Environmental Modelling & Software 25 (2010) 1769e1780

Contents lists available at ScienceDirect

Environmental Modelling & Software

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

Flow guided interpolation e A GIS-based method to represent contaminant

concentration distributions in groundwater
M. Morio a, *, M. Finkel a, E. Martac b, c
a
Universität Tübingen, Center for Applied Geoscience, Sigwartstr. 10, D-72076 Tübingen, Germany
b
tgageTübinger Gesellschaft für Angewandte Geowissenschaften e.V., Sigwartstr. 10, D-72076 Tübingen, Germany
c
geo-log GmbH, Bahnhofstraße 9, D-72147 Nehren, Germany

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

Article history: This paper introduces a new interpolation method to estimate the spatial distribution of contaminant
Received 25 November 2009 concentrations in groundwater. The method is intended to identify areas of risks in early investigation
Received in revised form stages when groundwater sampling data is typically scarce and available interpolation methods fail to
11 May 2010
provide reasonable results. As a consequence, the method does not only incorporate available sampling
Accepted 24 May 2010
Available online 7 July 2010
data, but also makes use of information about the groundwater flow field, in order to “guide” the
interpolation with e.g. ordinary kriging or inverse distance method. The guidance includes the
augmentation of available data by auxiliary point data and the segmentation of the estimated plume area
Keywords:
Interpolation
into a series of sectors. The method is evaluated for several settings and different sampling data sets.
Groundwater contamination Each data set reflects a specific level of field investigations at the model site, an abandoned military base
Contaminant transport in Potsdam near Berlin, Germany. The results reveal that flow guidance improves the representation of
Chlorinated hydrocarbons contaminant distribution for all cases examined in this study compared to “unguided” interpolation.
TCE These findings are underpinned by the results of the method’s application to real sampling data. The
GIS method especially shows its strength when data of only a few sampling points are available.
Particle tracking Ó 2010 Elsevier Ltd. All rights reserved.
Ordinary kriging
Numerical groundwater flow modeling

1. Introduction methods other than the estimation of the concentration distribu-

tion (e.g. Boufassa and Armstrong, 1989; MacKay, 1990; Kerry and
Different approaches exist to estimate the spatial distribution of Oliver, 2007).
contaminants dissolved in groundwater. Proposed approaches Data assimilation techniques such as Particle Filter and
range from experience-based practitioners’ methods, such as best Ensemble Kalman Filtering have gained considerable interest in the
guess delineation of source zones and plumes based on small last decade for utilization of available measurement data to update
amounts of sample data and experiences from similar sites and mathematical model predictions of groundwater flow and plume
cases, to advanced methods involving geostatistics (e.g. Isaaks and propagation (e.g. Evensen, 1994; Eigbe et al., 1998; Porter et al.,
Srivastasa, 1989; Deutsch and Journel, 1997; Michalak and 2000; Chang and Jin, 2005; Chang and Latif, 2009). Filtering
Kitanidis, 2004a, 2005), modeling (e.g. Shlomi and Michalak, methods seem to be best suited for transient problems of
2007; Prommer et al., 2002; D’Affonseca et al., 2008) and data groundwater system state estimation when time series of
assimilation techniques e.g. using Bayesian filtering techniques (e.g. measurements are to be repeatedly (i.e. sequentially) assimilated
Kalman, 1960; Chen, 2003). The appropriateness of individual into mathematical models (e.g. Liu et al., 2008; Huang et al., 2009;
methods is mostly dependent on the type, amount and quality of Chang and Latif, 2010).
available data as well as on the particular objectives. Plume delin- Geostatistical interpolation methods have been widely applied
eation focusing on detecting the plume’s extent (e.g. Meyer et al., in the past decades. Mehrjardi et al. (2008) proposed the use of
1994; Storck et al., 1997; McGrath and Pinder, 2003) may call for ordinary kriging and cokriging for interpolation of contaminants in
groundwater, while Reed et al. (2004) evaluated different inter-
polation methods to estimate the distribution of perchloroethylene
* Corresponding author.
(PCE) in three heterogeneous test cases of different size and
E-mail addresses: max.morio@uni-tuebingen.de (M. Morio), michael.finkel@ complexity for a groundwater plume with differing amounts of
uni-tuebingen.de (M. Finkel), e.martac@geo-log.com (E. Martac). non-gridded sampling data. They recommend quantile kriging and

1364-8152/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.envsoft.2010.05.018
1770 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780

