You are on page 1of 12

Journal of Contaminant Hydrology 251 (2022) 104097

Contents lists available at ScienceDirect

Journal of Contaminant Hydrology


journal homepage: www.elsevier.com/locate/jconhyd

Evaluating the effectiveness of a geostatistical approach with groundwater


flow modeling for three-dimensional estimation of a contaminant plume
Shizuka Takai a, b, *, Taro Shimada a, Seiji Takeda a, Katsuaki Koike b
a
Nuclear Safety Research Center, Japan Atomic Energy Agency, 2-4 Shirakata, Tokai, Ibaraki 319-1195, Japan
b
Department of Urban Management, Graduate School of Engineering, Kyoto University, Katsura C1-2-215, Kyoto 615-8540, Japan

A R T I C L E I N F O A B S T R A C T

Keywords: When assessing the risk from an underground environment that is contaminated by radioactive nuclides and
Contaminant plume estimation hazardous chemicals and planning for remediation, the contaminant plume distribution and the associated un­
Geostatistical inversion certainty from measured data should be estimated accurately. While the release history of the contaminant
Gibbs sampling
plume may be unknown, the extent of the plume caused by a known source and the associated uncertainty can be
Groundwater contamination
Groundwater flow
calculated inversely from the concentration data using a geostatistical method that accounts for the temporal
correlation of its release history and groundwater flow modeling. However, the preceding geostatistical ap­
proaches have three drawbacks: (1) no applications of the three-dimensional plume estimation using concen­
tration data from multiple depths in real situations, (2) no constraints for the estimation of the plume
distribution, which can yield negative concentration and large uncertainties, and (3) few applications to actual
cases with multiple contaminants. To address these problems, the non-negativity constraint using Gibbs sampling
was incorporated into the geostatistical method with groundwater flow modeling for contaminant plume esti­
mation. This method was then tested on groundwater contamination in the Gloucester landfill in Ontario,
Canada, using three-dimensional contaminant transport model and concentration data from multiple depths. The
method was applied to three water soluble organic contaminants: 1,4-dioxane, tetrahydrofuran, and diethyl
ether. The effectiveness of the proposed method was verified by the general agreement of the calculated plume
distributions of the three contaminants with concentration data from 66 points in 1982 (linear correlation co­
efficient of about 0.7). In particular, the reproduced peak of 1,4-dioxane corresponding to the large disposal in
1978 was more accurate than the result of preceding minimum relative entropy-based studies. The same peak
also appeared in the tetrahydrofuran and diethyl ether distributions approximately within the range of the
retardation factor derived from the fraction of organic carbon.

1. Introduction underlying physical processes that govern contaminant transport and


the contaminant plume distribution must be correctly understood.
Over recent decades, the underground environment in various areas There are three main challenges when developing practical methods
worldwide has been contaminated by anthropogenic chemical, radio­ to estimate the extent and concentrations of actual contaminant plumes.
logical, and microbiological substances. There are numerous reports of First, the amount and location of measured concentration data may be
contamination of groundwater by substances that migrate easily into limited because of temporal and budgetary constraints. If there are
groundwater. Examples include contamination by volatile organic enough data, the plume distribution can be estimated by interpolation or
compounds, heavy metals, and radionuclides such as 3H and 90Sr in extrapolation of the data using a spatial estimation method, such as
nuclear facilities including Hanford (Washington, USA), Sellafield kriging. To date, several types of kriging, such as ordinary and indicator
(United Kingdom), and Chalk River (Canada) (OECD, 2014). To assess kriging, cokriging (for multivariate data), and spatio-temporal kriging
the risk of groundwater contamination and support plans for reme­ have been applied in various settings (e.g., Saby et al., 2006; Juang et al.,
diating contaminated sites, the hydrogeological structures and 2008; Iskandar et al., 2012; Antunes and Albuquerque, 2013; Boente

* Corresponding author at: Nuclear Safety Research Center, Japan Atomic Energy Agency, 2-4 Shirakata, Tokai, Ibaraki 319-1195, Japan.
E-mail addresses: takai.shizuka@jaea.go.jp (S. Takai), shimada.taro@jaea.go.jp (T. Shimada), takeda.seiji@jaea.go.jp (S. Takeda), koike.katsuaki.5x@kyoto-u.ac.
jp (K. Koike).

https://doi.org/10.1016/j.jconhyd.2022.104097
Received 27 March 2022; Received in revised form 6 October 2022; Accepted 16 October 2022
Available online 19 October 2022
0169-7722/© 2022 Elsevier B.V. All rights reserved.
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

