Within the critical zone, regolith plays a key role in the fundamental hydrological functions of water collection, storage, mixing and release. Electrical resistivity tomography (ERT) is recognized as a remarkable tool for characterizing the geometry and properties of the regolith, overcoming limitations inherent to conventional borehole-based investigations. For exploring shallow layers, a small electrode spacing (ES) will provide a denser set of apparent resistivity measurements of the subsurface. As this option is cumbersome and time-consuming, larger ES – albeit offering poorer shallow apparent resistivity data – is often preferred for large horizontal ERT surveys. To investigate the negative trade-off between larger ES and reduced accuracy of the inverted ERT images for shallow layers, we use a set of synthetic “conductive–resistive–conductive” three-layered soil–saprock/saprolite–bedrock models in combination with a reference field dataset. Our results suggest that an increase in ES causes a deterioration of the accuracy of the inverted ERT images in terms of both resistivity distribution and interface delineation and, most importantly, that this degradation increases sharply when the ES exceeds the thickness of the top subsurface layer. This finding, which is obvious for the characterization of shallow layers, is also relevant even when solely aiming for the characterization of deeper layers. We show that an oversized ES leads to overestimations of depth to bedrock and that this overestimation is even more important for subsurface structures with high resistivity contrast. To overcome this limitation, we propose adding interpolated levels of surficial apparent resistivity relying on a limited number of ERT profiles with a smaller ES. We demonstrate that our protocol significantly improves the accuracy of ERT profiles when using large ES, provided that the top layer has a rather constant thickness and resistivity. For the specific case of large-scale ERT surveys the proposed upgrading procedure is cost-effective in comparison to protocols based on small ES.
Within the critical zone, the architecture and properties of the regolith, as well as its distribution across the landscape, play a key role in how rainfall is collected, stored and finally released to generate streamflow (Schoeneberger and Wysocki, 2005; Lin, 2010; Ghasemizade and Schirmer, 2013; Brooks et al., 2015). Factors such as the depth and composition of the soil cover and the rock weathering determine water pathways, storage capacity, residence times in the subsurface and subsequent interactions with surface water bodies (Freer et al., 2002; Hopp and McDonnell, 2009; Graham et al., 2010; Gabrielli et al., 2012; Lanni et al., 2013; Ameli et al., 2016).
However, limited access to the subsurface is a major hurdle to acquiring this information, meaning that often even the most basic data are missing, such as the transitions from the soil to the hard bedrock (Brooks et al., 2015). It is the complexity and spatial variability of the subsurface that make its characterization very challenging. Conventional investigation techniques (i.e. soil pits, drillings) of regolith are known to be invasive and of limited spatial representativeness – a trait causing them to be ignored in the vast majority of catchment studies (Burt and McDonnell, 2015; Parsekian et al., 2015). Several authors have also recently pointed out the subsurface as being the greatest knowledge gap in the understanding/modelling of hydrological processes, with a greater investment in “seeing” the subsurface needed to provide the Earth system modelling community with critical guidance on how to parameterize model subsurface structure depths and properties (Fan et al., 2019).
Geophysical prospection techniques have received increasing attention in recent years within the hydrological sciences community, thanks to their non-destructive character and ability to provide information on subsurface features over large areas. These investigative tools are now recognized as being essential for accurately characterizing the subsurface and studying water partitioning (Robinson et al., 2008; Loke et al., 2013; Binley et al., 2015; Brooks et al., 2015; Parsekian et al., 2015; Singha, 2017). Among the geophysical prospection techniques at hand, electrical resistivity tomography (ERT) is commonly used to characterize subsurface environments. This well-known technique is based on the injection of an electrical current through a pair of electrodes and the measurement of the resulting electrical potential between a second pair of electrodes along a line of dozens or hundreds of grounded electrodes. Through inversion schemes, ERT data are used to generate 2D and 3D electrical resistivity maps of the subsurface (see e.g. Binley and Kemna, 2005, for a detailed explanation of the ERT method).
The electrical resistivity tomography of the subsurface provides a weighted average of the electrical resistivity of its mineral grains, liquid and air (Archie, 1942; Keller and Frischknecht, 1966; Reynolds, 2011). Constitutive relationships can be used to link electrical resistivity to several properties and states that are of major interest to hydrologists, e.g. textural properties (Tetegan et al., 2012), porosity (Leslie and Heinse, 2013; Comte et al., 2018), hydraulic conductivity (Slater, 2007; Farzamian et al., 2015), water content (Brunet et al., 2010; Alamry et al., 2017) or solute concentrations (Bauer et al., 2006; Comte and Banton, 2007). While these constitutive relationships are essential for reliable hydrological interpretations (Binley et al., 2015), their accuracy largely depends on the resolution of the ERT images (Day-Lewis et al., 2005).