multigaussian kriging to be most robust and least biased compared information basis that is typically available in early stages of revi-
to ordinary kriging, intrinsic kriging and inverse distance weighted talization projects at contaminated sites. The method proposes to
methods. Journel and Rossi (1989) showed that universal kriging add extra sampling points in a standardized way through coupling
(also called “kriging with trend”) yields similar results to ordinary of flow data and existing samples plus information about possible
kriging on data sets with trend when the trend component is source zone extents to enable the application of kriging methods.
unknown and kriging is conducted in local neighborhoods for non- The purpose of the proposed method is to bridge the gap between
stationary data sets. elaborate, data intensive approaches and subjective and often non-
Cooper and Istok (1988) discusses the requisite of data prepa- reproducible methods, which are frequently applied in practice if
ration and analysis, including additivity, stationarity and amount of available data is scarce. The FGI method builds upon a groundwater
samples for the estimation using geostatistical interpolation flow model, assuming that basic information about the ground-
methods. Sufficient data is required, e.g. to make use of an empirical water flow regime can be made available at relatively low cost. This
semivariogram (e.g. Deutsch and Journel, 1997; Fuest et al., 1998; model is not supposed to provide a highly sophisticated repre-
Kitanidis and Shen, 1996; Reed et al., 2004; Kistemann et al., sentation of the flow situation but is supposed to show the major
2008). A practical rule for the minimum amount of samples is i.e. characteristic features. Guidance by groundwater flow is
given by Journel and Huijbregts (1978): intended to improve the interpolation especially if contaminant
concentration data is scarce. The idea is to incorporate upgradient
information in a sequential downgradient-moving interpolation
NðhÞ > 30  50 with : jhj < ðL=2Þ; (1)
procedure. The relevance of the quantity of available data is
where jhj denotes the magnitude of separation vector h for N addressed through analyzing the FGI method’s performance for
sample pairs and L stands for the longest dimension of the different knowledge states, i.e. different sample amounts.
contaminant plume in the direction of h. For a complete list of The remainder of this paper is organized as follows: the FGI
symbols used in this article, please refer to Table A.5 method is described in section 2 by use of pseudo-code algorithms
Interpolation may utilize concentration measurements either in to explain the sequential procedure execution; section 3 gives
terms of point observation data from distributed monitoring a description of the model site in which the FGI method was
networks (e.g. Sudicky et al., 1983; Warrick et al., 1998) or in terms applied; the evaluation of the FGI method for different parameter
of data from so-called control planes or monitoring fences (e.g. settings is discussed in section 4; results of the case study are
Schwarz et al., 1998; King and Barker, 1999; Bockelmann et al., presented in section 5; conclusions are made in section 6.
2001; Basu et al., 2006; Kübert and Finkel, 2006; Bayer-Raich
et al., 2009). 2. Methodology
Several suggestions were made to improve interpolation by
The FGI method is implemented into the geographic information system ArcGIS
including additional information in the estimation process, e.g.
(ArcMap, Version 9.1 or higher, (ESRI, 1992e2005) with VBA and ArcObjects (Razavi,
hydraulic gradient or head data from sampling campaigns (e.g. 2004; Burke and Arana, 2003). The method is comprised of a sequence of procedures
Burger and Schafmeister, 2000; Shlomi and Michalak, 2007). with major portions that are controlled by VBA modules.
Other approaches are based on coupling numerical transport In short, the following steps are performed: (1) delineation of the known or
models to interpolation methods (Michalak and Kitanidis, 2004a, expected source zone of contamination using particles, which are equidistantly
distributed along the source zone edges; (2) delineation of the plume fringe by
2004b). Neupauer and Wilson (2004) used a probabilistic tracing these particles advectively downgradient; (3) segregation of the plume into
numerical flow and transport model to relate concentration several sectors according to the plume’s tortuosity and curvature; (4) employment
measurements to possible upgradient source locations. Rautman of sector-wise flow guided interpolation, utilizing sampling data in the sector, as
and Istok (1996) and Istok and Rautman (1996) proposed well as auxiliary sampling point data along the plume fringe and along the boundary
to the previously processed sector; (5) merging of the results of all sectors into
stochastic geostatistical modeling of contaminant plumes as an
a single grid by mosaicking.
approach to derive probabilities of having a contamination at A series of three pseudo-code algorithms describe the process of (i) plume
a certain point with respect to a specific concentration threshold fringe delineation based on the previously defined source zone (Algorithm 1, see
and probability cutoff. also Fig. 1a), of (ii) segmentation of the derived plume shape dependent on
A common problem in practice is that available contaminant groundwater flow direction (Algorithm 2, see also Fig. 1b) and of (iii) sector-wise
interpolation to estimate the spatial concentration distribution within the plume
concentration measurement data is not sufficient to make sensible (Algorithm 3, see also Fig. 1ceg). The procedures are described in more detail
use of the abovementioned geostatistical interpolation methods. below. Please note that multiple source zones and corresponding plumes can be
Numerical transport modeling (e.g. Anderson and Cherry, 1979; Chu considered. For the sake of clarity we limit the description to one source zone and
et al., 1987), which further requires comprehensive information one plume. The delineation of the plume fringe requires polygon data of the source
zone, a numerical groundwater flow model, and a particle tracking module. The
about aquifer and transport properties, is also inappropriate for this
flow model is required to calculate the groundwater flow field, which in turn is
reason e.g. (Batu, 2006). Limited availability of contaminant data is required to calculate pathlines of groundwater flow by particle tracking. We used
characteristic of early project stages in tiered decision-making MODFLOW 96 (Harbough and McDonald, 1996) and MODPATH 3.0 (Pollock, 1994),
procedures, when information on subsurface contamination only respectively. Note that other models, e.g. a version of ESRI’s Groundwater Modeling
stems from historical data and some initial site investigation. This is application (ESRI, 2009), could be used as well after some adaption work (see also
the concluding discussion of this section further below). The identification of the
particularly true, e.g. when extent and complexity of a site require an source zone is based on desk work examining information on former use of the
early identification and prioritization of focal areas and origins of site, possible locations of contaminant spills, the geological and hydrological
risks in order to drive further decisions on detailed investigation situation, and subsurface sampling information. Source zone polygons are then
programmes and remediation measures (compare Triad approach, created manually as polygon shape files in GIS, based on the assembly of given
data. As shown in Algorithm 1, the source zone polygons are converted into a set of
e.g. in Crumbling et al., 2001; Mack et al., 2004; Critto et al., 2007;
equidistantly distributed points, which are then used as water particle starting
O’Reilly and Brink, 2006). Thus, especially for early site investiga- locations for particle tracking with MODPATH. The pathline data returned by
tion stages, there is a need for enhancement and appropriate pro- MODPATH is automatically converted to a polyline shape file consisting of polylines
cessing of sparse amounts of available data in order to produce the for each particle starting point and a corresponding point shape file bearing
premise for a reasonable application of interpolation methods such information on the pathline time steps. The pair of outermost polylines in the
shape file represents the plume fringe and is converted to a polygon shape file. The
as kriging.
resulting polygon has to be cut in order to account for the expected plume age.
In this paper we present a flow guided interpolation (FGI) Using travel time information from the point shape file, a polygon is created that
method that has been specifically adapted to the type, scale, and represents the plume extent corresponding to the given time after spill.
 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780 1771

Fig. 1. Flow guided interpolation: (a) potential source zone with outermost groundwater flow pathlines, (b) segmentation according to curvature of plume’s center pathline (here: 6
sectors), (c) interpolation within sector downgradient of the source zone, (deg) consecutive sector-wise interpolation in downgradient direction. Blue colored circles denote
groundwater wells, orange circles stand for direct push samples.