et al., 2017; Dalla Libera et al., 2017; Liang et al., 2018; Masoudi et al., sampling over power transformation has been verified by a case
2021). However, these spatial estimation methods may not be suitable study of the atmospheric inverse modeling of trace gas emissions
for accurately reproducing the plume distribution in groundwater (Miller et al., 2014). Such study demonstrated the great flexibility
because the monitoring wells are not always optimally sited to allow the of the Gibbs sampler in determining the shape of the posterior pdf
delineation of plume structures formed by advection, dispersion, and and its computational efficiency. So far, however, the Gibbs
sorption along with the groundwater flow (Shlomi and Michalak, 2007). sampling method has not been applied to estimate plumes in
Therefore, flow and transport modeling is indispensable for reproducing groundwater flow.
the plume distribution in a groundwater system. Second, the contami­ (3) Although contamination is commonly caused by multiple con­
nant source (source location or release history) for accidental contami­ taminants, most studies tackle only one contaminant such as 1,4-
nation incidents is usually not identified, and release history records are dioxane (14DNE; Michalak and Kitanidis, 2002) and hexa­
particularly rare. For example, when remediating nuclear sites in the chlorocyclohexane (Gyzl et al., 2014). Michalak and Kitanidis
past, source locations have been detected by preliminary surveys or (2003) studies the release histories of two contaminants, tetra­
historical site assessments (U.S. EPA, 2000) in drain lines, sumps, pipes, chloroethene (PCE) and trichloroethene (TCE), but did not
and tanks, and the transient release histories have been unknown in all discuss the consistency between the results of the two contami­
cases (OECD, 2014). In such a situation, either the plume cannot be nants. Therefore, we are not sure if it is feasible to analyze mul­
estimated from only the flow and transport simulation or the uncertainty tiple components in the same study.
associated with the estimation will be large. Third, risk assessments
should include a quantitative evaluation of the uncertainty of the plume To address these problems, the non-negativity constraint using Gibbs
estimation. The uncertainty of the estimated plume distribution can aid sampling was incorporated into the geostatistical method with
in determining the number and location of new measurement points groundwater flow modeling for estimating the distribution of contami­
(Snodgrass and Kitanidis, 1997). nant plume and the associated uncertainty. The method was then tested
To overcome these issues, a geostatistical approach that incorporated on groundwater contamination in the Gloucester landfill in Ontario,
groundwater flow and transport modeling was proposed to estimate the Canada, using three-dimensional contaminant transport model and
distribution of a contaminant plume and the associated uncertainty multiple concentration data in 1982. The method was applied to three
(Shlomi and Michalak, 2007). This approach assumes that the source water soluble organic contaminants: 14DNE, tetrahydrofuran (THF),
locations are known but the release history is unknown. The plume and diethyl ether (DEE). The research flows from an estimation of the
distribution is calculated from the release history estimated by geo­ release history of a conservative tracer, 14DNE, and its comparison with
statistical inversion that accounts for its temporal covariance (e.g., the known large disposal in 1978; reproduction of the estimated peak of
Snodgrass and Kitanidis, 1997). If groundwater contamination is a result 14DNE by calculating the retardation factors of THF and DEE and
of multiple actions of several parties, the reproduced release history can verifying the factors using fraction of organic carbon; to three-
assist in determining how to divide the remediation cost among the dimensional plume estimations of the three contaminants and their
responsible parties (Skaggs and Kabala, 1994; Michalak, 2002; Michalak verification by comparing the obtained values with concentration data
and Kitanidis, 2002; Butera and Tanda, 2003). However, there are three from 66 points.
problems with the preceding geostatistical approaches:
2. Methods
(1) Although actual contamination is three-dimensional, the pre­
ceding geostatistical approaches have so far been applied to es­ 2.1. Geostatistical inversion
timate contamination and release histories in one dimension
along depth profiles (e.g., Michalak and Kitanidis, 2003) or two- The contaminant plumes were calculated with the inverse/forward
dimensional aquifer modeled with one layer in which a homo­ modeling method of Shlomi and Michalak (2007), which uses a known
geneous vertical contaminant distribution is assumed (e.g., source location and a steady groundwater flow for the transport of
Butera and Tanda, 2003; Gyzl et al., 2014). This problem is contaminant materials. The source intensity is estimated at a time tj (j =
common to geostatistical and other stochastic approaches such as 1, ⋯, mt) using the measured concentration data z* ∈ ℝn×1 (n-dimen­
minimum relative entropy and backward probability tracking, sional real space) at a sample location xi (i = 1, ⋯, n) and time tmes (>tmt).
because most preceding studies (e.g., Colombo et al., 2020) have Assuming steady flow, z* can be related linearly to the release history, s
generally been based on two-dimensional aquifer models ∈ ℝmt×1, as
(Gomez-Hernandez and Xu, 2022).
(2) The concentrations that are produced from plume estimations z* = H * s + v, (1)
can include negative values and large uncertainties because of a
lack of constraints. Negative values frequently appear in practical where H* ∈ ℝn×mt is the Jacobian matrix and v ∈ ℝn×1 is the measure­
applications where the problem is ill-conditioned and has large ment and model error. This linear relationship is illustrated schemati­
errors. While negative values can be avoided by transforming cally in Fig. 1(a). Because the Jacobian matrix can be defined in
data with a power law (Box and Cox, 1964) (e.g., Snodgrass and advance, typically by simulating the groundwater flow, the contaminant
Kitanidis, 1997; Gyzl et al., 2014), data transformation can plume distribution can be calculated forward over an area by applying
change linear inversion problems to nonlinear, and can yield inverse analysis of the unknown release history s with limited data z*
highly asymmetric probability density functions for the estimated and the linear system of Eq. (1).
parameter values in the untransformed space that are unrealistic The release history can be estimated by geostatistical inversion
in some cases (Michalak and Kitanidis, 2003). Alternative ap­ methods, as demonstrated earlier (e.g., Kitanidis and Vomvoris, 1983;
proaches include the Markov chain Monte Carlo (MCMC) Kitanidis, 1995), where s and v were assumed to be random vectors.
methods, which use Bayes' rule to constrain the posterior prob­ Suppose that s is a multivariate Gaussian distribution with a structural
abilistic density function (pdf) (Michalak and Kitanidis, 2002, parameter θ:
2003, 2004a, 2004b, 2005; Michalak, 2008). Michalak (2008) s = Xs βs + ε, ε ∼ N (0, Qs (θ) ), (2)
showed that an application of the non-negativity constraint using
Gibbs sampling, an MCMC method, improved the accuracy and where Xsβs is the drift (i.e., the expected value of s), Xs ∈ ℝmt×ps is a
precision of the estimates of the arsenic release history in the known matrix of basis functions, βs ∈ ℝps×1 are ps unknown drift co­
North Fork of the Humboldt River. The superiority of Gibbs efficients, and Qs(θ) ∈ ℝmt×mt is a covariance matrix of s. On the basis of

2
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Fig. 1. Illustration of (a) the linear relationship between the measured concentrations and the source intensity and (b) the steps taken to estimate the release history
(step 1) and the plume distribution (step 2) by geostatistical inversion.

these settings, the unknown release history is estimated using the described as follows:
following two steps. In the first step, the optimal θ is specified . The
̂s = Λs z* , (10)
posterior pdf p(s| z*) can be proportional to the product of the likelihood
p(z*| s) and the prior pdf of p(s) as in Bayes' rule:
V̂s = Qs − Qs H * ΛTs − Xs Ms . (11)
T

* *
p(s|z )∝p(z |s)p(s). (3)
Once ̂s and V̂s are solved, the estimated plume distribution over a
Assuming that v~N (0, R), where R∈ ℝn× is an error covariance
n

study area ̂z ∈ ℝm×1 and its uncertainty V̂z can be solved:


matrix of z*, p(s) and p(z*| s) can be expressed as:
1
[
1
] ̂z = Ĥs , (12)
p(s) = mt /2
|Qs |− 1/2 exp − (s − Xs βs )T Qs − 1 (s − Xs βs ) , (4)
(2π) 2
V̂z = HV̂s H T , (13)
[ ]
1 1 *
p(z* |s) = |R|− 1/2
exp − (z − H * s)T R− 1 (z* − H * s) . (5) where H ∈ ℝn×m is the Jacobian matrix and m is the number of esti­
(2π)n/2 2
mation points. The steps for optimal estimation of the release histories
The optimal structural parameter θ is estimated using a restricted and plume distributions are illustrated in Fig. 1(b).
maximum likelihood approach which maximize the following marginal In most cases, the release history is unknown and the drift shape and/
∫ ( ∫ ⃒ )
likelihood L (θ) = p(z∗ |βs , θ)dβs   p(z∗ ) = p(z∗ ⃒s)p(s)ds , in which or the covariance of ̂s are difficult to identify and determine, as in the
the unknown βs is eliminated by integrating. This is equivalent to present case study. Accordingly, the simplest linear trend (Xs = [1 t])
minimizing L(θ) (the negative log of L (θ)): and a cubic generalized covariance (Qs(ti, tj| θ) = θ|ti − tj|3) were used,
following Kitanidis and Lee (2014). This was done by assuming that the
1 1 ⃒ ⃒ 1 T
best ̂s estimate is continuously differentiable and smooth, that is,
L(θ) = ln|Σ| + ln⃒Xs T H * Σ − 1 H * Xs ⃒ + z* Ξ − 1 z* , (6)
T

2 2 2 equivalent to cubic spline interpolation (Dubrule, 1984). While R con­


tains measurement and model errors, only the uncorrelated measure­
Σ = H * Qs H * + R, (7)
T
ment error was considered simply as R = σR 2 I, where σ R 2 is the variance
( )− 1 of measurement errors and I ∈ ℝn×n is an identity matrix.
Ξ = Σ − 1 − Σ − 1 H * Xs Xs T H * Σ − 1 H * Xs Xs T H * Σ − 1 . (8)
T T

The optimal θ is calculated iteratively using the Fisher scoring


2.2. Non-negativity constraint using Gibbs sampling
method (Kitanidis, 1995). Then, the unknown release history ̂s is esti­
mated. By maximizing the posterior pdf p(s| z*), the system to be solved
The non-negativity release history can be achieved by constraining
is described as:
p(s| z*) to be non-negative using Gibbs sampling. The sampling can be
( )( T ) ( * ) executed sequentially from the marginal (one-dimensional) conditional
Σ H * Xs Λs H Qs
= , (9) distribution because the unconstrained pdf is explicitly derived by a
(H * Xs )T 0 Ms XsT
product of multidimensional Gaussian distributions, as described in
detail by Michalak (2008). The marginal posterior pdf p(si| z*, s) can be
where Λs ∈ ℝmt×n is a weight matrix and Ms ∈ ℝps×mt is a Lagrange
expressed using Bayes' rule:
multiplier. The form of Eq. (9) is equivalent to that of universal kriging
(or kriging with trend), which is frequently used as a spatial estimator in p(si |z* , s)∝p(z* |si , s)p(si |s)H (si ) (i = 1, ⋯, mt ), (14)
geostatistics. The best estimate and posterior covariance of ̂s are

3
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097


⎨ 1 si > 0 calculated to be infinite because μ was too low (negative) and τ was
H (si ) = 1/2 si = 0 (15) narrow, the result was set to zero.

