Journal of Applied Geophysics 68 (2009) 522–532

Contents lists available at ScienceDirect

Journal of Applied Geophysics

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / j a p p g e o

4-D inversion of DC resistivity monitoring data acquired over a dynamically changing

earth model
Jung-Ho Kim a,⁎, Myeong-Jong Yi a, Sam-Gyu Park a, Jae Gon Kim b
a
Geoelectric Imaging Laboratory, Korea Institute of Geoscience and Mineral Resources, 30 Gajeong-dong, Yuseong-gu, Daejeon 305-350, South Korea
b
Geologic Hazards Department, Korea Institute of Geoscience and Mineral Resources, 30 Gajeong-dong, Yuseong-gu, Daejeon 305-350, South Korea

a r t i c l e i n f o a b s t r a c t

Article history: Geophysical monitoring is used principally to interpret the locations and amounts of ground condition
Received 17 July 2008 changes. To achieve these objectives, differences are computed and examined using time-lapse images
Accepted 2 March 2009 calculated under the time-invariant static assumption, that any material property changes during the data
measurement can be practically ignored. These monitored data, however, can be contaminated with noise
Keywords: and frequently generate false anomalies of ground condition changes. Furthermore, the assumption of the
4-D inversion
static model can be invalid if the material property changes significantly during data acquisition. To alleviate
Geophysical monitoring
Resistivity tomography
these problems, we developed a new least-squares inversion algorithm that allows for the subsurface
Dye tracer experiment properties to continuously change in time. We define the subsurface structure and the entire monitoring data
Dynamic earth model in the space–time domain, allowing us to obtain a four-dimensional space–time model using just one
inversion process. We introduce the regularizations not only in the space domain but also in time, resulting in
reduced inversion artifacts and improved stability of the inverse problem. We demonstrated the performance
of the proposed algorithm through numerical experiments that assumed several scenarios of ground
condition changes and data acquisition sequences. Finally, the applicability to field data was proven by
applying the developed algorithm to the monitoring data of crosshole resistivity tomography jointly
performed with a dye tracer flooding experiment. This experiment had a small enough scale that we could
not ignore the change of material properties during the data measurement.
© 2009 Elsevier B.V. All rights reserved.

1. Introduction resultant time-lapse images, however, may be strongly dependent on

this a priori model. Another possible source of errors is the
Geophysical monitoring, which images the changes of subsurface assumption of a time-invariant static subsurface model. This model
material properties with time, has been widely applied to many supposes that data can be acquired very quickly compared with any
environmental and engineering problems. In particular, direct current changes of the subsurface material properties. If the material property
(DC) resistivity monitoring has been actively used to obtain hydro- changes significantly during data acquisition, this assumption of a
geological properties, since the resistivity of subsurface material is static model will be invalid; an example is when fluid migrates very
easily affected by conductive or resistive fluid injection (Daily and quickly in a highly permeable medium. In such cases, the changes of
Ramirez, 1995a,b; Daily and Ramirez, 2000; Slater et al., 2000; Alpak material properties during data acquisition cannot be ignored without
et al., 2004; Vargemezis et al., 2007). An approach commonly used for the resultant subsurface images becoming severely distorted. Day-
interpreting monitoring data is to independently invert the measured Lewis et al. (2002) proposed an algorithm of time-lapse inversion to
data acquired at each monitoring phase and to reconstruct time-lapse overcome this problem in the monitoring of cross-hole radar
images. Then, the ground condition changes are studied using tomography. However, their approach may not be easily applied to
primarily the difference images computed from the time-lapse other types of geophysical monitoring data, such as DC resistivity or
images. However, the time-lapse images may be contaminated with electromagnetic data.
inversion artifacts due to measurement error; moreover, these In order to address these problems, we propose a new four-
inversion artifacts are significantly amplified in the difference images dimensional (4-D) least-squares inversion algorithm which provides
(Kim, 2005). Several studies have attempted to minimize any possible reasonable subsurface images even when changes of the subsurface
inversion artifacts by using the inversion result of the background or properties during the measurement cannot be ignored. The subsur-
of the initial data as an a priori model for inverting the data at each face model is assumed to change continuously in time; thus, the
monitoring phase (Loke, 1999; LaBrecque and Yang, 2001). The subsurface medium is defined in a space–time domain. The coordi-
nates of the measured data are also defined in a 4-D space, in which
⁎ Corresponding author. Tel.: +82 42 868 3172; fax: +82 42 861 9721. the data measured at the same spatial coordinates but at different
E-mail address: jungho@kigam.re.kr (J.-H. Kim). times are regarded as different data. Including the time dimension