Algorithm 1. FGI-work flow: plume delineation invoking an ArcGIS-MODPATH These steps are processed mostly automatically by the VBA code. The center pathline
interface is provided as a result of Algorithm 1 in terms of a polyline and its corresponding
Require: Source zone polygon data {GIS (multiple sources at a site possible)} point shape file. Sseg serves as a criterion for the maximum deflection of the center
Require: Numerical flow model {here: MODFLOW 96} pathline that causes the creation of a new sector. Sector boundary points are
Require: Advective transport i.e. particle tracking module {here: MODPATH 3.0} determined along the center pathline. The deflection of a center pathline point i is
Read model extent and discretization {ArcGIS-MODPATH interface, works auto- quantified as normal distance lcrit, i of this point to an ’elongation line’ representing
matically, if not noted otherwise.} the center pathline’s initial direction (cf. Fig. 2b). If lcrit, i > Sseg, a new boundary point
Convert source polygon to points (equidistantly distributed, here: point distance will be created midway between point i and the initial point from where the
Dl ¼ 10 m) elongation line starts (cf. purple arrow in Fig. 2b). This process of segmentation is
Create MODPATH particle starting locations repeated in downgradient direction to determine all sector boundaries until the tip
Run MODPATH, output: pathline file of the plume is reached. Lines perpendicularly oriented to the center pathline are
Import MODPATH pathlines in GIS, Convert to: created on the newly derived boundary points. These lines are then used to clip the
B ESRI raster grid representing cells passed by particles plume polygon into distinct sectors. The number of sectors resulting from plume
B Polyline shape file {pathlines} segmentation increases as the Sseg value is lowered and as the plume shape becomes
B Point shape file with attribute: cumulative travel times more winding (cf. Fig. 2b). The sector polygons are then converted into points which
for all source zone(s) do are subsequently used to create auxiliary sampling points (ASPs) at the upgradient
Select outermost pathlines and Convert {Manually}: sector boundary, and at the lateral boundaries along the plume fringe. The plume
segmentation provides additional geometrical variables, which serve as input for the
B Polyline to polygon shape file representing plume
subsequent interpolation procedure, namely the total plume length Ltot, and for each
B Polyline to point shape file with points on model cell edges
sector s the mean width Ws, and anisotropy angle as.
Select center pathline’s polyline and Convert {Manually}:
B Polyline to point shape file with points on model cell edges Algorithm 2. FGI-work flow: plume segmentation based on flow direction
With point shape files: Select points that equals spill age {Attribute: cumu- Require: Centerline point and polyline shape file {output from algorithm 1 with n
lative travel times} no. of points on center pathline}
Extract selected points and convert to polyline Read Gauss Krüger coordinates of vertex points
Split polygon by polyline {Result: Polygon with plume extent} Calculate descriptive variables to describe center pathline: lengths of line
end for elements li and angles bi of li, the directions of lines between adjacent nodes
ðxi ; yi Þðxiþ1 ; yiþ1 Þ against horizontal
Please note that this purely advective approach may underestimate the real s ¼ 1 {sector number}
plume extent. However, assuming that macroscale heterogeneity is sufficiently for i ¼ 1/n do
described in the flow model used, the neglected effect of hydrodynamic dispersion is nstep ¼ 0 {additional increment variable}
believed to be minor. while lcrit;iþnstep < Sseg ^i þ nstep < n do
The segmentation of the plume into distinct sectors is based solely on the lcrit;iþnstep ¼ jðyiþ1  yi Þ$ðxi  xiþnstep Þ  ðxiþ1  xi Þ$ðyi  yiþnstep Þj=li {Calcu
curvature of the plume’s center pathline and one further input parameter, the so- late normal distance of plume center pathline node i þ nstep to elongation
called segmentation criterion, Sseg. The segmentation of the center pathline is line (cf. Fig. 2a)}
illustrated in Fig. 2a and is comprised of diverse steps as described by Algorithm 2. nstep ¼ nstep þ 1
1772 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780

Fig. 2. Illustration of the plume segmentation process dependent on the curvature of the plume’s center pathline.(a) Principles of segmentation process. Righthand-side triangle
shows an enlarged section of the pathline on the left-hand side of the figure. bi denotes the angle between two adjacent pathline elements. (b) Results of plume segmentation for
different values of Sseg (from left to right 50 m, 2.5 m, 0.5 m, and 0.05 m).