0 si < 0
3. Site description and physical model
where si means s(ti), and H (si ) is the Heaviside function. Based on a
multidimensional Gaussian distribution of the likelihood of the mea­ 3.1. Background information about the Gloucester landfill
surements p(z*| s) in Eq. (5), the marginal likelihood p(z*| si, s) can be
expressed as: The Gloucester landfill in Ottawa, Ontario, Canada, was a municipal
1
(
1 (si − μL )
) and industrial landfill where hazardous wastes containing organic
p(z* |si , s) = √̅̅̅̅̅̅̅̅̅̅exp − , (16) wastes were disposed of from 1969 to 1980 in a special waste compound
2πτL 2 τL
(SWC) along the western edge of the landfill that extended across
aT b σR 2 approximately 3000 m2. The organic wastes were principally organic
μL =
bT b
, τL = T ,
b b solvents (e.g., chloroform, carbon tetrachloride, methanol, acetone,
benzene, and hexane). Soluble organic solvents such as 14DNE caused

mt widespread contamination in the aquifer above the top of the limestone
a j = z* j − Hjk sk , bj = Hji . (17) bedrock at between 25 and 30 m deep (Fig. 2(a)) (e.g., Jackson et al.,
1985; Patterson et al., 1985; Jackson et al., 1989). While there is very
k=1,k∕
=i

Then, the marginal prior pdf p(si| s) is defined as: little information about the release history, records show that approxi­
( ) mately 1 ton of organic solvents (1 ton = 1000 L aqueous solution) was
1 1 (si − μP )2 disposed of in May 1978 (Jackson et al., 1985).
p(si |s) = √̅̅̅̅̅̅̅̅̅̅exp − , (18)
2πτP 2 τP The site sits on top of a complex sequence of Quaternary fluvioglacial
and coastal deposits (Fig. 2(b)) that was classified into five units
where μP and τP are obtained as the solution of the following system for (Jackson et al., 1985). In ascending order, there were a limestone
si: bedrock unit (Unit A); a basal till unit up to 2 m in thickness (Unit B); a
( )( ) ( ) thick confined, outwash aquifer up to 25 m in thickness composed of
Qs k∕=i,k∕=i Xsk∕=i Qs k∕=i,i
Λ
= , (19) interstratified and interfingering silts, sands, and poorly sorted gravels
Xs Tk∕=i 0 − M Xs Ti (Unit C); a clayey silt and silt unit acting as an aquitard or a confining
layer (Unit D); and surface deposits of fine sands to gravels up to 10 m in

m t− 1 ∑
m t− 1 ps
∑ thickness, containing unconfined groundwater at several meters in
μP = Λj sj , τP = Qs ii − Λj Qs ji + Xsil Ml . (20)
j=1 j=1 l=1
depth (Unit E). The contaminants laterally flowed down to the confined
aquifer from the SWC along the topography, approximately from west to
Finally, the marginal posterior pdf p(si| z*) can be expressed by a east (Fig. 2(a)) (MacFarlane et al., 1983; Jackson et al., 1985).
truncated-Gaussian distribution: Of the contaminants in the waste, there was particular concern about
( ) three compounds (14DNE, THF, and DEE) because of their ability to
1 1 (si − μ)2
p(si |z* )∝√̅̅̅̅̅̅̅̅exp − H (si ), (21) migrate rapidly in aquifers (Lessage et al., 1990). 14DNE exhibits the
2πτ 2 τ greatest mobility because of its low KOW and high solubility (Table 1).
( ) Because 14DNE can be regarded as a conservative tracer, two previous
τ=
τP τ L μ μ
,μ = τ P + L . (22) studies estimated its release history three-dimensionally, using the
τP + τL τP τL minimum relative entropy (MRE) approach (Woodbury et al., 1998) and
To summarize the above calculations, the non-negativity- a geostatistical inversion with a non-negativity constraint based on a
constrained release history, contaminant plume, and their uncertainty Metropolis–Hasting algorithm (Michalak and Kitanidis, 2002). Howev­
can be obtained as follows: er, the estimations of plume distribution and the related uncertainty
were out of scope of these studies.
Here, the plume distributions of the three contaminants and the
1. Set initial values as the unconstrained solution ̂s cc 0 =
associated uncertainties were quantified and assessed, with comparing
(̂s uc (t1 ) , ⋯, ̂s uc (tmt ) )T .
the estimated release history of 14DNE with those from previous studies.
2. Draw a conditional realization ̂s cc l+1 from ̂s cc l at each time step by: The estimation comprised three steps. The first step was to estimate the
( ) release history of 14DNE and compare the result with those from pre­
̂s cc l+1 (t1 ) ∼ p ̂s cc l (t1 ) |̂s cc l (t2 ) , ⋯, ̂s cc l (tmt ) ,
l+1 ( l l+1 l ) ceding studies that used the same flow and transport model. Next, the
̂s cc (t2 ) ∼ p ̂s cc (t2 ) |̂s cc (t1 ) , ⋯, ̂s cc (tmt ) , (23) retardation factors of THF and DEE were specified by matching with the

l+1 ( l l+1
̂s cc (tn ) ∼ p ̂s cc (tn ) |̂s cc (t1 ) , ⋯, ̂s cc (tmt − 1 ) ,l+1 ) release peak of 14DNE. Finally, the three-dimensional plume distribu­
tions and the associated uncertainties for the three contaminants were
estimated.
and calculate ̂z cc l+1 = Ĥs cc l+1 .

3. Repeat step 2 until the calculation is converged (l = ls). 3.2. Contaminant transport in groundwater
4. Calculate the best estimate and 95% confidence interval (difference
(
between the upper and lower limits) of ̂s cc and ̂z cc from ̂s cc lburn +1 , ⋯ Contaminant transport is generally modeled using the advec­
) ( ) tion–dispersion equation. For three-dimensional contaminant transport
, ̂s cc ls and ̂z cc lburn +1 , ⋯, ̂z cc ls . The first lburn steps until the algorithm
in porous media with a steady, uniform groundwater flow along the x-
reaches convergence are discarded as a burn-in period. Convergence
axis, the transport is described as:
is evaluated by tracking the mean values of likelihood and prior
distribution, as in Michalak (2008). ∂c ∂c ∂2 c ∂2 c ∂2 c
Rf + Vx − Dx 2 − Dy 2 − Dz 2 = − λc (24)
∂t ∂x ∂x ∂y ∂z
*
Sampling from p(si| z ), a truncated-Gaussian distribution, was
implemented based on Johnson et al. (1994). If the sampled value was where c is the concentration; t is the time; Vx is the actual groundwater
velocity; Dx, Dy, and Dz are the dispersion coefficients in the x