0926-9851/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.jappgeo.2009.03.002
 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532 523

allows the simultaneous inversion of the entire monitoring data sets Using the velocity vector of model change, the model vector at an
to obtain a 4-D model using just one inversion process. This model arbitrary time t (τk ≤ t ≤ τk + 1) is expressed as
enables us to describe the subsurface properties as they continuously
change in time. Furthermore, we introduce the regularizations in both pðt Þ = uk + ðt − τk Þvk : ð4Þ
the space domain and in the time domain, to effectively reduce
inversion artifacts and to improve the stability of the inverse problem. Even though the 4-D subsurface structure is defined by several
To efficiently describe a 4-D subsurface model, we introduce several reference models, a huge amount of computation is needed to obtain
reference space models along the time axis, using the approximation the forward model response, since the model in the space domain
that the material property varies linearly in time between two changes for each instance of measurement. To eliminate this problem,
subsequent reference models. This allows us to approximate and define the forward model response at an arbitrary time t between two
the distribution of material properties in the entire space–time domain reference times, τk and τk + 1, is approximated with the weighted
using the several reference space models. Even though the 4-D average of the first order Taylor expansions of the forward modeling
subsurface structure is defined by these reference models, a huge responses at two reference times, F(uk) and F(uk + 1).
amount of computation is inevitable to obtain the forward model
response, since the model in the space domain changes for every Gðt Þ;
τk + 1 − t
τk + 1 − τk
fFðuk Þ + ðt − τk ÞJk vk g +
t − τk
τk + 1 − τk

F uk f + 1

ð5Þ
instant of measurement. To alleviate this heavy computational effort,
we approximate the forward model response at an arbitrary time using
a first order Taylor expansion of the forward modeling results based on
 
+ t − τk + 1 Jk + 1 vk ; g
the reference models.
where Jk = AFAu
ðuk Þ
are the partial derivatives of the model response at
The performance of the proposed algorithm is demonstrated by k
reference time τk with respect to the reference model uk. To solve the
numerical experiments using synthetic data from DC resistivity
least-squares inversion, we must calculate the partial derivatives of
tomography surveys. In addition to these numerical experiments,
the responses in the entire 4-D domain in which the monitoring data
we applied the developed algorithm to cross-well resistivity monitor-
are defined. This is easily accomplished by differentiating the Eq. (5)
ing data acquired by a dye tracer experiment. In this case, the material
with respect to a reference model ui, where i = 1,2,…,m.
properties vary within the measurement times and this variation
Using the approximations above, an inverse problem to find the
cannot be neglected. The applicability to field data is investigated by
entire space–time model can be simplified into an inverse problem
comparing the images by the proposed algorithm with those by a
that seeks the optimum of several reference space models, i.e., U. Let
conventional algorithm. These images are also compared with the dye
us define the error vector e and the predicted error vector e′ as
stain images captured directly over the tomographic planes vertically
excavated after the dye tracer experiment. e = d − GðUÞ; e0 = d − GðU + ΔUÞ; ð6Þ

2. 4-D inversion algorithm where d is the measured data vector. Note that the measured data
must be expressed in both space and time coordinates. Obviously, data
One possible way of describing a 4-D subsurface structure is to measured at the same spatial coordinates but at different times are
sample the subsurface model at regular intervals which are spaced regarded as separate data.
close enough to capture how the subsurface structure changes We want to minimize the l2 norm of the predicted error by the
continuously in time. The space–time model vector P is defined as inversion process. Since both the data and the model are defined using
space–time coordinates, we are able to adopt two regularizations in