end while The sector-wise flow guided interpolation itself is described in Algorithm 3.
i ¼ i þ nstep Sectors are distinguished by index s. The interpolation starts in the sector located
if i < n directly downgradient of the source zone (s ¼ 1). The processing of the first sector
{Determine sector boundary coordinate on the middle between two existing is different compared to the subsequent ones because there is no upgradient
points.} neighboring sector except for the source zone. The source zone boundary is
xbd;s ¼ ðxiþ1Pnstep $0:5þ0:5R þ xiPnstep $0:5þ0:5R Þ=2 populated with ASPs, the location of which coincide with the particle starting
ybd;s ¼ ðyiþ1Pnstep $0:5þ0:5R þ yiPnstep $0:5þ0:5R Þ=2 locations. To each of these ASPs a concentration value needs to be assigned that
end if represents the source zone conditions. In this study we assign a unique value, the
s ¼ s þ 1 {Increment number of sector} mean of available concentration measurements in proximity of the source zone, to
end for all points. If detailed information is available, spatially varying concentration
Calculate azimuth of straigt line defined by adjacent sector boundary coordi- values may be used. Sample values at ASPs along the lateral boundary of the sector
nates ðxbd;s ybd;s ; xbd;sþ1 ; ybd;sþ1 Þ {Yields anisotropy angle as for each sector} (plume fringes) are set to some lower threshold concentration Ctc (e.g. the
Calculate cumulative length Ltot of point to point distances li detection limit of the contaminant of concern). All ASPs and the real samples
Create new boundary points table with xbd, s ybd, s as available in the sector in process are added to the sample list to be used for
Create polyline shape file with lines perpendicular to center pathline through interpolation. If no real sample is available, real samples (but not the ASPs) of the
each of its vertices {Done by VBA script} neighboring sector downgradient are added to the sample list. If the neighboring
Select a subset of lines that intersect sector boundary points on center pathline downgradient sector also does not contain real sample points, the next down-
by new sector boundary points from new boundary points table {Done by SQL gradient sector is considered, and real samples of this sector are added to the
type of GIS query function} sample list of the sector in process. Note that the additivity assumption of sample
Clip plume polygon with subset of lines to create sectors data is fulfilled since effective porosity and thickness of the aquifer are assumed to
Calculate mean Width Ws for each sector. be constant for this study (compare Cooper and Istok, 1988). Note further that
Create equidistantly distributed points along sector boundaries (here: point sample data outside the plume fringe are not considered for interpolation. If there
distance Dl ¼ 10 m). is evidence for a contamination outside the delineated plume fringe belonging to
 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780 1773

Table 1
Available sampling data and sector-specific interpolation parameters in plume sectors for different cases of segmentation and investigation levels.

Sectorsa Sector no. No. of Samplesb Mean sector flow directionc Anisotropy ratio Total plume length Mean sector width in sector s
 
() s () Wells,DP1,DP2/Tree-core as ( , E ¼ 0 , CCW) RAs() Ltot(m) Ws(m)
2 1 4, þ3, þ1/13 41.0 4.4 400 90
2 0, þ0, þ0/9 25.0 3.4 115

1 1, þ1, þ0/4 47.0 4.4 91

2 3, þ2, þ1/9 60.5 4.4 90
5 3 0, þ0, þ0/2 31.8 4.0 400 100
4 0, þ0, þ03 17.8 3.2 125
5 0, þ0, þ0/4 26.5 3.0 135

1 1, þ1, þ0/4 48.6 4 91

2 0, þ0, þ0/0 51.1 5.5 90
3 3, þ2, þ1/9 63.4 4.2 94
8 4 0, þ0, þ0/1 47.2 3.8 105
5 0, þ0, þ0/1 14.2 3.8 400 106
6 0, þ0, þ0/3 17.4 3.1 127
7 0, þ0, þ0/2 22.8 3.2 125
8 0, þ0, þ0/2 27.1 3.2 105
a
This columns indicates the amount of sectors, into which the plume is segmented.
b
This column denotes the amount of samples that are available in the particular sector s after investigation campaign “Wells” and the amount of samples that became
additionally available after campaigns “DP1” and “DP2”. The numbers given for investigation phase “Tree-core” stands for on own and are not to be added to the samples of the
other investigation campaigns.
c
i.e. anisotropy angle.

the given source zone we suggest to consider a revision of the flow model or of the sector are noted in Table 1. Please note that a trend within the sample data, i.e.
source zone extent. a component or tendency of the data to change their values according to their spatial
Ordinary kriging is then conducted with a sector-specific values of anisotropy position, is not considered here. Due to the small amount of sample data available
ratio RAs (derived from the ratio Ltot/Ws) and angle as, which is provided by within the individual sectors, it does not seem practicable to estimate a reliable
Algorithm 2. The respective parameters for the interpolation procedure in each trend component and to model a corresponding variogram. The grid of estimated

Fig. 3. Field site with model boundaries, area of interest and zoomed-in study area. Colored circles illustrate magnitude of sampled TCE concentrations (orange) and semi-
quantitative tree-core sample values (purple).
1774 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780