4
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Fig. 2. (a) Plan view of the study area


(Gloucester landfill in Ottawa, Ontario,
Canada) and (b) geological cross-section
showing the distribution of the five
stratigraphic units along the cross-
section (red line) (modified from Jack­
son et al., 1985). The locations of the
contaminant source (special waste
compound), multilevel samplers with
the numbers used in Jackson et al.
(1985), and assumed groundwater flow
direction are indicated. (For interpreta­
tion of the references to colour in this
figure legend, the reader is referred to
the web version of this article.)

z1 and z2 are the top and bottom boundary depths of the source, and y0 is
Table 1
the half-width of the source. For a discretized time of the release history,
Solubilities and octanol-water partition coefficients of 14DNE, THF, and DEE.
t (0 < t < T, T is the measurement time), the analytical solution of the
Contaminant* 14DNE THF DEE Jacobian matrix H is expressed as:
Molecular formula C4H8O2 C4H8O (C2H5)2O [ ]
Solubility+ Miscible (> 8.0 × 105) Miscible (1 × 106) 6.9 × 104 x(z2 − z1 ) 1 (x − Vx (T − t) )2
H(x, y, z, T − t) = √ ̅̅̅̅̅̅̅̅ exp − − λRf (T − t)
logKOW++ − 0.27 0.46 0.89 4B πDx (T − t)1.5 4Dx (T − t)
+
Solubility in water (mg/L) (Yalkowsky et al., 2010; Wishart et al., 2007;
ICSC, 2017). [ ( ) ( )]
y − y0 y + y0
++
Log of octanol-water partition coefficient (Hansch et al., 1995; Brooke • erfc ( )1/2 − erfc ( )1/2
et al., 1998). 2 Dy (T − t) 2 Dy (T − t)
*
14DNE: 1,4-dioxane; THF: tetrahydrofuran; DEE: diethyl ether.
[ ]
x 1 (x − Vx (T − t) )2
+ √ ̅̅̅̅̅̅̅̅ exp − − λR (T − t)
(longitudinal), y (transverse horizontal), and z (vertical) directions, π1.5 4Dx (T − t)1.5 4Dx (T − t)
f

respectively; Rf is the retardation factor; and λ is the first-order


biochemical decay constant. To facilitate comparison, the contaminant [ ( ) ( )]
y − y0 y + y0
transport model with a three-dimensional analytical solution used in this • erfc ( )1/2 − erfc ( )1/2
study was the same as those used in the previous studies discussed 2 Dy (T − t) 2 Dy (T − t)
earlier. We assume that a contaminant is sourced in a rectangular area in [ ( ) ( )] ( ) [ 2 2 ]
∑∞
the yz plane and transported from there in an aquifer. The extent of •
1
sin
lπz2
− sin
lπz1
cos
lπz
exp −
l π Dz (T − t)
. (26)
aquifer is semi-infinite, infinite, and finite with thickness B in the x-, y-, l=1
l B B B B2
and z-directions, respectively. The initial and boundary conditions of the
The values of the parameters in Eq. (26) used in the two previous
solution are as follows:
studies discussed earlier, derived from the analysis of reported data by
c(x, y, z, 0) = 0, Jackson et al. (1985), were used in this study (Table 2). The top
{
s(t) ( − y0 ≤ y ≤ y0 z1 ≤ z ≤ z2 ) boundary of the calculation domain was set to the ground surface
c(0, y, z, t) = ,
0 (otherwise) because the water table, the aquitard, and the top of the bedrock were
approximately parallel. The top and bottom depths of the confined
c(∞, y, z, t) = 0, c(x, ±∞,z, t) = 0, outwash aquifer were assigned to z1 and z2, respectively. Woodbury
∂c ∂c⃒
⃒ (25) et al. (1998) calculated the transport velocity Vx on the basis of an
= 0, ⃒⃒ =0
∂zx,y,B,t ∂z x,y,0,t analysis of the center of mass of the plume, which was generally
consistent with the reported values: A borehole dilution survey in a
where z is the depth from the ground surface, s(t) is the release history, single piezometer showed that the mean groundwater velocity in the

5
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Table 2
Parameter values used for the contaminant transport model using the advection–dispersion equation.
Parameter Vx (m/d) αx (m) αy (m) αz (m) Dm (m2/d) λ (/d) y0 (m) z1 (m) z2 (m) B (m)