P = p1 ; N pi ; N pn ; ð1Þ the time and space domains to stabilize the inversion and to obtain a
more meaningful time–space model. Consequently, the objective
where pi is the space model vector at a single time i. However, this function to be minimized by the inversion process can be expressed as
definition of the space–time model is hugely impractical due to the follows:
enormous amount of computation space and time needed. To solve
Φ = jje Wd e jj + λW + αC;
this problem, we randomly sample the 4-D subsurface structure at 0T 0 2
ð7Þ
pre-selected times along time axis. The pre-selected times are called
the reference times, and the models in the space domain at those where Ψ and Γ are the two regularizations in the space and time
times are called the reference space models. Hence, the 4-D subsur- domains, respectively, λ and α are the Lagrangian multipliers for
face model U is defined using the reference model vectors as controlling these two regularizations, and Wd is the data weighting
matrix. We adopt a roughness functional in the space domain of Ψ to
U = fu1 ; N uk ; N ; um ð2Þ achieve a smooth model. For the constraint in the time domain, we
assume that the model change between two consecutive reference
where uk is a reference space model vector at a pre-selected reference models, i.e., uk and uk + 1, will not be significant (Kim, 2005). Ψ and Γ
time τk, and m (m b b n) is the number of reference times. The reference are defined as
times should be chosen based on the number of surveys, the total
 n T  n 
elapsed time of the monitoring survey, the time needed to obtain the W = A ΔU A ΔU ; ð8aÞ
data at each phase, and the amount of model change expected. We are
concerned with a 4-D structure in which the material properties and
continuously change in time, and we assume that the material property
at a fixed position changes linearly between two consecutive reference X
m −1  
times. This assumption of a linear temporal change in material property C= jjðuk + Δuk Þ − uk + 1 + Δuk + 1 jj2 ð8bÞ
i=1
enables us to define a velocity vector of the model change using the T
reference models at two reference times, τk and τk + 1, as = fMðU + ΔUÞg MðU + ΔUÞ;

where M is a square matrix in which only the diagonal and one sub-
du u − uk diagonal element have values, 1 or − 1. We can regard the functional Γ
vk = = k+1 ð3Þ
dt τk + 1 − τk as a measure of the model roughness in the time domain. In this study,
524 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532

the Lagrangian multiplier α which controls the regularization in the second regularization is the minimization of the model roughness
time domain is set to be constant; however, it could instead be defined along the time axis. In this study, the model roughness in the time
as a variable in reference time. domain is defined as the model changes between two consecutive
Minimizing the objective function (7) with respect to the model reference times; its measure is expressed in Eq. (8b). The amount of the
perturbation vector yields the following normal equation, contribution of this regularization about the inverse model is controlled
by assigning a different value to the Lagrangian multiplier (α). Hence,
   
T T T T T the larger the value we choose for the Lagrangian multiplier, the more
A Wd A + C ΛC + αM M ΔU = A Wd e − αM MU ; ð9Þ
similar are the reference models that result from the 4-D inversion.
Various values of the multiplier in the time domain were tested in both
where A are the partial derivatives of the responses in the 4-D domain numerical experiments and field applications. All of the tests showed
in which the entire monitoring data are defined, and C is the differential that introducing the constraint in the time domain resulted in a more
operator in the space domain. The active constraint balancing (ACB) stable convergence. It was found that 0.05–0.2 was the proper value for
method (Yi et al., 2003) is adopted to balance the constraints in the the Lagrangian multiplier α; this proper value should be chosen by
space domain; thus, the Lagrangian multiplier λ is expressed as a considering the signal to noise (S/N) ratio and the amount of model
diagonal matrix Λ in the above equation. Solving the normal Eq. (9) change expected. The value 0.1 was used in both the numerical
iteratively results in the final inverted space–time model. experiments and the field application. With the numerical experiments,
One of the important issues in an inverse problem is the the rms error from the final inverted model resulting from the 4-D
optimization of the Lagrangian multipliers through which we control inversion was nearly the same or slightly higher than the rms error
the contributions of regularizations about the solution to be obtained. using the conventional approach. Conversely, with the field data
In this study, two regularizations are introduced in both the space and applications, the rms error from the 4-D inversion was slightly lower
time domains; thus, two Lagrangian multipliers (λ, α) should be than that using the conventional approach.
chosen as optimal values. Since we adopted the ACB technique, the
Lagrangian multiplier (λ) for controlling the contribution of the 3. Numerical experiments
regularization in the space domain is not a constant, but is a spatially-
dependent variable. These spatially-varying multipliers were auto- To verify the performance of the proposed algorithm, we
matically calculated and optimized during each iteration step via the conducted numerical experiments in which the subsurface models
parameter resolution matrix and the spread function analysis. The are categorized into the following two cases: 1) a static model, in