ERT has also been successfully used to characterize regolith architecture by delineating areas showing similar resistivity patterns (Crook et al., 2008; Comte et al., 2012; Leopold et al., 2013; Cassidy et al., 2014; Holbrook et al., 2014; Hübner et al., 2015; Uhlemann et al., 2015; Wainwright et al., 2016; Scaini et al., 2017). An increasing number of studies use automated edge detection approaches to delineate these key interfaces within the subsurface (Nguyen et al., 2005; Hsu et al., 2010; Chambers et al., 2012, 2013, 2014, 2015; Audebert et al., 2014; Ward et al., 2014; Uhlemann et al., 2015; Wainwright et al., 2016; Scaini et al., 2017). However, it has also shown that the application of these methods can fail – even when the true interface is sharp – because of insufficient sensitivity and accuracy in the vicinity of the interface (Chambers et al., 2013, 2014).
Ultimately, the geophysical information on the subsurface that can be derived from ERT investigations, either in terms of geometry or hydraulic properties, may be used to feed process-based hydrological/hydrogeological models in order to improve their rightness/realism in terms of spatial variability (e.g. Mastrocicco et al., 2010; House et al., 2016; Loritz et al., 2017; Comte et al., 2018). However, when resistivity models form the basis for constructing such process-based models, the accuracy of the geophysical information and its interpretation is a critical issue that can lead to seriously wrong models and correspondingly wrong model predictions (Andersen et al., 2013).
The characterization of subsurface properties and the delineation of structural units within it should thus go hand in hand with a suitable resolution of ERT images. Otherwise, the results can be inaccurate (Chambers et al., 2013, 2014; Clément et al., 2009, 2014; Ward et al., 2014). Chambers et al. (2014) emphasize that using ERT to detect thin surficial layers remains challenging. Indeed, when shallow structures are investigated, a small electrode spacing (ES) is required, as it delivers denser and well-discretized measurements of the shallow subsurface (Reynolds, 2011; Chambers et al., 2014). However, for large horizontal ERT surveys, small ESs are often not a viable option, as their implementation remains time-consuming. Depending on the regolith architecture and the size of the investigated area, it is thus challenging if not impossible to balance the requirement for shallow-layer characterization (i.e. smaller ES) against the competing need to cover the area of interest within a reasonable amount of time and cost (i.e. larger ES) (Chambers et al., 2014). Moreover, as shown by Kunetz (1966) or Clément et al. (2009), oversizing the ES might not only be inappropriate for the characterization of thin surface layers, but it may also affect the characterization of deeper layers by causing a depth-based resistivity bias, as a result of the inversion process which can be affected by a lack of shallow data.
In this study, we focus on which ES should be considered to characterize the entire regolith accurately. More specifically, we are concerned whether deep structures are well defined (within the limits of the intrinsic resolution limitation of the ERT method that translates to a drop in resistivity model sensitivity with depth) if the shallow structure is not well sampled (i.e. if a larger ES is used). These issues should be ideally addressed prior to the design of fieldwork campaigns to avoid any misinterpretation of field data. Ultimately, we aim at carefully choosing the ES parameter for adequately disentangling the subsurface architecture and properties (both for the shallow part of the subsurface and for the deeper layers). We intend to increase the potential for large horizontal ERT surveys – based on oversized ES – to equally deliver detailed knowledge on shallow subsurface structures. Within this work, we investigate the following.
Detailed map of topography and investigations made in the Weierbach catchment (background aerial photography from
How the inverse solution reconstruction is affected by the ES parameter (i.e. impact of the lack of shallow apparent resistivity data induced by
the use of an oversized ES on inverted ERT image accuracy) and which is the most appropriate ES value for accurately characterizing the entire
regolith (i.e. for both surface and deeper horizons). The potential for a new approach to improve the accuracy of ERT surveys relying on a large ES by adding interpolated levels of surficial
apparent resistivity based on a limited number of measurements with a small ES.
To this end, we use as a reference case study the Weierbach catchment, where an earlier ERT survey has been carried out in order to shed new light on the spatial heterogeneity of the subsurface but for which we have been facing ES choice-related issues. In addition to the field dataset, we investigate a set of synthetic soil–saprock/saprolite–bedrock models using a classical geophysical approach based on numerical modelling to corroborate and reinforce the results. While the former represents a field reality in terms of heterogeneity, the latter provides important information under controlled conditions and a priori exact knowledge. The assessment of the ERT images obtained from these two datasets is carried out considering the accuracy of the inverted resistivity distribution and the derived interface depths.