concentration values has the same resolution as the flow model grid. The grid extent 3. Model site: investigation data and flow model
is clipped to the sector in process and sampled at the downstream boundary points
to generate the ASPs required for interpolation in the next sector. The interpolation
of the 2nd and any further sector follows the same procedure. The grids calculated
The model site is an abandoned military base covering 120 ha
for the single sectors are finally mosaicked, i.e. sector-specific results are merged to on the outskirts of Potsdam near Berlin, Germany. The site was
a single raster grid, to produce the grid of concentration values for the entire plume. turned into a brownfield in the early 1990s. It was used by German
Algorithm 3. FGI-work flow: estimating contaminant distribution and Russian armed forces until 1945 and 1991, respectively. Gas
Require Csource, Ctc {Source and plume fringe concentration}, xbd, s, ybd, s, as, Ws, Ltot, stations and a dry cleaning facility represent the major sources of
plume sector polygons and according boundary points {Result of algorithm 2} subsurface contamination. Since 1992, the site has been revisited
for s ¼ 1/stot do and investigated several times by various groups of consultants
Assign concentration values to auxiliary sampling points on upper sector
and researchers. Four major investigation campaigns have resul-
boundary
if s ¼ 1 then ted in an increasing net of monitoring wells, from a rather scarce
Cub, s ¼ Csource {Use source concentration values in 1st sector} net of observation points in 1996 to a quite dense net after
else a comprehensive direct push campaign and sampling of tree-cores
Cub, s ¼ Clb, se1 {Use results from previous sector}
in 2007.
end if
Extract Cub, s, add to XYZ-file of sector s {Using ArcGIS Toolbox
The presently known groundwater contamination is dominated
Sample_SpatialAnalyst} by volatile chlorinated hydrocarbons (CHC, mainly trichloroethy-
Assign Ctc to lateral plume boundary as ASP in sector s (add to XYZ-file of sector s) lene (TCE) and cis-dichloroethylene, up to more than 170 mg/l).
if Samples in sector s d then Contaminants originate from diverse source zones, as shown in
Select samples within sector (add to XYZ-file of sector s)
Fig. 3. Contaminated groundwater flows from the site towards two
else
Select samples from next sector containing samples from (add to XYZ-file of lakes, nature reserves and other potential receptors (Fig. 3).
sector s) A two-dimensional finite difference flow model was used to
end if simulate groundwater flow in the unconfined aquifer consisting of
CalculateRAs ¼ Ltot/Ws {anisotropy ratio}
fine sand to silty deposits. The model domain is represented by
as anisotropy angle, calculated in algorithm 2
Perform Ordinary Kriging on XYZ-file of sector s using RAs and as
a cell-grid with an extent of 2500  4000 m using a discretization of
end for square cells of 10  10 m. A fixed head boundary condition is defined
Mosaic sectors’ concentration rasters to one seamless raster for entire plume. along gaining streams and lakes. No-flow boundary cells are set
To ease the application of the FGI method, it is recommended to process and where model boundaries coincide with groundwater flow lines (see
manage all available site data in the GIS. The FGI method can also be implemented Fig. 3). Hydraulic conductivity values of the calibrated model range
in GIS packages other than ArcGIS. Open source GI-Systems and packages include from K ¼ 6.23$108 m/s to K ¼ 1.32$103 m/s (geometric mean
e.g. GRASS-GIS, Quantum GIS, the geodata abstraction library GDAL (http://www. Kgeom ¼ 1.34$105 m/s), aquifer thickness from 3.6 to 20 m. The
gdal.org) or R with the rgdal-package (GRASS Development Team, 2009;
Quantum GIS Development Team, 2009; R Development Core Team, 2009). The
model calibration is based on hydraulic heads measured at 27
pseudo-code algorithms presented above may serve as a guideline for such an groundwater monitoring wells and 20 temporary direct push wells
implementation process. The FGI method presented in this paper may also be (Rein et al., Submitted for publication). The groundwater level varies
adapted for using other codes to simulate groundwater flow and advective trans- approx. between 2 and 6 m below ground surface.
port (particle tracking).

Fig. 4. Modeled TCE concentration distribution in groundwater in area of interest. C(sat) denotes saturation concentration.
 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780 1775

Table 2 In correspondence to this stepwise site investigation since 1996,

Sample types and reference concentrations, for X,Y-coordinate, refer to Table 3. four different knowledge states (investigation stages), representing
Type Name Samples reference (mg/l) different levels of extent and density of concentration data, can be
Well B 18 313666 distinguished (Schädler et al., 2008). For TCE, which is used here as
Well B 20 946668 a model compound, the first level of information is comprised of
Well B 21 181586 concentration measurements at 4 groundwater wells. This level is
Well B 22 1848830
hereafter denoted as ”Wells”. Two direct-push campaigns provided
DP1 PK72 141450
DP1 PK73 135161 further data (three additional points in campaign #1 and another
DP1 PK79 976230 one in campaign # 2) and result in investigation levels ”DP1” and
DP2 72-B-3 248389 ”DP2”. Finally, an intensive sampling of tree-cores (Holm, 2007,
Tree-core T72 125708 2009; Rein et al., 2008) in the area of interest strongly improved
Tree-core T73 135161
Tree-core T79 1200000
the measurement network (level ”Tree-core”). Tree-core sampling
Tree-core T128 38730 utilizes contaminant accumulation in plants providing semi-
Tree-core T129 36260 quantitative information about shallow tetrachloroethene and tri-
Tree-core T133 2473 chloroethene contaminations in soils and groundwater (e.g.
Tree-core T134 11709
Schumacher et al., 2004; Larsen et al., 2008).
Tree-core T141 2397
Tree-core T151 860 In this study we focus on one contaminant plume in the
Tree-core T152 218 northern part of the site. Since little information about the vertical
Tree-core T153 847 contaminant distribution was available, only a two-dimensional
Tree-core T154 193 analysis was performed.
Tree-core T155 86
Tree-core Tz8 1733

Fig. 5. Results of the FGI method at different investigation levels in comparison to ordinary kriging (OK) and the reference distribution. Number of samples at different investigation
levels increases in the following order: “Wells”, “DP 1”, “DP 2”, and “Tree-core”.
1776 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780

4. Method evaluation
RMSE
NRMSE ¼ (3)
4.1. Purpose max cest  min cest
with max cest and min cest being the maximum and minimum
In order to examine the appropriateness and performance of the values of estimated concentrations in all grid elements covering the
proposed FGI method, it was applied to a synthetic close-to-reality plume area.
test case. Through the creation of a data grid of reference values of
TCE concentration, we could evaluate the performance of the 4.4. FGI performance
method for different values of segmentation criterion Sseg and
different investigation stages, simply by comparing the estimated The FGI method was applied to the test case with different sets
concentration distribution with the assumed ‘true’ distribution, as of sampling data and different values of segmentation criterion Sseg.
given in the form of the reference data grid. The latter was varied between Sseg ¼ 50 and Sseg ¼ 0.5. The resulting
segmented plumes consist of 2, 5, and 8 sectors, accordingly