Fig. 1. A test model in which only the rectangular conductor beneath the surface layer is changing in resistivity. The background or initial resistivity of the conductor is set as 100 Ω m
(Phase 1).
 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532 525

which the change of subsurface material properties can be safely 2) Increase of the resistivity of the anomalous body: 100 (Phase 1), 200
ignored during measurements of the time-lapse data. 2) a dynamic (Phase 2), 400 (Phase 3), 800 (Phase 4), and 1600 Ω m (Phase 5).
model, in which these changes with time cannot be neglected.
Assuming 2-D structures in the space domain, we calculated the The adopted array was dipole–dipole, and the measurement interval
monitoring data using several scenarios of ground condition change was 8 m. There were five monitoring surveys. Synthetic data for the
and data acquisition sequence. surface measurements were calculated using the 2.5-dimensional finite
element modeling code based on the above scenarios and assembled
3.1. Static model into one set of 4-D data. In these experiments, the static model assumes
that the earth does not change continuously; the earth remains
Based on the subsurface model of Fig. 1, two monitoring unchanged for the duration of the measurement at each time-lapse.
experiments were performed based on the following scenarios of This assumption can be easily implemented in our algorithm, by
the material property change. holding constant the time coordinates of the data measured at each
phase, and by fixing the number of reference models to that of the
1) Increase of the conductivity of the anomalous body: 100 (Phase 1), monitoring phases. Random electrical noise of 5 mV/A peak-to-peak
65 (Phase 2), 40 (Phase 3), 20 (Phase 4), and 10 Ω m (Phase 5). amplitude was added to the synthetic potential difference data to

Fig. 2. (a) and (c) are the difference images between Phases 1 and 2 when the anomalous body of Fig. 1 becomes conductive, while (b) and (d) are those between Phases 2 and 3
when the body becomes resistive. (a) and (b) are reconstructed using time-lapse images generated by the conventional separate inversion, and (c) and (d) are by the 4-D inversion
algorithm.
526 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532

simulate the field data. The input data for the inversion are not potential approach (Fig. 2a and b), severe inversion artifacts contaminate the
differences, but are apparent resistivity; thus, the noise characteristics images. The artifacts may contribute to a misinterpretation of the
of the input data are not random. The S/N ratio drastically decreases as ground condition change. These severe artifacts are mainly due to the
the distance between the current and the potential dipoles increases. non-uniform noise of apparent resistivity, which becomes drastically
The synthetic monitoring data were inverted by the proposed 4-D stronger as the electrode separation index increases. Alternatively, in
algorithm. As a comparison, they were also simultaneously inverted the difference images reconstructed from the 4-D inversion results
without the constraint in the time domain (α = 0 in Eq. (9)); this (Fig. 2c and d), we can clearly and easily recognize the places where
simulated the conventional approach of separately inverting the data the change of material properties take place, even although the input
at each time instance. The difference images between an arbitrary pair data are contaminated with severe and non-uniform noise. These
of inversion results at different time lapses were reconstructed by remarkable improvements of difference images are ascribed to the
computing the ratios relating the resistivity distribution at one time to regularization in the time domain used in the 4-D inversion algorithm.
that at another time.
The time-lapse images inverted by the two different inversion 3.2. Dynamic model
approaches did not vary much from each other, and the rms errors
were also nearly the same. The difference images, however, had large In this subsection we handle the case when the subsurface material
variations between them, as illustrated in Fig. 2. Although the properties continuously and quickly change, so that we cannot ignore
conductivity increase and decrease of the anomalous body can be the variation of the subsurface structure in time (i.e., the dynamic
recognized in the difference images generated by the conventional model). Since the coordinates of the measured data are expressed in