Synthesis scheme of the regolith in the Weierbach catchment (with
Our experimental test site is the forested Weierbach headwater catchment, located in the Luxembourgish Ardennes Massif (0.45 The solum is a stony loam soil with a mean thickness of 50 The subsolum has two parts. The upper subsolum (from 50 to 90 The slate hard bedrock starts, on average, at a depth of 140
Given this intrusive point-scale investigation knowledge, the solum–subsolum electrical resistivity interface is expected to be sharp, while the subsolum–hard bedrock interface is most probably more gradational. In addition, point-scale observations suggest a probable spatial variability of the subsurface electrical resistivity. Indeed, mirroring multiple weathering phases in the Luxembourgish Ardennes Massif over geological timescales (Moragues-Quiroga et al., 2017; Demoulin et al., 2018), cores obtained from deep drilling campaigns (Fig. 2d) reveal different weathering degrees in the Weierbach catchment (Fig. 2e). The top of the substratum presents a high weathering degree in the upper part of the basin (north and west of the catchment, morphologically expressed by a plateau). Elsewhere in the catchment, bedrock weathering is less pronounced (on hillslope position and along the eastern limit). This difference also implies contrasted surface-layer properties. As observed in soil pits on the plateau, slate fragments are smaller and less consistent and the clay content of the matrix is higher (Fig. 2f; regolithic saprolite subsolum type as per Juilleret et al., 2016). Elsewhere, soil pits exhibit bigger and more coherent slate fragments and less clay in the matrix (Fig. 2g; regolithic saprock subsolum type as per Juilleret et al., 2016).
For a characterization of the subsurface of the entire Weierbach catchment, we built a mesh of several large ERT profiles using the roll-along
technique (white lines drawn in Fig. 1, cumulative length of about 12
All measurements were taken with a Syscal Pro 120 (10-channel) resistivity meter from IRIS Instruments with multicore cables attached to stainless steel rod electrodes. A pulse duration of 500
To assess data accuracy, we measured however 25 % of the quadrupoles in a normal configuration. Reciprocal error calculation (defined as the percentage standard error in the average of the normal and reciprocal measurements; Wilkinson et al., 2012), together with the analysis of the SDs
obtained for the repeated measurements, allows us to characterize the measurements as both very precise and accurate (99.4 % of the SDs do not exceed 1 % and 99.1 % of the reciprocal errors are below 5 %; mean SD and reciprocal error values of 0.10 % and 0.68 %,
respectively). Even though the overall quality of the data was good, we applied a cleaning procedure (removal of obvious apparent resistivity outliers
and all quadrupoles presenting a measured potential lower than 10
First, all available processed apparent resistivity data were used for the inversion of each profile. Second, to match the set-up of the catchment-scale survey and document the associated loss in resolution, only quadrupoles measured with an ES of 2
As mentioned previously (see Sect. 2.1.1), for a given geological substratum, and according to rock weathering and pedological processes, the regolith
can be partitioned into three main units (Velde and Meunier, 2008; Juilleret et al., 2016), namely, from top to bottom,
the solum, which is the “true soil”, where pedogenetic processes are dominant, the subsolum, corresponding to weathered materials where the original rock structure is preserved and geogenic processes still dominate
(depending on the degree of weathering, the saprock and/or saprolite can be distinguished), and the hard bedrock.
Based on this three-layered subsurface conceptual model, and according to the specificity of the Weierbach catchment, we generated 25 1D “conductive solum–resistive subsolum–conductive bedrock” conceptual models to investigate different scenarios with varying resistivity and thickness contrasts. Solum and bedrock resistivity was set to 1000
It is worth noting that we have opted for the use of a 1D synthetic model structure but that the subsequent forward modelling and inversion processes will be then done in 2D in order to evaluate not only the accuracy, but also the precision of the inversion results. This would not have been possible using a 1D inversion scheme.
To mimic apparent resistivity measurements with the synthetic models, we simulated the electric field distribution resulting from current injections using the electric field distribution theory (Maxwell's equation) and the finite-element method. We performed numerical simulations using the AC/DC module of Comsol Multiphysics, complemented with a forward 3D modelling (F3DM) MATLAB script (Clément et al., 2011; Audebert et al., 2014). This script assesses automatically, for a quadrupole sequence, the electric potential between the two potential electrodes, for a given current. To achieve a realistic dataset reflecting the properties of a field survey, we applied a systematic Gaussian noise distribution with 3 % SD relative error to the apparent resistivity dataset to simulate the noise commonly recorded with the resistivity meter.
In addition to the Wenner–Schlumberger array which was used for the Weierbach catchment survey, the dipole–dipole array was also used here to simulate apparent resistivity from the resistivity models in order to broaden the modelling findings. The dipole–dipole and Wenner–Schlumberger arrays represent the two most commonly used ERT arrays (Carrière et al., 2017). Their successful application in field studies is mainly due to their
surveying efficiency and sensitivity (Dahlin and Zhou, 2004). In order to assess the effect of the lack of shallow apparent resistivity measurements
related to the ES choice, simulations of apparent resistivity for both arrays were conducted using five different ESs (0.25, 0.5, 1, 2 and 4
Illustration of the lack of shallow apparent resistivity data posed by the ES parameter-related choice exemplified by
Proposed approach to upgrade ERT datasets relying on a large oversized ES by adding interpolated levels of surficial apparent resistivity based on a limited number of measurements with a small accurate ES.