4.2. Evaluation test case

The continuous reference values of TCE concentration in the

area of interest (see Fig. 4 and Table 2) were generated by imple-
menting a transport model using MT3DMS (Zheng and Wang,
1999). Data input regarding the heterogeneous flow field was
taken from the MODFLOW results (see previous section). Transport
parameters were set based on field data (e.g. from tracer tests and
soil column analyses). The total organic carbon content foc varies
between 0.02 and 1.17%. The effective porosity ne ranges from 0.10
to 0.34. The distribution coefficient Kd for TCE was set as the
product of ’measured’ foc values and literature values of the sorp-
tion constant Koc (e.g. Mehran et al., 1987; Schwarzenbach et al.,
2003), assuming that pure forward modeling serves the purpose
of generating a reasonable close-to-reality contaminant concen-
tration data field. In accordance with observations made in the
field, a steady-state plume and a constantly emitting source zone
were assumed. The constant concentration boundary condition
equals the saturation concentration of TCE in groundwater
(Csource ¼ 1.2$106mg/l). Longitudinal and lateral dispersivity were set
to 1 m and 0.01 m, respectively. In-situ degradation of TCE was
modeled as 1st order process using a degradation rate constant of
0.002 d1, in accordance with literature (Gorder et al., 1996, 1997).
The generated plume is shown in Fig. 4. It should be pointed out
that the transport model is used here only to produce the reference
data for the method’s evaluation. It is, however, not needed for the
application of the FGI method itself.
The generated plume (i.e. the modeled concentration data grid)
was then sampled at the sample locations given by the different
investigation campaigns. For each of the four campaigns one set of
synthetic concentration point measurements was produced (see
Table 2).

4.3. Evaluation method

The quality of the FGI method was quantified by comparing the

estimated concentration values of all grid elements with the
respective values in the reference data grid. We utilized the
normalized root mean square error (NRMSE) to obtain an overall
measure of quality (Eq. (3)). The root mean square error (RMSE) is
determined by:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2ffi
uP
u j ref
t i ¼ 1 ci  ci
est
RMSE ¼ ; (2)
j
where j is the number of grid node elements considered within the
Fig. 6. Normalized root mean square error NRMSE for concentration estimates as
plume, cref est
i and ci denotes the concentration values of the refer- a function of increased knowledge status (cf. Fig. 5). Comparison of results from FGI
ence distribution and of the estimated distribution, respectively. conducted with ordinary kriging (OK) and the inverse distance weighted method
The NRMSE is further calculated by: (IDW), and from ’un-guided’ interpolation with OK and IDW.
 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780 1777

(compare Fig. 2b). The number of samples that are located in the propagation of estimates from upgradient sectors. Further analysis
individual sectors and the parameters used for sector wise inter- reveals that FGI particularly outperforms OK in the case of sparse
polation are listed in Table 1. When using a plume divided into only sampling data. Quantitative assessment also supports these find-
two sectors, for example, the second sector contains no real ings (see Fig. 6). While OK (standard interpolation) exhibits
samples for the investigation levels “Wells”, “DP1”, and “DP2”, a distinct dependency on available sampling information, the FGI
whereas the fourth investigation level, “Tree-core”, provides 9 method shows a rather consistent performance. Guiding the
sample points. Similar conditions apply to the cases with 5 and 8 interpolation by flow information obviously compensates for
sectors, respectively, where sectors located further downgradient missing data in downgradient plume regions. Fig. 6 also indicates
also do not contain sample information for the first three investi- a shift in the ranking of the results obtained with the three
gation levels. Note also that the density of available data may different degrees of segmentation. Although the NRMSE decreases
fluctuate from sector to sector depending on the distribution of with increasing number (density) of available sampling points for
given sample points (e.g. sequence of sectors 1e2e3e4 in the case all degrees of segmentation, more data (levels “DP2” and “Tree-
of 8 sectors). The interpolation was performed with the commercial core”) seems to favor a finer segmentation while a coarser
computer program Surfer 8 (Golden Software) using ordinary segmentation apparently performs better when little data is
(point) kriging without modeling a variogram nor incorporating available (levels “Wells” and “DP1”). For the purpose of compar-
slope or trend. The resulting estimation grids were converted to ison, interpolation was also done by the inverse distance weighted
ESRI ASCII grids with the free software GridConvert (www. method (IDW) using squared distances and the same anisotropy
geospatialdesigns.com) for further processing within GIS. The angle and ratio as was used for OK (see Table 1). Although IDW
comparison of the estimations to the reference distribution was does not perform as well as OK, again flow guidance clearly
calculated using MatlabÓ(R2009a, The MathWorks, Natick, improves the quality of the results as can be seen from the NRMSE
Massachusetts). values given in Fig. 6.
Fig. 5 shows the matrix of evaluation results obtained with the
FGI method depending on the number of sectors (matrix rows) at 5. Method application
different investigation levels (matrix columns). For the purpose of
comparison, the last row displays the results obtained with ‘un- Based on the results of the evaluation described in the previous
guided’ ordinary kriging (OK). A visual inspection of the results section, the FGI method was applied to the real sampling data sets
suggests that the FGI method provides reasonable concentration available for the area of interest. Segmentation settings that per-
distributions for all cases. Missing sampling points in the farther formed best in the evaluation were chosen for each of the four
downgradient sections result in slightly overestimated concen- investigation levels (comp. Fig. 6): 2, 5, and 8 sectors were
trations in these sectors. This might be explained by the respectively used for levels “Wells”, “DP1”, and both “DP2” and

Fig. 7. Results of segmented interpolations using field data within plume.