Fig. 3. A test subsurface model in which the resistivity distribution is continuously changing in time. This figure shows the changing feature of the test model in snapshots when t = 0
(a), t = 0.2 (b). t = 0.4 (c), t = 0.6 (d), t = 0.8 (e), and t = 1 (f), where t is the time normalized by the total measurement period. The circles denote the locations of electrodes.
 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532 527

both the space and time domains, the proposed algorithm can be consisted of the data acquired through three kinds of acquisition
applied not only to multiple sets of time-lapse data, but also to only one methods: crosshole, inline, and hole-to-surface. The adopted array
data set. However, in principle it is much more difficult to obtain the was pole-dipole, and the measurement interval was 2 m.
subsurface model if there is only one data set provided. For general When a subsurface model continuously changes during the
geophysical exploration, a single data set can be considered to be a measurement, the sequence of data collection becomes very impor-
simple survey rather than geophysical monitoring. With the assump- tant (Day-Lewis, 2002). The sequence is particularly important for
tion that the earth is continuously changing, yet only one complete data finding an optimal 4-D subsurface model from just one set of data. We
set is available, we performed numerical experiments of a cross-hole presumed two strategies for the data collection:
resistivity survey to verify the performance of our proposed algorithm.
Fig. 3 illustrates the subsurface model used for the numerical 1) Random acquisition scenario. The measurement is performed
experiments of the dynamic model, which has data taken as snapshots randomly so that the data obtained by each acquisition method
at certain elapsed times. An example situation that requires this 4-D may contain the responses of all 30 different space models.
model is if conductive fluid is infiltrating into the basement rock from 2) Sequential acquisition scenario. First, a crosshole survey is
the soil layer boundary, and is accumulating inside the resistive performed, next the inline method, and finally, the hole-to-surface
basement. To simulate the forward response to the dynamically acquisition method. Correspondingly, the responses of the early
changing earth, the space–time model was sampled as 30 space time model (for example, Fig. 3a and b) are dominant in the
models (n = 30). From these 30 sets of modeled data, we assembled crosshole data, while those of the late time model (for example,
one 4-D data set. One complete data set of the tomography survey Fig. 3e and f) are dominant in the hole-to-surface data. Obviously,

Fig. 4. An illustration of the 4-D inversion results as snapshots at different elapsed times, t = 0 (a), t = 0.2 (b). t = 0.4 (c), t = 0.6 (d), t = 0.8 (e), and t = 1 (f), where t is the elapsed
time normalized by the total measurement period. Input data were calculated based on the model illustrated in Fig. 3, and the random data acquisition was presumed.
528 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532

in this scenario it is much more difficult to find the 4-D model than clear snapshot images such as those from the random scenario, the
it is in the first scenario. images are still reasonable and we can easily recognize the consider-
able changes of subsurface structure (migration of the conductive
Fig. 4 illustrates the inversion results obtained using the random zone) that happened during the data collection (Fig. 5).
acquisition scenario. There were only two reference models chosen, The same data were further inverted after increasing the number of
and the corresponding pre-selected reference times were τ1 = 0.17 reference models to three; this allowed us to examine the effect of the
and τ2 = 0.83, where τ is the elapsed time normalized by the total number of reference models. The results were nearly the same as those
measurement period. Although the entire subsurface change event using two reference models. From these experiments, we were able to
had been completed during one run of the geophysical survey, the know that either two or three reference models per one monitoring phase
snapshots reconstructed from the inverted results reasonably repro- are sufficient to reconstruct the subsurface images and to reasonably
duced the changing features of the subsurface structure, as shown in reflect the continuous change of underground structure with time.
Fig. 3; this mimics the migration of a conductive zone.
A crosshole resistivity survey would commonly use the sequential
acquisition scenario rather than the random scenario. Fig. 5 shows the 4. Application to field data: crosshole resistivity monitoring of dye
inversion results for the case where the tomography data were tracer infiltration
measured using the sequential scenario. The number of reference
models and their reference times were set to match those of the To examine the applicability of electrical resistivity tomography to
random acquisition case. Although the inverted results do not show the monitoring of water movement in a soil profile, crosshole