As exemplified in Fig. 3, the use of a larger ES leads to less apparent resistivity measurements that possibly induce a critical lack of shallow apparent resistivity data. In the event of an ERT survey carried out with a large ES – and for which the first acquisition level (i.e. quadrupoles whose external electrode separation is of the smallest possible extension; see Fig. 3) is too deep to properly characterize the subsurface structure's top layer (in our case the solum) – we propose to take advantage of the potential relationships between this first acquisition level and additional surficial apparent resistivity acquisition levels (i.e. quadrupoles with smaller external electrode separations; see Fig. 3) obtained from a reduced number of ERT profiles with a smaller ES. If the top layer has a rather constant thickness and resistivity, we consider that such relationships exist and could then be transposed to areas where the larger ES have been used and where data gaps prevail in the shallow subsurface. Illustrated in Fig. 4, this approach may eventually reduce the oversized ES-related effects.
Here, we assess the proposed approach by applying it to the ERT profiles with an ES of 2 Step 1 in Fig. 4 – from the set of apparent resistivity data measured with an ES of 0.5 Step 2 in Fig. 4 – we use the dataset created in the first step to assess (using regression analysis) whether robust relationships exist between the apparent resistivity data of the shallower levels obtained from the external electrode separations of 1.5, 2.5, 3.5 and 4.5 Step 3 in Fig. 4 – if such relationships exist, we ultimately use them to upgrade the ERT profiles relying on an ES of 2
Both (synthetic and field) datasets were inverted with the same procedure. Inverse solution reconstruction of the interpreted resistivity distribution
relied on the BERT code (Boundless Electrical Resistivity Tomography; Günther and Rücker, 2016). This code is based on the finite-element–forward modelling and inversion method described in Rücker et al. (2006) and Günther et al. (2006). The aim of the inversion process
is to calculate a resistivity model that satisfies the observed apparent resistivity data. A homogeneous starting model is generated with the median-measured apparent resistivity, for which a response is calculated and compared to the observed data. The starting model is then modified iteratively
until an acceptable convergence between the model response and the observed apparent resistivity is achieved. The root mean square misfit error (Loke
and Barker, 1996) and the
The constraints placed on the resistivity model during the inversion had to be carefully considered. Here, we used the same 2D inversion settings for
all apparent resistivity datasets: a smooth inversion optimization method (L2-norm), a z-weight factor of 1 for generating isotropic resistivity
distribution and a regularization parameter
For the synthetic dataset, we evaluate the agreement between true synthetic resistivity models and interpreted resistivity distributions using the
Nash–Sutcliffe model efficiency coefficient (NSE):
We compared the true interface depths with those that can be derived from inverted ERT images as an additional way to assess the accuracy of the results. In ERT image analysis, isosurface and derivative methods are the two groups of methods commonly used for this purpose. Here, we retained the group of derivative methods despite it being shown that these methods can fail because of insufficient sensitivity and accuracy in the vicinity of the interface (Chambers et al., 2013, 2014). Indeed, derivative methods represent the most universal way to extract interfaces because their use is relevant in both homogeneous and heterogeneous subsurface contexts (Chambers et al., 2014). It is worth noting also that derivative methods have already been used with success in other ERT studies, even when using an L2-norm (smooth) model constraint (e.g. Hsu et al., 2010; Chambers et al., 2012; Ward et al., 2014). Derivative methods assume that interfaces are located where changes in image properties are at a maximum. These changes can be detected using either the first or second derivatives, targeting maximum gradients or zero values, respectively (Marr and Hildreth, 1980; Torreão and Amaral, 2006; Sponton and Cardelino, 2015). In this study, we used the second derivative of ERT images and targeted zero values (e.g. Hsu et al., 2010) with Paraview software (Ahrens et al., 2005). To be consistent with the inverse solutions delivered by BERT (Günther et al., 2006), we calculated the second derivative on the logarithm of resistivity. Finally, we defined interfaces by following second-derivative zero contour continuity and horizontality as well as the resistivity distribution and its associated gradient (first derivative). Note that the delineation of some interfaces results from the merging of several zero contours to ensure their continuity.
The same accuracy criteria (NSE and interface depths derived from second-derivative zero values) were used for the field dataset of the Weierbach catchment to assess and compare the accuracy of the inverted ERT profiles obtained from the quadrupoles measured with an ES of 2
Nash–Sutcliffe model efficiency coefficient and mean interface depths resulting from the inversion of the 25 synthetic apparent resistivity models using the Wenner–Schlumberger array with the five different ESs. In plots showing the estimated interface depths, thick black lines indicate the expected values.