Value 0.127 2.6 0.5 0.1 0.0 0.0 25.0 6.0 24.0 24.0

target area was 0.07 m/day, while a pumping test showed it was 0.15 m/ significantly change the original depth–profile trend (Fig. 4).
day (Jackson et al., 1985, 1989). The dispersivity coefficients can be The measured concentration data were shown by the least significant
expressed as (Dx, Dy, Dz) = (Dm + Vxαx, Dm + Vxαy, Dm + Vxαz), where Dm digit of 0.1 (μg/L), but Jackson et al. (1985) did not provide information
is the molecular diffusion coefficient and αx, αy, and αz are the longitu­ about data precision. Assuming that this digit included uncertainty and
dinal, transverse and, vertical dispersivities, respectively. The values of the measurement errors were normally distributed, the error variance
the dispersivity coefficients were defined from the plume moments and was assigned by σR 2 = 0.25 (μg/L)2, to correspond with the 95% confi­
by referring to classic theory (Domenico and Schwartz, 1990), and dence interval of ±1 (μg/L) (Michalak, 2008). Meanwhile, the detection
corresponded to the estimated values of (αx, αy) = (2.0, 1.5) (m) by limit of 14DNE using the conventional purge-and-trap coupled with GC/
calibrating with the hydraulic head and concentration data (Gailey MS was reported to range from 3 to 20 (μg/L) because of the poor purge
et al., 1991). The λ value was set to 0. The start date of the special waste efficiency (Issacson et al., 2006; Mohr, 2001; Draper et al., 2000).
activities, May 1, 1969, was set to t = 0. The release history was Therefore, we have also considered the large measurement error for
recovered until the beginning of the measurement period, September 14DNE in our discussion, i.e., σR= 3.0 (μg/L), assuming the detection
1982, or t = 4870 days. limit was 10 (μg/L) and this limit was equal to three times σR (Loconto,
The following empirical relationship can be used to define Rf: 2001). The time interval was one day. The Gibbs sampling was set to
/ 5000 realizations and the first 100 realizations were regarded as the
Rf = 1 + ρb • Kp ε, Kp = fOC • KOC , (27)
burn-in period. The release histories and plume distributions were
calculated for the rectangular area covering x = [0,300] (m) × y =
where ρb is the soil bulk density, ε is the porosity (0–1), Kp is the par­
[− 150, 150] (m) as shown in Fig. 2(a).
titioning coefficient, fOC is the fraction of organic carbon, and KOC is the
organic carbon partitioning coefficient related to the octanol–water
4. Results
partitioning coefficient, KOW, derived from the study of the sorption of
chlorinated benzenes for sediments with an fOC > 0.001 (Schwarzenbach
4.1. Release history of 14DNE
and Westall, 1981):
logKOC = 0.72logKOW + 0.49. (28) For the recovered release history of 14DNE, the value of θ used in the
cubic covariance was 1.3 × 10− 3 (mg2/L2/day3). The resulting release
The Rf values were derived as shown in Table 3 by setting ρb = 2000
histories without a non-negativity constraint and those constrained
kg/m3, and ε = 0.35 based on the values of 0.3 to 0.35 for the outwash
using the Gibbs sampling are compared in Fig. 5(a) and (b). It is worth
aquifer (Jackson et al., 1985), and using the KOW values of each
noting that the peak corresponding to the large disposal in May 1978
contaminant (Table 2). The resulting values were the same as those of
was clearly reproduced by the non-negativity constraint. The peak
Jackson et al. (1985). Because the Rf of 14DNE was close to 1 regardless
concentration was much lower than the solubility limit >8 × 105 mg/L
of fOC, the sorption of 14DNE was not considered as in earlier studies
(Yalkowsky et al., 2010). It is also worth noting that the 95% confidence
(Woodbury et al., 1998; Michalak and Kitanidis, 2002). In contrast, the
interval was almost equivalent to the best estimate, and supports the
Rf values of THF and DEE depended on fOC. Therefore, we calculated the
accuracy of the inversion result. In contrast, the best estimate without
value of Rf, by reproducing the estimated peak of 14DNE and verified it
the non-negativity constraint contained unrealistic negative concentra­
against a value range derived from the fOC.
tions. In addition, the 95% confidence interval was much larger than the
best estimate by a factor of two orders (e.g., the peak concentration of
3.3. Measurement and estimation of the contaminant plume the best estimate was 0.4 mg/L and had a confidential interval of
approximately 206 mg/L), which meant that the uncertainty was very
We used the concentration data collected by Jackson et al. (1985). At large and the accuracy of estimates was low.
that time, the organic compounds in the groundwater were measured Comparison of the recovered release histories showed that the pro­
using bundle-type multilevel samplers (Cherry et al., 1983) with screens posed method was superior to the MRE approach of Woodbury et al.
set at depth intervals of approximately 1 m (Nos. 14, 17, 21, 34, 37, 38, (1998) that assumed Gaussian or uniform prior distributions (Fig. 5(c)).
39, 47, and 54; Fig. 2) or 4 m (Nos. 12, 18, 33, 40, 52, and 56) along 15 When assuming the Gaussian prior, the latter peak in the two estimated
observation wells. The organic compounds in the groundwater samples peaks corresponded to the large disposal in 1978 (Woodbury et al.,
were analyzed using gas chromatography/mass spectrometry (Finnigan 1998). However, the resulting release histories were strongly dependent
4000 GC/MS system). From all the data, 176 points measured from on the prior distribution. In contrast, we were able to reproduce the
September to December 1982 were selected (the blank was treated as large spill without any prior information, which is an advantage of
zero, as in Jackson et al., 1985). Then, the 1-m interval data were geostatistical inversion (Michalak and Kitanidis, 2002).
adjusted to the 4-m interval data by averaging the original data over The results presented in this paper differed from those presented by
each 4-m depth range, to account for the variation in the sample location Michalak and Kitanidis (2002), even though the geostatistical inversion
distributions and measurement time intervals. The distributions of the approach was used. This is mainly because they also used the data from
66 data points that resulted are shown in Fig. 3. The averaging did not 1983 onwards; the difference in measurement data results in a different
setting of the covariance structure (θ) and the error (σR). The ground­
water flow and contaminant conditions may be changed after 1983 by
Table 3
Definitions of the retardation factor (Rf) values of the three contaminants using the GTC purge well test in October 1983 (Jackson et al., 1985), soil
the two fractions of organic carbon (fOC). remediation from 1987 to 1989, and contaminant removals from the
deep and shallow groundwater plume after April 1992 (Franz Environ­
Contaminant 14DNE THF DEE
mental Inc., 2007). Therefore, calculations using the data after 1983
Rf (fOC = 0.0035) 1.0 1.1 1.3 under steady-flow conditions can be considered misleading. In addition,
Rf (fOC = 0.01) 1.1 1.4 1.8
Michalak and Kitanidis (2002) constrained the non-negativity with the

6
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Fig. 3. Distribution of groundwater concentration data from 1982 for the three contaminants: (a) 14DNE, (b) THF, (c) DEE.

Fig. 4. Examples of the measured and averaged concentration data for the three contaminants at depth intervals of 4 m at the three wells. The solid and dashed lines
denote the original 1 m interval data and the averaged concentrations, respectively (Jackson et al., 1985).

Fig. 5. The release histories of 14DNE recovered using geostatistical inversion (a) with no constraint and (b) with a non-negativity constraint. The red and dashed
black lines denote the best estimate (BE) and the 95% confidential interval (95% CI), respectively. (c) Comparison of the best estimates with those of two previous
studies using a minimum relative entropy approach (Woodbury et al., 1998) and geostatistical inversion using the Metropolis–Hasting algorithm (Michalak and
Kitanidis, 2002). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Metropolis–Hasting algorithm, an MCMC method. The Gibbs sampling is the constrained solution can be obtained efficiently using the Gibbs
considered superior to the Metoropolis–Hasting algorithm (the random sampling. Previously, residual variance values of 22.5 (μg/L)2 (Wood­
walk procedure containing the acceptance test) because of its compu­ bury et al., 1998) and 12.3 (μg/L)2 (Michalak and Kitanidis, 2002) were
tational efficiency and shorter computation time. If the posterior pdf for reported. Only a rough comparison is possible because of the difference
each time can be calculated using Eq. (21) in advance, as in this study, in the used data; however, the residual variance of our result was 1.07