1778 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780

Table 3 Naturally, no absolute quantitative assessment of the results

Sample locations and concentrations used for interpolation procedures at field could be carried out since true values at unsampled locations are
application.
unknown. Still, visual inspection reveals evident differences
Type Name GK coordinates Samples field Tree-core between the results obtained with the FGI method and OK. The
(mg/l) () FGI method yielded plausible results at all investigation levels.
X Y
Well B 18 3366921 5814676 77440 7 Results at low investigation levels appear to be confirmed by
Well B 20 3366868 5814635 119703 8 results at higher levels. The series of estimated concentration
Well B 21 3366898 5814708 72962 7 distributions exhibits a successive refinement (see Fig. 7, from left
Well B 22 3366945 5814682 174181 8
to right). In contrast, OK results were found to be rather arbitrary
DP1 PK72 3366874 5814708 6089 6
DP1 PK73 3366959 5814684 820 5
and strongly dependent on available data. This disparity between
DP1 PK79 3366859 5814632 121605 8 the methods’ results is particularly apparent when comparing the
DP2 72-B-3 3366917 5814697 8229 6 plumes’ shape at level “DP2” with the shape at level “Tree-core”. It
Tree-core T72 3366885 5814719 e 7 should be mentioned that the FGI conducted for level “Tree-core”
Tree-core T73 3366951 5814689 e 6
was performed in the same way as for the groundwater concen-
Tree-core T79 3366859 5814627 e 9
Tree-core T128 3366991 5814731 e 6 tration values of investigation levels “Wells”, “DP 1” and “DP 2”.
Tree-core T129 3366884 5814752 e 6 Thus, the inputs for “Tree-core” are scaled using integer intervals
Tree-core T133 3367091 5814723 e 4 but the estimated results on the grid consist of floating-point
Tree-core T134 3367044 5814761 e 3 numbers.
Tree-core T141 3367116 5814772 e 6
Tree-core T151 3367130 5814823 e 5
Tree-core T152 3367190 5814830 e 5 6. Summary and conclusions
Tree-core T153 3367074 5814849 e 3
Tree-core T154 3367157 5814875 e 3 In this paper we introduced and tested a new method, flow
Tree-core T155 3367118 5814893 e 3
guided interpolation (FGI), for generating continuous concentra-
Tree-core Tz8 3367027 5814837 e 4
tion data in the area of contaminant plumes in groundwater. In
comparison to pure interpolation by ordinary kriging, FGI offers
clear advancement in the use of modeled data on flow direction and
“Tree-core”. The application results are depicted in Fig. 7. Once
the respective shape of the contaminant plume: (i) the sampling
again, FGI results are compared with the results from OK.
point list used for interpolation is complemented by auxiliary
Please note that the application at the “Tree-core” level had to
sampling points along the plume fringe and (ii) the plume area is
take into consideration that tree-core sampling provides only
segmented into a series of plume sectors, depending on the plume’s
semiquantitative values. Results from tree-core sampling were
curvature. The interpolation is therefore termed “flow guided”. The
provided on a scale from 1 (below detection limit) to 9 (high
new method also features a sequential sector-wise interpolation in
concentration in groundwater close to solubility) representing peak
downgradient direction. Results generated in one sector are used
areas determined by gas chromatography. To get a complete and
for subsequent interpolation of the next sector, i.e. information is
consistent sample list for the “Tree-core” level, we converted the
virtually propagated downgradient. In this way, the FGI method
groundwater concentration values measured at the sampling
counteracts the decline in available measurement data with
points of the first three investigation stages to the semiquantitative
increasing distance to the source zone, as is typically observed in
tree-core scale (Table 3) using an approximative log-scaled
practice.
conversion scheme (Table 4). Comparison of converted values with
The FGI method has been implemented in ArcGIS using VBA
tree-core results at pairs of sampling points located close to each
code to automize major steps of the method’s procedures. Based on
other (T79 vs. PK79: 9 vs. 8; T72 vs. B21: 7 vs. 7; T73 vs. PK73: 6 vs.
the detailed descriptions and illustrations given in this paper, these
5) shows a good agreement and confirms the adequacy of this
procedures may easily be transferred to other geographical infor-
approximation for this particular case.
mation systems that allow for expanding and customizing
The ASP values along the downgradient edge of the source zone
functionalities.
were set to Csource ¼ 1.2 $ 10−5 mgl1 representing the mean measured
The evaluation of the method using varying segmentation
TCE concentrations in wells B20 and PK79, located in close prox-
settings and sample data sets of different size showed the
imity to the source zone. For level “Tree-core” the source zone
robustness of the method. Estimated concentration distributions
concentration was accordingly set to Csource ¼ 9 (compare Tables 3
agree well with the reference distribution, independent of the
and 4). The threshold concentration values Ctc assigned to the ASP setting and the data set which have been chosen. Results from the
on the lateral sector boundaries were set to zero at all investigation
method’s application to a model site confirm its practicability. The
stages. advantage of the FGI method over “unguided” ordinary kriging
was shown to be especially large when available sampling data is
scarce. Although a rigorous analysis of the FGI method’s perfor-
Table 4 mance by means of numerical experiments remains to be
Conversion scheme for tree-core sample analogues at well and DP locations. completed, the results presented in this paper obviously demon-
Tree-core () Hypothetical analog concentration strate that the method meets the demands of a realistic repre-
for groundwater (mg/l) sentation of contaminant plumes in groundwater at early
1 Unverifiable investigation stages.
2 <Detection limit (<0.1)
3 <10
Acknowledgments
4 10e99
5 100e999
6 1000e9999 The authors gratefully acknowledge the support of the German
7 10000e99999 BMBF research priority program REFINA (contract no. 0330757C).
8 100000e999999 Funding for this study also has been provided by the Helmholtz
9 >1000000
Centre for Environmental Research e UFZ in Leipzig (contract no.
 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780 1779

4500029698). The authors would also like to thank Margaret Hass

for language editing of the manuscript.