The 300 resistivity models resulting from the inversion of the synthetic resistivity models are provided as supporting information (Figs. S1–S6 and
S7–S12 in the Supplement for the Wenner–Schlumberger and dipole–dipole arrays, respectively). Depending on the models, the inversion process was terminated after 1 to 11 iterations. As suggested by the root mean square misfit error (average: 0.89 %, range: 0.40 %–2.12 %) and the
Scatter plot relating the apparent resistivity data corresponding to the first pseudo-depth acquisition level with an ES of 2
Nash–Sutcliffe model efficiency coefficient and mean interface depths resulting from the inversion of the 25 synthetic apparent resistivity models using the Wenner–Schlumberger array with an ES of 2
Nash–Sutcliffe model efficiency coefficient (NSE) and interface depths (avg
Nash–Sutcliffe model efficiency coefficient (NSE) and interface depths (avg
Note that since the results obtained for both arrays are very similar, we only present those obtained from the Wenner–Schlumberger array and that fit the field case study (Figs. 5–7, Tables 1–2). The same figures and tables resulting from analyses carried out on inversion results from the dipole–dipole array have nonetheless also been produced but are provided in the Supplement (Figs. S13–S15, Tables S1–S2 in the Supplement).
The visual examination of the inversion results (Fig. S1) and NSE values (Table 1, Fig. 5) obtained for the smallest ES (0.25
Although the information delivered when using an ES of 1
Finally, mean NSE values for the 25 synthetic models obtained from ERT images using ESs of 2 and 4
As shown in the scatter plots of Fig. 6, each of the four selected surficial apparent resistivity levels acquired with an ES of 0.5
The visual examination of the inversion results (Fig. S6) and NSE values obtained using the four surficial interpolated levels indicate an overall
good match between the ERT images and synthetic resistivity models (Table 2, Fig. 7). The mean NSE value for the 25 synthetic models is equal to 0.35. This value is much better than the one obtained when using the standard apparent resistivity datasets (i.e. 0.00). However, as indicated by negative NSE values, results for the lowest subsolum resistivity contrast (i.e. 1250
Regarding interface delineation, a strong overall improvement is also observed when adding the four surficial interpolated levels to the apparent
resistivity datasets using an ES of 2
Inversion results obtained for the 12 plot-scale ERT profiles measured in the Weierbach catchment using an ES of 0.5
Median resistivity as a function of depth for the 12 plot-scale ERT profiles measured in the Weierbach catchment using an ES of 0.5
The inversion results obtained for the 12 ERT profiles from the Weierbach catchment, with the two standard apparent resistivity datasets and the
upgraded dataset, are presented in Fig. 8. Four to 12 iterations were necessary to achieve the inversion process. In each case, an acceptable convergence between the calculated and simulated apparent resistivity data was reached, as indicated by the root mean square misfit error (average:
2.54 %, range: 0.94 %–4.82 %) and the
As shown in Figs. 8a and 9 (blue thick lines), the variability of resistivity with depth obtained using an ES of 0.5
A clear distinction between the different stages of weathering affecting the regolith is also possible, as revealed by soil pits and drillings. We are
able to identify two groups of profiles (Figs. 8a and 9). Profiles P05, P06, P07, P09, P10 and P12, located in the north and the west of the
catchment, were characterized by overall lower resistivity values for each of the subsurface layers than profiles P01, P02, P03, P04, P08 and P11,
which are located on steep slopes and the eastern catchment boundaries. For instance, the resistivity of the solum ranges from about 1500 to
2000
Interface depths (avg
Solum thickness and depth to hard bedrock derived from ERT images obtained with an ES of 0.5
Scatter plots relating the apparent resistivity data corresponding to the first pseudo-depth acquisition level with an ES of 2
For all 12 profiles, the scatter plots in Fig. 10 relate the first apparent resistivity acquisition level using an ES of 2
Distribution of the ratios calculated between the inverted resistivities obtained using an ES of 2
NSE values comparing standard ERT images obtained with an ES of 2
As shown in Figs. 8c and 9 (green thick lines), the enrichment of the apparent resistivity datasets using an ES of 2
Coloured density scatter plots (red – high density – to blue – low density) showing solum thickness and depth to hard bedrock derived from the 12 plot-scale ERT profiles measured in the Weierbach catchment using an ES of 0.5
Distribution of differences between interface depths obtained using an ES of 2
The use of standard ERT images obtained with an ES of 2
As for the accuracy of resistivity distributions, the enrichment of the apparent resistivity datasets using an ES of 2
In this study, we investigate a sequence of soil–saprock/saprolite–bedrock. The chosen synthetic three-layered “conductive solum–resistive subsolum–conductive bedrock” structure describes the subsurface of many natural contexts, such as the Weierbach catchment. Through our modelling exercise and the Weierbach catchment case study, we documented the ability and the limitations of ERT to correctly untangle such a typical regolith structure according to the ES parameter. Our results confirm that the choice of the ES is fundamental for obtaining accurate results, but most importantly it allows us to understand in detail from which ES threshold and why and how the accuracy of the inverted ERT images is affected.
