut-off value. Using this statistical method in combination with the resolution, when interpreting dc resistivity surveys, provides the geophysicist valuable information to avoid over- or misinterpretation of ERT images.

Introduction

Nowadays, direct current (dc) resistivity surveys are frequently used to map soil electrical properties in two to three dimensions in a wide range of engineering, hydrogeological and risks management applications and to monitor dynamic subsurface processes through time-lapse measurements.

Ill-posed problems including electrical resistivity tomography (ERT) inversion are inherently difficult. The general idea behind data inversion is to find a model that fits to the data given a defined error criterion which is generally estimated from noise estimates (see for example LaBrecque et al., 1996). However, this model is non-unique since an infinite number of models can fit to the data equally. Moreover, ERT inversion is an ill-conditioned inverse problem that leads to instability of the solution. This means that a small change in the data (e.g., a small increase of noise) can lead to severe changes in the resulting model. To overcome the non-uniqueness problem as well as to introduce stability in the inversion process, regularization is added (e.g. Tikhonov and Arsenin, 1977). Regularization consists of adding some model constraints that bias the solution, generally in a way to satisfy available a priori information of the subsurface.

ERT data inversion consists then of finding the model that fits to the data subject to a model constraint such as the widely used smoothness constraint (Constable et al., 1987, deGroot-Hedlin and Constable, 1990). The resulting image must then be appraised in order to find which parts are well resolved and therefore physically interpretable. One aspect of appraising electrical images consists of finding its depth of investigation (DOI).

The depth of investigation is defined by the depth to which the data are particularly sensitive to the physical properties of the subsurface (Oldenburg and Li, 1999) or in other terms, the depth below which the data are no longer influenced by the electrical resistivity, in our case. This means that below this depth, inverted structures are not related to surface data anymore and should be interpreted with caution.

Several resolution indicators can be used to estimate the depth of investigation or to identify possible artifacts in the electrical structures. The model resolution matrix (e.g. Alumbaugh and Newman, 2000, Friedel, 2003, Hilbich et al., 2009, Oldenborger and Routh, 2009, Ramirez et al., 1995) is the approximation for non-linear problem of the resolution matrix defined theoretically for linear problems. Ideally, the matrix should be close to the identity matrix to achieve good resolution. The cumulative sensitivity matrix (e.g. Christiansen and Auken, 2012, Kemna, 2000, Nguyen et al., 2009, Park and Van, 1991, Robert et al., 2012) quantifies the sensitivity of all measurements subject to changes in the electrical structures. Generally, areas with poor sensitivity or resolution are considered as less reliable. The main difficulty consists of finding an objective and quantitative way to select a threshold value which will define “well” and “poorly” resolved areas, especially since ground truth information is not always available.

The depth of investigation (DOI) index was introduced by Oldenburg and Li (1999) in order to discriminate the electrical structures related to surface data from the ones related to the choice of a reference model introduced during inversion. The idea of Oldenburg and Li (1999) is to force the inverted model to resemble a reference model. Analyzing resulting models by using different reference models (put mathematically in an index) allows finding the electrical structures that strongly depend on the choice of the reference model, and therefore are not constrained by surface data. The analysis of the DOI index was successfully used by several authors such as Hilbich et al. (2009), Marescot et al. (2003) or Robert et al. (2011) to appraise electrical images. However, the threshold value remains difficult to define. According to Oldenburg and Li (1999) and Marescot et al. (2003), DOI index values of 0.1 or 0.2 are recommended.

All these methodologies provide information on how well the regions of the subsurface are covered by the data. Recently, Caterina et al. (2013) compared quantitatively different resolution indicators to detect artifacts, to estimate the depth of investigation, to address parameters resolution and to appraise ERT-derived geometry. They compared numerical benchmark models to their corresponding ERT images to define the errors on inverted parameters and to evaluate them against resolution indicator values allowing the definition of threshold values. In the case of field applications, the errors on inverted parameters are not known except at borehole locations and the threshold value definition is uneasy. Caterina et al., 2011, Caterina et al., 2013 proposed a methodology to overcome this problem by defining a synthetic model that resembles the inverted model to appraise. Simulated data related to this synthetic model are then inverted (all acquisition and inversion parameters being the same) and compared to the true synthetic model. The errors on inverted parameter can then be evaluated against a resolution indicator as long as there is a correlation between both resolution indicators (the one related to the real data inverted model and the one related to the synthetic data inverted model). This methodology is time-consuming but works well for the analysis of the diagonal values of the resolution matrix or the cumulative sensitivity matrix.