Fig. 5. An illustration of the 4-D inversion results as snapshots at different elapsed times, t = 0 (a), t = 0.2 (b). t = 0.4 (c), t = 0.6 (d), t = 0.8 (e), and t = 1 (f), where t is the elapsed
time normalized by the total measurement period. Input data were calculated based on the model illustrated in Fig. 3, and the sequential data acquisition was presumed.
 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532 529

Fig. 6. Layout of the DC resistivity monitoring survey and dye tracer flooding
experiment. Large filled circles denote the boreholes, and small circles denote the
surface electrode locations. The dye tracer solute was flooded onto the shaded rectangle
with size 1 × 1 m.

resistivity monitoring and a dye tracer experiment have been jointly

carried out. This allowed the time-lapse images from the geophysical
monitoring to be easily compared with the stain images formed by the
migration of the dye solution. In this section, we discuss the inversion
experiments using these monitoring data.

4.1. Field works

The topmost layer of the test site consisted of residual soil

developed from Jurassic granite. Six boreholes were drilled down to
about 1 m, and three tomography sections connecting borehole pairs
were established, as shown in Fig. 6. Brilliant Blue FCF solution, with a
conductivity of 0.09 S/m, was flooded three times during the
monitoring period. The flooding and monitoring history is summar-
ized in Table 1. After three days of the flooding and monitoring, the
test site was vertically excavated at 10 cm intervals to record the stain
images and to examine the infiltration pattern.
The sequence used to collect one set of the tomography data at
each monitoring phase consisted of 1) inline, 2) crosshole, 3) surface-
to-hole, 4) hole-to-surface, and 5) surface measurements. The
electrode interval was set at 0.1 m. A pole-dipole array was adopted, Fig. 7. Time-lapse tomograms of Tomo4–5 at the time of beginning (a) and ending (b)
the measurement of Phase 2, and the time of completing Phase 6 (c).
and the dipole length and measurement interval were 0.2 m and
0.1 m, respectively. Since the pole-dipole array is an asymmetric
electrode array, we acquired forward-crosshole data as well as experimental scale was very small and the residual soil was the
reverse-crosshole (Yi et al., 2003). The data were measured using a medium of the infiltration of dye tracer.
SuperSting R8/IP™ equipped with a 108 channel switch box made by
AGI Inc. At every monitoring phase, we obtained the data along the 4.2. Inversion results and discussions
three tomographic sections shown in Fig. 6. It took about 85 min to
obtain one complete data set (2450 samples) for each monitoring The data acquired in Phase 1 unfortunately had a too low S/N ratio
phase. It was expected that the subsurface structure changed too for obtaining a reasonable subsurface image. Therefore, one set of
quickly to ignore the time variations during measurement, since the time–space data was assembled using the data acquired from Phases
2–6, excluding Phase 1. The data were inverted using the algorithm
developed in this study. Consequently, the inverted time–space data
consisted of data acquired only after the first flooding of the
Table 1
conductive fluid. As the reference times for the reference models,
History of dye tracer flooding and tomography monitoring.
we selected the starting and ending times of the data measurement
Survey step Elapsed time (min) Description for each monitoring phase of a 2-D tomography survey. Hence, there
Phase 1 0–85 Before flooding were ten reference models for each 2-D tomography section.
Flooding 1 85–105 100 l flooding Since the inversion resulted in one space–time model in this study,
Phase 2 105–195
we could reconstruct the time-lapse images at any moment from the
Flooding 2 195–230 100 l flooding
Phase 3 230–315 inverted result. The time-lapse images in Fig. 7 clearly show the
Flooding 3 315–363 100 l flooding flooding zone by low resistivity. There were low and vertical resistivity
Phase 4 363–448 anomalies developing along the boreholes; these were mainly due to
Phase 5 1002–1087
the salt water that was injected into the boreholes to enhance the
Phase 6 1087–1585
electrode–earth coupling, just prior to the monitoring. A thin
530 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532