Our results indicate first, for both arrays and whatever the ES retained, that resistivity and thickness contrasts play a key role in the resulting inverted ERT images. In general, for lower resistivity contrasts and shallower structures, the resulting inverted ERT images lead to relatively less well-resolved and fuzzy three-layer structures. Moreover, mainly for the lowest resistivity contrast, local resistivity artefacts are produced and disturb the accuracy of ERT images. At the opposite end of the scale, the higher the resistivity contrast and the deeper the structure, the more the ERT images tend towards a sharp, well-defined three-layered structure in our area of interest. However, in this case, for the strongest resistivity contrasts, the interpreted structures shift in depth, resulting in a decrease in ERT image accuracy. These relationships between resistivity contrasts and interpreted resistivity distributions logically affect the interface depths that are extracted from the second derivative of ERT images.
Our study also emphasizes the critical role of the ESs. The impact of these inverse problem effects as a function of the resistivity and thickness
contrasts on the accuracy of the geophysical information delivered does indeed largely depend on the ES parameter. While these effects are rather
negligible for the smallest ES, they increasingly deteriorate the accuracy of the ERT images with increasing ES values. More specifically, we observed
a threshold effect at an ES value of 0.5
Indeed, considering the depth of investigation of collinear symmetrical four-electrode arrays using the dipole–dipole or Wenner–Schlumberger arrays (Roy and Apparao, 1971; Barker, 1989), this ES allows a vertical resolution for the shallow parts of the subsurface of about 0.25
Recently, Chambers et al. (2014) highlighted the very significant challenges in using ERT to detect thin surface layers and suggested that a reliable resolution of surface layers with a thickness of less than one-third of ES should not be expected. This conclusion was based on the interface delineation accuracy but not on that of the resistivity distribution. Moreover, the use of derivative methods had failed in their case, and only isosurface methods gave good results. These methods, which consist of selecting a resistivity threshold value on the basis of intrusive measurements (Chambers et al., 2013, 2014; Wainwright et al., 2016) or using statistical analysis of the ERT images (Audebert et al., 2014; Ward et al., 2014), are indeed less dependent on the sensitivity of ERT images and have shown a greater ability than derivative methods in several cases (Ward et al., 2014; Chambers et al., 2013, 2014). The success of the application of isosurface methods is however restricted to specific case studies, resulting from the homogeneity of targeted resistivity layers which imply consistent interfaces (Chambers et al., 2013, 2014; Ward et al., 2014). In other cases, they provide poor results (Ward et al., 2014; Chambers et al., 2012). Our results are therefore not contradictory with the findings of Chambers et al. (2014).
We ideally recommend using an ES that is close to the thickness of the top subsurface layer in ERT surveys to mirror the architecture and properties of the subsurface correctly. This choice, which is obvious for the characterization of the shallower layer, is also relevant for characterizing the subsurface in its entirety – even when solely aiming for the characterization of deeper layers. However, this recommendation results from one typical subsurface structure and should consequently be transposed to areas of similar characteristics. This means that a generalization of our findings and their interpretation about the inverse solution accuracy problem posed by ES parameter-related choices is limited. Nevertheless, the same methodology as followed in this work might be used for other case studies, such as for example the reverse case, i.e. “resistive solum–conductive subsolum–resistive bedrock”.
The design of an ERT survey consists of a compromise between the need for high resolution for the near-surface layer (which would suggest smaller ESs) and the need to cover the area of interest in a reasonable amount of time and to an investigation depth that is deep enough to reconstruct the architecture of the deeper layer (which would give a preference for larger ESs; Chambers et al., 2014).
Eventually, as proposed by Dahlin and Zhou (2004), a quadrupole sequence of apparent resistivity measurements with decreasing vertical resolution and horizontal scanning in depth can reduce operational time without a drastic loss of accuracy. Moreover, in recent years, there has been substantial development of algorithms dedicated to automatically determining non-conventional electrode configurations (Loke et al., 2013). Those algorithms can lead to inverted ERT images whose resolution is superior to or equal than, respectively, the same or a fewer number of measurements, to those using standard survey designs, as for example Wenner–Schlumberger or dipole–dipole arrays (e.g. Stummer et al., 2004; Furman et al., 2004, 2007; Wilkinson et al., 2006, 2012; Loke et al., 2014; Abdullah et al., 2018; Uhleman et al., 2018). In the scope of large-scale ERT surveys, such optimized non-conventional electrode arrays could also help reduce the operational measurement time without reducing the information content. However, setting up the electrodes remains time-consuming, and the depth of investigation may be insufficient (e.g. ERT device with a limited number of electrodes). If the competing needs to cover the area of interest are still not reached (i.e. cost and time constraints, adequate depth of investigation), a set-up with larger ESs must be preferred, but the accuracy of the resulting ERT images might be affected by inverse solution reconstruction issues related to the lack of shallow apparent resistivity data as documented within this work. An improvement of these results is nevertheless possible by filling the lack of information in the shallow part of the subsurface. For instance, the deployment of a fast-moving measurement device (Andrenelli et al., 2013; Guerrero et al., 2016) could be used in parallel to complement the apparent resistivity dataset. Another example, as shown by Clément et al. (2009), is the use of an advanced inversion constrained by a priori surficial information to improve the accuracy of ERT images.