Except for a few work such as the tracer experiment of Robert et al. (2012) where authors had access to ground truth information to objectively select a cut-off value, the depth of investigation is commonly estimated using a recommended but arbitrarily chosen cut-off value on the selected resolution indicator (diagonal of the resolution matrix, sensitivity or DOI index). Therefore, different practitioners could obtain significantly different depth of investigation.

In this paper, we present a new methodology to statistically determine the depth of investigation. This method is based on a modified DOI index approach where statistical distributions of DOI index values are computed to discriminate between reliably- and poorly-reconstructed cells. This is a quick and easy-to-use methodology allowing the identification of statistically poorly-reconstructed areas in depth. The efficiency of the methodology is tested on synthetic and field data where the statistically determined depths of investigation are compared to the depths estimated using recommended cut-off values.

Beyond this introduction, this paper is organized as follows. First, we define the three resolution indicators considered (resolution matrix, sensitivity matrix and DOI index). Then, we present our statistical approach based on a modified DOI index approach. This method is then tested and compared to other approaches of depth of investigation estimation. This comparison, based on synthetic and real-world data will then be discussed. Finally, we draw conclusions and recommendations for depth of investigation issues.

Section snippets

Inversion of dc resistivity data

The inversion of resistivity and IP data was carried out by using the 2D inversion program RES2DINV (Loke and Barker, 1996, Loke and Dahlin, 2002). The inversion software allows the application of least-squares or robust inversions. The Gauss–Newton smoothness-constrained least-squares algorithm is commonly used to determine smooth changes in the model parameters that would minimize the sum of the squares errors between the model response and the observed data values (deGroot-Hedlin and

Sensitivity and resolution matrix analysis

The Jacobian matrix defines the sensitivity of predicted data to changes in model parameters. The cumulative sensitivities are often defined as the sum of the absolute sensitivities of the arrays used, since negative sensitivity values also give information about the subsurface. Based on this definition, the cumulative sensitivity of each model block is a measure of the amount of information or “coverage” about the resistivity of the block that is contained in the measured data set. If the

Results and discussions

We benchmarked the statistical DOI methodology on synthetic examples and tested its robustness with field data. The results obtained are compared to the depths of investigation estimated using frequently used cut-off values on model resolution matrix and scaled DOI index values.

Conclusions

Compared to the existing image appraisal tools, where the depth of investigation is commonly defined using an arbitrarily chosen cut-off value, we proposed a statistical method which allows the identification of well-constrained cells of the model without using any arbitrary threshold. This method is based on the principles of the DOI index method. This method is thus data-driven and sensitive to inversion errors. It can easily be implemented and automated. Even if we only considered 2D surveys

Acknowledgments

This work was partly conducted under the auspices of the Walloon Region Ministry (FIRST SPIN–OFF GRANSPE visa no 091/6974). It was also a part of a Ph.D. thesis funded by the Research Institute for the Science and Management of Risks at University of Mons (Faculty of Engineering, Geology and Applied Geology Department). The authors thank the Editor and two anonymous reviewers for their relevant comments and suggestions that improve the overall quality of the manuscript.

References (39)

  • T. Dahlin et al.

    A numerical comparison of 2D resistivity imaging with 10 electrode arrays

    Geophys. Prospect.

    (2004)
  • F.D. Day-Lewis et al.

    Applying petrophysical models to radar travel time and electrical resistivity tomograms: resolution-dependent limitations

    J. Geophys. Res.

    (2005)
  • C. deGroot-Hedlin et al.

    Occam's inversion to generate smooth, two-dimensional models from magnetotelluric data

    Geophysics

    (1990)
  • L.S. Edwards

    A modified pseudosection for resistivity and induced polarization

    Geophysics

    (1977)
  • R.G. Ellis et al.

    Applied geophysical inversion

    Geophys. J. Int.

    (1994)
  • C.G. Farquharson et al.

    Non-linear inversion using general measures of data misfit and model structure

    Geophys. J. Int.

    (1998)
  • S. Friedel

    Resolution, stability and efficiency of resistivity tomography estimated from a generalized inverse approach

    Geophys. J. Int.

    (2003)
  • C. Hilbich et al.

    Applicability of electrical resistivity tomography monitoring to coarse blocky and ice-rich permafrost landforms

    Permafr. Periglac. Process.

    (2009)
  • A. Kemna

    Tomographic inversion of complex resistivity: theory and application

    (2000)
  • Cited by (0)

    View full text