Fig. 8. The stain image taken over the vertically excavated tomographic section, Tomo4–5, after completing the dye tracer experiment. Grid pattern is overlapped to compare the
stained image to the images of resistivity monitoring, and the grid size is 0.1 × 0.1 m.

horizontal layer with high conductivity develops directly beneath the solute (Fig. 8). This implies that the migration of the conductive fluid
flooding zone, which we interpret as corresponding to the zone of and its path must be interpreted using difference images instead of the
homogeneous matrix flow. However, we cannot precisely identify the individual time-lapse images; thus, the precision of the difference
kinds of phenomena that have taken place by flooding the conductive images is very important.
fluid. Furthermore, the time-lapse images themselves do not show a Fig. 9 shows examples of the difference images using the 4-D
clear correlation with the stain image formed by the infiltration of dye inversion proposed in this study; Fig. 10 shows examples using the

Fig. 9. Resistivity changes of Phase 3 (a), Phase 4 (b), Phase 5 (c), and Phase 6 (d) with respect to Phase 2, which are expressed in terms of the resistivity ratio between each phase and
Phase 2. These difference images are computed using the 4-D inversion results, i.e., the snapshots of the subsurface conductivity distribution at the moment initiating the data
acquisition at each monitoring phase.
 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532 531

Fig. 10. Resistivity changes of Phase 3 (a), Phase 4 (b), Phase 5 (c), and Phase 6 (d) with respect to Phase 2, which are expressed in terms of the resistivity ratio between each phase
and Phase 2. These difference images were computed using the time-lapse subsurface models by individually inverting the time-lapse data based on the time-invariant static model.

conventional approach in which each monitoring data set is inverted

separately based on the static time-invariant earth model. Comparing
these two figures, it can be seen that our algorithm generates much more
reasonable difference images with little inversion artifact. Furthermore,
the resultant difference images more closely match the stain image in
Fig. 8 than those calculated by the conventional inversion approach.
Directly beneath the flooding zone, there is an uppermost horizontal
layer with a thickness of about 20 cm; its conductivity remains nearly
constant through Phase 4 and decreases afterwards. This implies that
this layer was already fully saturated by the first flooding, since all the
difference images in Fig. 9 are based on the resistivity distribution of
Phase 2, just after the flooding in Phase 1. This indicates that the
uppermost layer corresponds to the zone of the homogeneous matrix
flow. We can observe from Fig. 9 that the very conductive zone has been
developing downward to the left borehole, with a gentle dip and
widening until Phase 4. After Phase 4, the area of this highly conductive
zone is slightly reduced, but its conductivity is increased. From these
features observed in the difference tomograms, we surmise that the
conductive solute has been flowing mainly from the right to the left part
of the tomography section; this was interpreted as a lateral flow along a
sloping layer interface (Walter et al., 2000; Kim et al., 2004). The
difference image in Fig. 9a suggests that, at an early stage (Phases 2 and
3), the conductive fluid has migrated very quickly from the left side of
the zone of lateral flow, to the left borehole, and finally along the
borehole. This interpretation is consistent with the fact that the borehole
was drilled and artificially filled with soil, so that the borehole and its
adjacent part consist of a highly permeable medium.
Fig. 11 illustrates the subsurface change in resistivity ratio which
occurred during the data measurement. The figure shows clearly that
a significant change took place during the Phase 2 data acquisition,
while small changes occurred during Phase 6. Moreover, the resistivity Fig. 11. Resistivity changes occurring during the data measurement of Phase 2 (a) and
change during Phase 2 occurred mainly along and adjacent to the Phase 6 (b): difference images between the initial and final times of each phase.
532 J.-H. Kim et al. / Journal of Applied Geophysics 68 (2009) 522–532