Both the synthetic and Weierbach catchment datasets demonstrated the potential for our novel upgrading procedure to improve the accuracy of large-scale ERT surveys based on large ESs. By adding four surficial apparent resistivity levels to the standard datasets using an ES of 2
The main constraint of the proposed upgrading procedure is that it is only applicable if the shallower layer is relatively homogeneous in terms of resistivity and thickness, which was the case for the synthetic models used. It is indeed this homogeneity that induces the good correlation between the surficial apparent resistivity levels and a deeper level. In the case of the Weierbach catchment, the solum is relatively homogeneous, as indicated by the point-scale investigations available and the 12 plot-scale ERT profiles whose locations were distributed between locations with different geomorphological characteristics. It is important to note that local inconsistencies are expected in places where the shallower part of the subsurface will not satisfy the overall solum homogeneity criteria. For instance, in the riparian zone, where solum and subsolum have been eroded, at forest roads, where the soil has been extensively modified (road cut, ballast), or in grasslands surrounding the catchment, where the soil does not have the same characteristics as in the forest zone, the application of the method would most probably lead to erroneous results by inducing false inverted surficial resistivity layers.
A second limitation of the proposed method is pointed out by the synthetic modelling results. The proposed approach fails and even leads to worse results if there is a low resistivity contrast between layers. Indeed, we were able to show that linear regressions leading to the interpolated levels of surficial apparent resistivity crossed each other in this case and led to an unsuitable variation pattern in the apparent resistivity (Fig. 7). Ultimately, this causes the formation of false resistivity layers by the inversion process (Figs. S6 and S12 in the Supplement). This problem could be solved to some extent by constraining the linear regressions with respect to each other, so that they do not cross. Other improvements of the method can be anticipated, such as a weighting procedure for the inversion of the interpolated levels of surficial apparent resistivity levels, depending on how well they correlate. Relying on a cross-validation to assess the efficiency of the dataset used for the calibration of the regression models for delivering robust estimates of the surficial levels of apparent resistivity would also be beneficial. This statistical method, which relies on splitting the full dataset into two parts, namely the training and validation datasets, is widely used in prediction problems for estimating how accurately a predictive model will perform in practice (Hastie et al., 2008). It would allow for instance assessment of the effect of the dataset size (e.g. number of ERT profiles) and characteristics (e.g. location of the ERT profiles) on the prediction results.
Here again, it is worth recalling that our findings and their interpretation result from one typical subsurface structure. Extra work is needed to strengthen, and eventually adapt, our upgrading approach to a more general regolith pattern. We especially recommend assessing the proposed methodology for the reverse case “resistive solum–conductive subsolum–resistive bedrock”. For instance, a set of three-layered “resistive–conductive–resistive” synthetic models might be explored to confirm/infirm the linear regression logic we highlight in this study.
The ideal design of an ERT survey exploring the architecture and properties of the Weierbach catchment's regolith should rely on an ES of
0.5
Other geophysical methods exist that might be more efficient than ERT in exploring the regolith over large areas (Binley et al., 2015; Parsekian et al., 2015; Fan et al., 2020). For instance, ground-penetrating radar (GPR) allows usually a higher spatial resolution and data collection rate. In several studies, GPR has allowed us to accurately delineate interfaces of several relevant structures in the critical zone (e.g. Carrière et al., 2013; Hare et al., 2017; Guo et al., 2019; Šamonil et al., 2020). However, as GPR requires sliding of the instrument on the ground, its use is much more ticklish and time-consuming over long distances in forests than in open areas, such as grassland or cropland. Furthermore, GPR surveys were done in the close vicinity of the Weierbach catchment. The structural analysis only revealed soil layering, but it did not show the depth to bedrock due to chaotic reflection patterns, which are common for this type of geologic setting (Jackisch et al., 2017; Allroggen et al., 2020). The electromagnetic induction (EMI) method could also have been another way to quickly characterize the shallow subsurface at a large scale. Indeed, unlike GPR or ERT, EMI systems do not require a direct coupling with the ground, which allows much faster acquisitions, even in forested areas. Despite their limited spatial resolution and depth of investigation, multi-depth EMI devices have been used in several studies for the characterization of subsurface structures and properties (e.g. Brosten et al., 2011; Saey et al., 2012; Rejiba et al., 2018; Simon et al., 2020). A field test using a multi-frequency domain EMI device (Profiler EMP-400, GSSI) was done in the Weierbach catchment. Unfortunately, the results were inconclusive, as contrasting shallow patterns observed with ERT were in fact not distinguishable due to the overall too electrically resistive nature of the subsurface.