7
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

(μg/L)2, which was approximately on the same scale as that of the error at around 1 × 10− 3 (mg/L) for THF and DEE (Fig. 7(b)). These data were
setting σ R 2 = 0.25 (μg/L)2. collected from shallow depth (z = 4 to 12 m) at sampling point No. 34
(Fig. 2(a)), which is located in the stratified low permeability zone of the
confined aquifer (Unit C, shown in Fig. 2(b)). Such unit-dependent dif­
4.2. Verification of the retardation factors of THF and DEE
ferences in permeability were not considered in the analytical solution,
and so may have caused the overestimation. This is considered in more
The retardation factors of THF and DEE estimated by accounting for
detail in the discussion.
the fraction of organic carbon (Table 2) were verified using the repro­
duced release peak of 14DNE. The release histories of THF and DEE
5. Discussion
(Fig. 6) were reproduced using the optimal θ of 1.8 × 10− 3 and 2.5 ×
10− 3 (mg2/L2/day3). The peaks of THF (25 mg/L) and DEE (77 mg/L)
5.1. Effect of averaging the measurement concentrations
are much lower than solubility limits of THF (1.0 × 106 mg/L) (Wishart
et al., 2007) and DEE (6.9 × 104 mg/L) (ICSC, 2017). The retardation
As described earlier, the original concentration data at the 176 points
factors were determined as 1.75 for THF and 1.8 for DEE to reproduce
were averaged over a 4-m depth range to calculate the release history
the peak of 14DNE caused by the large disposal in 1978 (Fig. 5(b)) under
using the measured concentrations at 66 points. To examine whether the
the following settings: from 1.1 to 1.4 for THF and from 1.3 to 1.8 for
averaging affected the calculation accuracy, the release history of
DEE (Table 3). This comparison supports the suitability of the parameter
14DNE was calculated using the original data with the nonnegativity
set for DEE but indicates that the retardation factor of THF was slightly
constraint. The value of θ in the cubic covariance was 1.2 × 10− 3 (mg2/
overestimated. This may reflect insufficient measurement locations,
L2/day3). Two release histories were estimated using the original and
which might have decreased the accuracy of the inversion result, or co-
averaged data (Fig. 8). While the release history based on the original
metabolic biodegradation of 14DNE with THF (Kohlweyer et al., 2000;
data had a slightly larger peak and more uncertainty than the one based
Vainberg et al., 2006; Sun et al., 2011), which may have affected the
on the averaged data, the distributions of the release histories were
shapes and concentrations of the 14DNE and THF plumes.
similar. The residual variance of the result was 1.55 (μg/L)2, which is
similar to that using averaged data (1.07 (μg/L)2). Consequently, this
4.3. Contaminant plume distributions of the three contaminants confirmed that the averaging had little effect on the release history and
plume distribution estimations.
The plume distributions of the three contaminants were calculated
(Fig. 7(a)) from the release histories estimated using the non-negativity
5.2. Effect of the large measurement error
constraint. The accuracy of the estimations was checked using scatter­
plots of the absolute error between the calculated and measured con­
The effect of the measurement error caused by the poor purge effi­
centrations (Fig. 7(b)), the mean absolute error (MAE), and the linear
ciency was examined using a σR value of 3.0 (μg/L) for 14DNE. The value
correlation coefficient (r) values of these concentrations (Table 4). With
of θ in the cubic covariance was 4.9 × 10− 2 (mg2/L2/day3). The esti­
the non-negativity constraint, the MAEs of all the contaminants
mated release history with a large measurement error showed a wider
decreased considerably, with decreases of 46%, 67%, and 59% for
peak concentration than the release history with a small measurement
14DNE, THF, and DEE, respectively, and the r values increased to be­
error (Fig. 9(a)), but the plume distributions were similar regardless of
tween 0.60 and 0.77. These results indicate a high reproduction accu­
the measurement error (Fig. 9(b)), in which MAE was 2.0 × 10− 2 (μg/L)
racy and demonstrate the effectiveness of the non-negativity constraint.
and r was 0.76. The maximum uncertainty of the estimated plume dis­
In addition, the low uncertainties of these distributions can be confirmed
tribution with the large measurement error was up to 1.9 times greater
from the 95% confidence intervals, which were one or more orders
than that with the small measurement error. However, the high uncer­
lower than the best estimates common to the three contaminants (Fig. 7
tainty area (the middle of the study area) was the same regardless of the
(a)). The uncertainties for the three contaminants were relatively high
measurement error, and corresponding to a large release peak and a lack
around the middle of the study area, probably because of the relatively
of measurement points.
high concentrations in that area associated with the large peak after
1978 and the sparse density of the measurement points. The uncertainty
distribution can be used to determine the number and location of 5.3. Verification of the release histories and plume distributions
additional measurements.
The disagreement in the measurements was probably caused by The release histories and plume distributions of the three contami­
using the analytical solution for contaminant transport. The absolute nants are examined in detail below. A vertical cross-section at y = 0 m in
errors between the calculated and measured concentrations were large the E–W direction (along the x-axis) was drawn for the three

Fig. 6. Comparison of the release histories of (a) THF and (b) DEE recovered using geostatistical inversion with a non-negativity constraint. The thick and dashed
lines denote the best estimate (BE) and the 95% confidential interval (95% CI), respectively.

8
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Fig. 7. (a) Perspective views of the plume distributions of the three contaminants (the top row shows the best estimates, the bottom row shows the 95% confidential
intervals). The top surface (z = 24 m) corresponds to the bottom of the aquifer. (b) Scatterplots showing the relationship between the absolute error and the measured
concentrations of the three contaminants using the non-negativity constraint (red circle) and without any constraint (black circle) with each mean absolute error
(dashed line). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

contaminants and overlain with the original (not averaged) measured


Table 4
concentrations (Fig. 10). The calculated plume distributions were
Comparison of the mean absolute error (MAE) and the linear correlation coef­
generally consistent with the measured concentrations. Any disagree­
ficient (r) of the relationships between the calculated and measured concen­
trations with and without the non-negativity constraint. ments between the calculated and measured concentrations may have
been caused by heterogeneity of the aquifer permeability. The geolog­
Contaminant 14DNE THF DEE
ical columnar section had more geological units and suggested inter­
2 1 1
No constraint MAE (mg/L) 3.4 × 10− 2.7 × 10− 4.0 × 10− stratified low permeability zones in the aquifer (Unit C; Jackson et al.,
r 0.60 0.39 0.56
2 2 1 1985). Because of this low permeability, the measured concentrations
Nonnegativity MAE (mg/L) 1.8 × 10− 9.0 × 10− 1.6 × 10−
r 0.77 0.60 0.71 were remarkably low in the area (e.g., low THF and DEE zones extended
through the aquifer at an elevation of around 85 m). In addition, the
groundwater flow direction was presumed to be heterogenous,
depending on the geological structure. However, a simple transport
model with a one-direction flow was adopted in this study to compare
our results with those from the studies of Woodbury et al. (1998) and
Michalak and Kitanidis (2002). The Jacobian matrix in Eqs. (1) or (12)
expresses the sensitivity of the concentration at each point and each
time. This can be calculated numerically by running a flow and transport
model with a release of a unit concentration pulse. Therefore, accurate
three-dimensional groundwater flow models that account for detailed
geological structures are needed to improve the estimation accuracy
instead of the analytical solution.
Differences may also occur by simplifying the contaminant transport
from the source (SWC) to the aquifer (Unit C). Although the SWC was at
a depth of about 10 m, we assumed that the source was uniformly
distributed at the edge of the aquifer in the analytical solution. This
Fig. 8. Comparison of the release history of 14DNE with the non-negativity assumption originated from the fact that the aquitard (Unit D) under the
constraint using all the original data (solid lines) and the data averaged over SWC was locally discontinuous because of heterogenous lithofacies or
each 4-m depth range (translucent lines). The thick and dashed lines denote the excavation for waste disposal (Jackson et al., 1985). This implies that
best estimate (BE) and the 95% confidential interval (95% CI), respectively. most contaminants may have moved rapidly through the discontinuities
in the aquitard, thereby causing the peak after the large disposal in

9
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Fig. 9. Estimated (a) release history and (b) plume distribution of 14DNE using a large measurement error setting. The thick and dashed lines denote the best
estimate (BE) and the 95% confidential interval (95% CI), respectively.

Fig. 10. E–W cross-sections showing the calculated concentrations of the three organic compounds along the x-axis at y = 0 m (Fig. 2(a)). Five hydrogeologic units
(A-E), measured concentrations at sampling locations along the observation wells (colored circles), a heterogeneous silt distribution in the confined aquifer (Unit C),
and the assumed groundwater flow direction in the confined and unconfined aquifer are overlapped.

1978. Meanwhile, the estimated release histories of THF and DEE After the large disposal in May 1978, rainy seasons with rainfall
showed second peaks in 1981 (Fig. 6) because of the high concentrations amounting to approximately 100 mm/month were recorded almost
measured in the upper part of the aquifer at around E.L. = 100 m near every year. Our steady flow assumption may be valid for long-term
the SWC. Although other large spills may have occurred in 1980, the contaminant transport. Therefore, the effects of rainfall infiltration
estimated second peak may be considered a result of the slow transport and seasonal changes in groundwater flow fields on the release condi­
of contaminants through the aquitard. Assuming that the hydraulic tion settings (amount and flow direction) of contaminants from a source
gradient (ih) was 0.002, the hydraulic conductivity of the aquitard (KD) needs to be confirmed.
was 7.9 × 10− 6 (m/s) and the porosity (ρD) was 0.4, following Jackson
et al. (1985), and the thickness of the aquitard (h) was 1.5 m, the 6. Conclusions
transport time calculated as h • Rf/(ih • KD/ρD), in the aquitard was
estimated as 769 days for THF and 791 days for DEE. The transport time The aim of this study was to demonstrate the applicability and
in the aquitard was approximately consistent with the time lags between effectiveness of a geostatistical approach that incorporated groundwater
the first and second peaks of the estimated release history, which was flow modeling to estimate the contaminant plume when the release
880 days for THF and 890 days for DEE. To improve the accuracy of the history is unknown. The approach was tested on groundwater contam­
estimations, the contaminant release from the SWC needs to be modeled. ination in the Gloucester landfill. The approach was effective, as the
Differences may also occur because of seasonal change in rainfall. accuracy of the estimation was improved by applying a non-negativity