borehole, in nearly the same pattern of change as it occurs between References

Phases 2 and 3, shown in Fig. 9a. All of these observations are quite
Alpak, F.O., Torres-Verdin, C., Habashy, T.M., 2004. Joint inversion of transient pressure
reasonable and verify that the new inversion approach can generate a and dc resistivity measurements acquired with in-situ permanent sensors: a
reasonable time–space model when the subsurface structure sig- numerical study. Geophysics 69, 1173–1191.
nificantly changes during the data measurement. Daily, W., Ramirez, A.L., 1995a. Electrical resistance tomography during in-situ
trichloroethylene remediation at the Savannah River Site. Journal of Applied
Geophysics 33, 239–249.
5. Conclusions Daily, W., Ramirez, A.L., 1995b. Electrical imaging of engineered hydraulic barriers.
Geophysics 65, 83–94.
Day-Lewis, F.D., Harris, J.M., Gorelick, S.M., 2002. Time-lapse inversion of crosswell
We developed a new least-squares inversion algorithm for radar data. Geophysics 66, 1740–1752.
simultaneously inverting entire data sets for geophysical monitoring, Kim, J.-H., 2005. Four dimensional inversion of dc resistivity monitoring data.
in which both the subsurface model and the monitoring data are Proceedings of Near Surface 2005, the 11′th European Meeting of Environmental
and Engineering Geophysics. European Association of Geoscientists and Engineers,
defined in a 4-D space–time domain. The algorithm was designed to
p. A006.
effectively reduce inversion artifacts and to provide reasonable Kim, J.G., Chon, C.-H., Lee, J.-S., 2004. Effect of structures and texture on infiltration flow
subsurface images even if changes in the subsurface properties during pattern during flood irrigation. Environmental Geology 46, 962–969.
the data acquisition cannot be ignored. Using numerical experiments LaBrecque, D.J., Yang, X., 2001. Difference inversion of ERT data: a fast inversion method
for 3-D in situ monitoring. Journal of Environmental and Engineering Geophysics 6,
of resistivity tomography monitoring, we demonstrated that the new 83–89.
inversion approach can generate difference images with little Loke, M.H., 1999. Time lapse resistivity imaging inversion. Proceedings of the 5th
inversion artifact, thus providing precise information on the location Meeting of the Environmental end Engineering European, Em1.
Slater, L., Binley, A.M., Daily, W., Johnson, R., 2000. Cross-hole electrical imaging of a
and amount of ground condition changes. We also demonstrated that controlled saline tracer injection. Journal of Applied Geophysics 44, 85–102.
the history of ground condition change can be reasonably reproduced Vargemezis, G.N., Tsourlos, P., Kim, J.-H., Yi, M.-J., 2007. Application of Electrical
using just one set of time-lapse data, in cases where the subsurface Resistivity Tomography in Monitoring Recycled Water Injection. Proceedings of
Near Surface 2007, Istanbul, Turkey, 3–5 September.
material properties continuously and significantly change in time. The Walter, M.T., Kim, J.S., Steenhuis, T.S., Parlange, J.Y., Heiling, A., Braddock, R.D., Boll, J.,
algorithm is also applicable to field data, as was demonstrated by 2000. Funneled flow mechanism in a sloping layered soil: laboratory investigation.
applying it to the monitoring data of crosshole resistivity tomography Water Resources Research 36, 841–849.
Yi, M.-J., Kim, J.-H., Chung, S.-H., 2003. Enhancing the resolving power of least-squares
jointly performed with a dye tracer flooding experiment. The scale of
inversion with active constraint balancing. Geophysics 68, 931–941.
the experiment was so small that we could not ignore the change of
material properties during the data measurement. Although we have
demonstrated our approach with only resistivity tomography in this
paper, we anticipate that the proposed algorithm will be applicable to
the inversions of a variety of other geophysical monitoring data.

Acknowledgements

This research was supported by the Basic Research Project of the

Korea Institute of Geoscience and Mineral Resources funded by the
Ministry of Science and Technology of Korea.