Hence, the upgrading procedure proposed in this study is particularly interesting in the context of the Weierbach catchment. Through our study, we
demonstrated that applying this new approach to the existing catchment-wide ERT dataset measured with an ES of 2
Yet the intrusive point-scale investigation suggests a potentially significant spatial variability of the subsurface hydraulic properties (in close relation to the observed soil clay content, subsolum slate fragment size and bedrock weathering heterogeneity; see Sect. 2.1.1) that might be derived from the catchment-wide ERT survey as suggested in this study (see Sect. 3.2.1). Moreover, as highlighted by Loritz et al. (2017) in a nearby
catchment with the same regolith structure, the bedrock topography plays a significant role in the interplay of water flow and storage in our study
area. Using a physically based hillslope modelling approach, they showed that a model with surface-parallel bedrock topographies performed considerably worse in matching streamflow than a model including a bedrock topography. Furthermore, in their model, the topography of the bedrock was
successfully constrained with an ERT survey using an ES of 0.5
The presented elements suggest that the application of the upgrading procedure to the catchment-wide ERT survey dataset relying on an ES of
2
In catchment studies, our understanding of the mechanistics and prediction of subsurface water flow paths, transit times and storage volumes remains fractionate, largely as a result of our poor knowledge on regolith structure and characteristics. The characterization of the subsurface is stymied by the invasive and “point-scale” characters of traditional investigation techniques, essentially because of time and cost constraints. ERT is one of the geophysical tools at hand to overcome this limitation. This technique is now commonly used in the critical zone to disentangle regolith properties and architecture, but its use should go hand in hand with a suitable resolution of ERT images.
In this paper, we discuss the importance of ESs for the quality of ERT images to adequately mirror subsurface resistivity distributions and accurately delineate interfaces. To this end, we investigated a synthetic “conductive–resistive–conductive” three-layered sequence of soil–saprock/saprolite–bedrock, which mirrors the subsurface of many natural contexts, in combination with the Weierbach catchment field dataset, as a reference case study. Inversion results obtained for different ESs were compared in terms of resistivity distribution accuracy. We also inferred interface depths from each ERT image using a derivative method and evaluated their accuracy.
Our results highlight the need to use an adapted vertical resolution to best mirror the structure of the subsurface. More specifically, we document the inverse solution reconstruction issues related to the lack of shallow apparent resistivity data induced by the use of an oversized ES. We found out that the thickness of the most superficial layer must be taken into consideration when choosing the ES. Specifically, we demonstrated that the best compromise consists of using an ES close to the thickness of the subsurface top layer. If a larger ES is retained, the accuracy of the results decreases rapidly in terms of both resistivity distribution and interface delineation. This choice, which is obvious for the characterization of the shallower layer, is also relevant for characterizing the subsurface in its entirety – even when solely aiming for the characterization of deeper layers. For instance, our observations obviously support previous findings and confirm that oversizing the ESs not only leads to an inappropriate vertical resolution for the delineation of thin surface layers, but that it also affects the outlining of deeper layers. In particular, we demonstrated that an oversized ES leads to overestimations of the depth to bedrock and that this overestimation is even more important for subsurface structures with high resistivity contrast.
To overcome this limitation, we propose adding interpolated levels of surficial apparent resistivity based on a limited number of ERT profiles with a small ES that satisfies the thickness of the top subsurface layer. We show that our protocol significantly improves the accuracy of ERT profiles based on large ESs provided that the top layer has a rather constant thickness and resistivity, such as the solum in the Weierbach catchment. Our results demonstrated that this upgrading procedure is promising for carrying out large-scale surveys in a cost-effective and more robust way, for instance to feed hydrological models with subsurface structure depths and properties at catchment scale. However, our findings and their interpretation result from one typical regolith logic, and extra work is needed to strengthen, and eventually adapt, our upgrading approach to a more general regolith pattern. We especially recommend assessing the proposed methodology for the reverse case “resistive solum–conductive subsolum–resistive bedrock”.
Data used to generate the resistivity models are the property of and archived by the Luxembourg Institute of Science and Technology. They are available on request from the corresponding author.
The supplement related to this article is available online at:
CH and LP secured funding that supported this research work. LG directed the study and conceptualized its experimental design. LG carried out the field work with support from CH and JJ. LG and RC designed the synthetic case study. LG and RC performed the forward and inverse modelling, respectively. LG analysed, interpreted and discussed the results with assistance from all the co-authors. LG prepared the manuscript, with contributions from all the co-authors in the reviewing and editing thereof.
The authors declare that they have no conflict of interest.
This work was funded by the Luxembourg National Research Fund (FNR) as part of the SOWAT (FNR/CORE/C10/SR/799842/SOWAT) and CAOS2 (FNR/INTER/DFG/14/02/CAOS2) projects as well as through the Luxembourg Institute of Science and Technology (LIST). We would like to thank Cyrille Tailliez and Jean-Francois Iffly for their help during the field campaigns. We also express our gratitude to T. Günther for providing the BERT2 software and supportive guidance.
This research has been supported by the Luxembourg National Research Fund (FNR) (grant nos. FNR/CORE/C10/SR/799842/SOWAT and FNR/INTER/DFG/14/02/CAOS2).
This paper was edited by Alberto Guadagnini and reviewed by two anonymous referees.