10
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

constraint using Gibbs sampling. With the geostatistical approach and a Boente, C., Matanzas, N., Garcia-Gonzalez, N., Rodriguez-Valdes, E., Gallego, J.R., 2017.
Trace elements of concern affecting urban agriculture in industrialized areas: a
non-negativity constraint, the three-dimensional plume distributions
multivariate approach. Chemosphere 183, 546–556.
were produced with adequate accuracy for the three organic solvents Box, G.E.P., Cox, D.R., 1964. An analysis of transformations. J. R. Stat. Soc. Se. B 26,
(14DNE, THF, and DEE) with linear correlation coefficients between the 353–360.
estimation and measurement of 0.60 to 0.77 for the concentration data Brooke, I., Cocker, J., Delic, J.I., Payne, M., Jones, K., Gregg, N.C., Dyne, D., 1998.
Dermal uptake of solvents from the vapour phase: an experimental study in humans.
from multiple depths. In addition, the reconstructed release peak for Ann. Occup. Hyg. 42, 531–540.
14DNE that corresponded to the large disposal in 1978 was more correct Butera, I., Tanda, M.G., 2003. A geostatistical approach to recover the release history of
than for the preceding studies, without any prior information of the groundwater pollutants. Water Resour. Res. 39 (12), 1372.
Chen, Z., Xu, T., Gomez-Hernandez, J.J., Zanini, A., 2021. Contaminant spill in a sandbox
release history. The reconstructed release histories for THF and DEE also with non-Gaussian conductivities: simultaneous identification by the restart normal-
indicated the same peaks with retardation factors that were generally score ensemble Kalman filter. Math. Geol. 53, 1587–1615.
consistent with the estimation using fraction of organic carbon. This Cherry, J.A., Gillham, R.W., Anderson, E.G., Johnson, P.E., 1983. Migration of
contaminants in groundwater at a landfill: a case study. 2. Groundwater monitoring
study may be the first application to estimate three-dimensional plume devices. J. Hydrol. 63, 31–49.
distributions and the associated uncertainties when the release history is Colombo, L., Alberti, L., Mazzon, P., Antelmi, M., 2020. Null-space Monte Carlo particle
unknown, targeting multiple contaminants. The proposed method can backtracking to identify groundwater tetrachloroethylene sources. Front. Environ.
Sci. 8, 142.
be applied widely to contaminated groundwater fields. Dalla Libera, N., Fabbri, P., Mason, L., Piccinini, L., Pola, M., 2017. Geostatistics as a tool
To increase the accuracy of the estimation, the modeling of the to improve the natural background level definition: an application in groundwater.
groundwater flow and the hydrogeologic structure should be improved. Sci. Total Environ. 598, 330–340.
Domenico, P.A., Schwartz, F.W., 1990. Physical and Chemical Hydrogeology, 1st ed.
The accuracy of the predictions of future plume transport could be
Wiley, New York.
improved with precise estimation of the hydrogeological structure, Draper, W.M., Dhoot, J.S., Remoy, J.W., Perera, S.K., 2000. Trace-level determination of
which would be useful when planning for remediation. The hydro­ 1,4-dioxane in water by isotopic dilution GC and GC-MS. Analyst 125, 1403–1408.
geological structure could be estimated from transient measurements of Dubrule, O., 1984. Comparing splines and kriging. Comput. Geosci. 10 (2–3), 327–338.
Franz Environmental Inc, 2007. Risk management plan: former Gloucester landfill. In:
groundwater concentrations. For example, the hydraulic parameters Final Report presented to Transport Canada.
used in the analytical solution could be calibrated by sensitivity analysis Gailey, R.M., Crowe, A.S., Gorelick, S.M., 1991. Coupled process parameter estimation
with an objective function, such as the reproducibility of measurement. and prediction uncertainty using hydraulic head and concentration data. Adv. Water
Resour. 14 (5), 301–314.
The Ensemble Kalman filter is one of the potential methods, which can Gomez-Hernandez, J.J., Xu, T., 2022. Contaminant source identification in aquifers: a
be also applied to a joint estimation of contaminant source and hydraulic critical view. Math. Geosci. 54 (2), 437–458.
conductivity (e.g., Chen et al., 2021). Geostatistic-based hydraulic to­ Gyzl, G., Zanini, A., Fraczek, R., Kura, K., 2014. Contaminant source and release history
identification in groundwater: a multi-step approach. J. Contam. Hydrol. 157,
mography using well hydraulic test data could be used to help account 59–72.
for heterogeneity of hydraulic conductivity from limited data (e.g., Lee Hansch, C., Loe, A., Hoekman, D., 1995. Exploring QSAR: Hydrophobic, Electronic, and
and Kitanidis, 2014). The estimations in this study could also be Steric Constants. American Cehmical Society, Washington, DC.
ICSC Database, 2017. International Chemical Safety Cards. International Labour
improved by considering spatial correlation of measurement data Organization, Geneva. https://www.ilo.org/safework/info/publications/W
(Shlomi and Michalak, 2007), in addition to the temporal correlation of CMS_113134/lang–en/index.htm (accessed 24 March 2022).
release history. However, the steady-flow assumptions may not be valid Iskandar, I., Koike, K., Sendjaja, P., 2012. Identifying groundwater arsenic contamination
mechanisms in relation to arsenic concentrations in water and host rocks. Environ.
if the transient change of groundwater flow increases (e.g., by pumping
Earth Sci. 65 (7), 2015–2026.
wells). In this case, our method should be repeated with temporal Issacson, C., Mohr, T.K.G., Field, J.A., 2006. Quantitative determination of 1,4-dioxane
measurement data to estimate the plume distributions. and tetrahydrofuran in groundwater by solid phase extraction GC/MS/MS. Environ.
Sci. Technol. 40, 7305–7311.
Jackson, R.E., Patterson, R.J., Graham, Bahr, J., Belanger, D., Lockwood, J., Priddle, M.,
CRediT authorship contribution statement 1985. Contaminant Hydrogeology of Toxic Organic Chemicals at a Disposal Site,
Gloucester, Ontario: 1. Chemical Concepts and Site Assessment. National hydrology
research institute. NHRI Paper No. 23, IWD Scientific Series, No. 141, 114p.
Shizuka Takai: Conceptualization, Data curation, Formal analysis,
Jackson, R.E., Crowe, A.S., Lesage, S., Priddle, M.W., 1989. Aquifer Contamination and
Investigation, Methodology, Validation, Visualization, Writing – orig­ Restoration at the Gloucester Landfill, Ontario, Canada, IAHS Publication, 185.
inal draft. Taro Shimada: Funding acquisition, Supervision, Writing – International Association of Hydrological Sciences, Walingford, Oxfordshire, U.K.,
pp. 181–188
review & editing. Seiji Takeda: Funding acquisition, Supervision,
Johnson, N., Kotz, S., Balakrishnan, N., 1994. Continuous Univariate Distributions, 2nd
Writing – review & editing. Katsuaki Koike: Supervision, Methodology, ed. Wiley, New York.
Validation, Writing – review & editing. Juang, K.-W., Liao, W.-J., Liu, T.-L., Tsui, L., Lee, D.-Y., 2008. Additional sampling based
on regulation threshold and kriging variance to reduce the probability of false
delineation in a contaminated site. Sci. Total Environ. 389 (1), 20–28.
Declaration of Competing Interest Kitanidis, P.K., 1995. Quasi-linear geostatistical theory for inversing. Water Resour. Res.
31, 2411–2419.
Kitanidis, P.K., Lee, J., 2014. Principal component geostatistical approach for large-
The authors declare that they have no known competing financial dimensional inverse problems. Water Resour. Res. 50, 5428–5443.
interests or personal relationships that could have appeared to influence Kitanidis, P.K., Vomvoris, E.G., 1983. A geostatistical approach to the inverse problem in
groundwater modeling (steady-state) and one-dimensional simulations. Water
the work reported in this paper. Resour. Res. 19, 677–690.
Kohlweyer, U., Thiemer, B., Schrader, T., Andreesen, J.R., 2000. Tetrahydrofuran
Data availability degradation by a newly isolated culture of Pseudonocardia sp. strain K1. FEMS
Microbiol. Lett. 186, 2, 301–306.
Lee, J., Kitanidis, P.K., 2014. Large-scale hydraulic tomography and joint inversion of
The data that has been used is confidential. head and tracer data using the principal component geostatistical approach (PCGA).
Water Resour. Res. 50, 5410–5427.
Lessage, S., Jackson, R.E., Priddle, M.W., Rlemann, P.G., 1990. Occurrence and fate of
Acknowledgments organic solvent residues in anoxic groundwater at the Gloucester landfill, Canada.
Environ. Sci. Technol. 24 (4), 559–566.
This study was funded by the Secretariat of Nuclear Regulation Au­ Liang, C.P., Chen, J.S., Chien, Y.C., Chen, C.F., 2018. Spatial analysis of the risk to human
health from exposure to arsenic contaminated groundwater: a kriging approach. Sci.
thority, Nuclear Regulation Authority, Japan.
Total Environ. 627, 1048–1057.
Loconto, P.R., 2001. Trace Environmental Quantitative Analysis, Principles: Techniques
References and Applications, 1st ed. CRC Press, Boca Raton, FL.
MacFarlane, D.S., Cherry, J.A., Gillham, R.W., Sudicky, E.A., 1983. Migration of
contaminants in groundwater at a landfill: A case study: 1. Groundwater flow and
Antunes, I.M.H.R., Albuquerque, M.T.D., 2013. Using indicator kriging for the evaluation
plume delineation. J. Hydrol. 63 (1–2), 1–29.
of arsenic potential contamination in an abandoned mining area (Portugal). Sci.
Total Environ. 442, 545–552.

11
S. Takai et al. Journal of Contaminant Hydrology 251 (2022) 104097

Masoudi, P., Desnoyers, Y., Grey, M., 2021. Spatio-temporal optimization of Patterson, R.J., Jackson, R.E., Graham, B.W., Chaput, D., Priddle, M., 1985. Retardation
groundwater monitoring network at Pickering Nuclear Generating Station. In: of toxic chemicals in a contaminated outwash aquifer. Wat. Sci. Tech. 17, 57–69.
Geostats 2021, July 12–16, 2021. Toronto, Canada. Saby, N., Arrouays, D., Boulonne, L., Jolivet, C., Pochot, A., 2006. Geostatistical
Michalak, A.M., 2002. Environmental contamination with multiple potential sources and assessment of Pb in soil around Paris, France. Sci. Total Environ. 367 (1), 212–221.
the common law- current approaches and emerging opportunities. Fordham Environ. Schwarzenbach, R.P., Westall, J., 1981. Transport of nonpolar organic compounds from
Law J. 14 (1), 147–206. surface water to groundwater, laboratory studies. Environ. Sci. Technol. 15,
Michalak, A.M., 2008. A Gibbs sampler for inequality-constrained geostatistical 1300–1367.
interpolation and inverse modeling. Water Resour. Res. 44, 1–14. Shlomi, S., Michalak, A.M., 2007. A geostatistical framework for incorporating transport
Michalak, A.M., Kitanidis, P.K., 2002. Application of Bayesian inference methods to information in estimating the distribution of a groundwater contaminant plume.
inverse modeling for contaminant source identification at Gloucester Landfill, Water Resour. Res. 43, 1–12.
Canada. In: Hassanizadeh, S.M., et al. (Eds.), Computational Methods in Water Skaggs, T.H., Kabala, Z.J., 1994. Recovering the release history of a groundwater
Resources XIV, 2. Elsevier, Amsterdam, Netherlands, pp. 1259–1266. contaminant. Water Resour. Res. 30 (1), 71–79.
Michalak, A.M., Kitanidis, P.K., 2003. A method for enforcing parameter nonnegativity Snodgrass, M.F., Kitanidis, P.K., 1997. A geostatistical approach to contaminant source
in Bayesian inverse problems with an application to contaminant source identification. Water Resour. Res. 33 (4), 537–546.
identification. Water Resour. Res. 39 (2), 1033. Sun, B., Ko, K., Ramsay, J.A., 2011. Biodegradation of 1,4-dioxane by a Flavobacterium.
Michalak, A.M., Kitanidis, P.K., 2004a. Application of geostatistical inverse modeling to Biodegradation 22 (3), 651–659.
contaminant source identification at dover AFB, Delaware. J. Hydraul. Res. 42, 9–18. U.S. EPA, 2000. Multi-Agency Radiation Survey and Site Investigation Manual
Michalak, A.M., Kitanidis, P.K., 2004b. A method for the interpolation of nonnegative (MARSSIM), Revision 1, EPA-402-R-97-016, August.
functions with an application to contaminant load estimation. Stoch. Env. Res. 18, Vainberg, S., McClay, K., Masuda, H., Root, D., Condee, C., Zylstra, G.J., Steffan, R.J.,
1–16. 2006. Biodegradation of ether pollutants by Pseudonocardia sp. strain ENV478.
Michalak, A.M., Kitanidis, P.K., 2005. A method for the interpolation of nonnegative Appl. Environ. Microbiol. 72 (8), 5218–5224.
functions with an application to contaminant load estimation. Stoch. Env. Res. Risk Wishart, D.S., Tzur, D., Knox, C., et al., 2007. HMDB: the human metabolome database.
A. 19 (1), 8–23. Nucleic Acids Res. 35 (Database issue), D521–D526, 2007 Jan.
Miller, S.M., Michalak, A.M., Levi, P.J., 2014. Atomospheric inverse modeling with Woodbury, A., Sudicky, E., Ulrych, T.J., Ludwig, R., 1998. Three-dimensional plume
known physical bounds: an example from trace gas emissions. Geosci. Model Dev. 7, source reconstruction using minimum relative entropy inversion. J. Contam. Hydrol.
303–315. 32, 131.
Mohr, T.K.G., 2001. Solvent Stabilizers, Santa Clara Valley Water District; San Jose, CA. Yalkowsky, S.H., He, Yan, Jain, P., 2010. Handbook of Aqueous Solubility Data, 2nd ed.
OECD, 2014. Nuclear Site Remediation and Restoration during Decommissioning of CRC Press, Boca Raton, FL.
Nuclear Installations: A Report by the NEA Co-Operative Programme on
Decommissioning. Radioactive Waste Management. NEA No. 7192.

12

You might also like