Appendix A. List of symbols

Table A.5 List of symbols, their definitions and units.

 
as Anisotropy direction angle for sector s ( , E ¼ 0 , CCW)
 
bi Angle between adjacent plume center pathline elements ( , E ¼ 0 , CCW)
cest
i Estimated concentration value on grid node j (m gl1)
cref
i Reference concentration value on grid node j (m gl1)
Clb, s Concentration at downgradient boundary of sector s (m gl1)
Csource Source concentration (m gl1)
Ctc Threshold concentration at lateral plume fringe (m gl1)
Cub, s Concentration at upgradient boundary of sector s (m gl1)
h Separation vector of sample pair (m)
i Generic increment variable
j Number of grid node elements in estimated raster grids
L Longest dimension of contaminant plume in direction of separation vector h (m)
lcrit, i Normal distance of plume center pathline node to elongation line (m)
li Length of center pathline element between to nodes (m)
Ltot Total plume length in direction of groundwater flow (m)
nstep Generic increment variable
n Total number of points on plume’s center pathline
N Number of sample pairs
RAs Anisotropy ratio in sector s
s Sector number
Sseg Segmentation criterion (m)
Ws Mean width of sector s (m)
x Point xecordinate (m Gauss-Krüger)
y Point yecordinate (m Gauss-Krüger)
xbd, s Sector boundary point xecordinate (m Gauss-Krüger)
ybd, s Sector boundary point yecordinates (m Gauss-Krüger)

References assessment for contaminated sites: Part i. a multi-criteria based system for the
selection of ecotoxicological tests and ecological observations. Science of the
Total Environment 379 (1), 16e33. doi:10.1016/j.scitotenv.2007.02.035.
Anderson, M.P., Cherry, J.A., 1979. Using models to simulate the movement of
Crumbling, D.M., Lynch, K., Howe, R., Groenjes, C., Shockley, J., Keith, L., Lesnik, B.,
contaminants through groundwater flow systems. Critical Reviews in Envi-
EJ.V., McKenna, J., 2001. Peer reviewed: managing uncertainty in environmental
ronmental Science and Technology 9 (2), 97e156.
decisions. Environmental Science & Technology 35 (19), 404Ae409A.
Basu, N.B., Rao, P.S.C., Poyer, I.C., Annable, M.D., Hatfield, K., 2006. Flux-based
doi:10.1021/es012490g.
assessment at a manufacturing site contaminated with trichloroethylene.
D’Affonseca, F.M., Blum, P., Finkel, M., Melzer, R., Grathwohl, P., 2008. Field scale
Journal of Contaminant Hydrology 86 (1e2), 105e127.
characterization and modeling of contaminant release from a coal tar source
Batu, V., 2006. Applied Flow and Solute Transport Modeling in Aquifers: Funda-
zone. Journal of Contaminant Hydrology 102 (1e2), 120e139. partial Source
mental Principles and Analytical and Numerical Methods. Taylor & Francis, Boca
Zone Removal.
Raton, FL.
Deutsch, C.V., Journel, A.V., 1997. GSLIB: Geostatistical Software Library and User’s
Bayer-Raich, M., Jarsjo, J., Teutsch, G., 2009. Breakthrough of attenuating contami-
Guide, second ed. Oxford University Press, New York.
nant plumes in pumping wells: analytical model and implications for integral
Eigbe, U., Beck, M.B., Wheater, H.S., Hirano, F., 1998. Kalman filtering in groundwater
pumping tests. Water Resources Research 45. doi:10.1029/2007wr006075.
flow modelling: problems and prospects. Stochastic Hydrology and Hydraulics
Bockelmann, A., Ptak, T., Teutsch, G., 2001. An analytical quantification of mass
12 (1), 15e32.
fluxes and natural attenuation rate constants at a former gasworks site. Journal
ESRI, 1992e2005. Arcmap [gis software]. Version 9.1.
of Contaminant Hydrology 53 (3e4), 429e453.
ESRI, 2009. Esri developer network documentation library e groundwater
Boufassa, A., Armstrong, M.P., 1989. Comparison between different kriging esti-
modeling. http://resources.esri.com/help/9.3/ArcGISDesktop/com/samples/
mators. Mathematical Geology 21 (3), 331e345.
SpatialAnalyst/groundwatermodeling/f63b8241-ad5d-45c1-af6b-
Burger, H., Schafmeister, M.-T., 2000. Gerichtete Interpolation zur verbesserten
ee1aa1ae042b.htm (last accessed 15.09.09).
Darstellung strömungsabhängiger Grundwasserbeschaffenheitsmerkmale.
Evensen, G., 1994. Sequential data assimiliation with a nonlinear quasi-geostrophic
Grundwasser 5 (2), 79e85.
model using Monte Carlo methods to forecast error statistics. Journal of
Burke, R., Arana, A., 2003. Getting to Know ArcObjects (With CD-ROM), pap/cdr
Geophysical Research 99 (C5), 143e153.
Edition. ESRI Press.
Fuest, S., Berlekamp, J., Klein, M., Matthies, M., 1998. Risk hazard mapping of
Chang, S.Y., Jin, A., 2005. Kalman filtering with regional noise to improve accuracy of
groundwater contamination using long-term monitoring data of shallow
contaminant transport models. Journal of Environmental Engineering-Asce 131
drinking water wells. Journal of Hazardous Materials 1998 (61), 197e202.
(6), 971e982.
Gorder, K.A., Dupont, R.R., Sorensen, D.L., Kemblowski, M.W., McLean, J.E., 1996.
Chang, S.-Y., Latif, S.M.I., 2009. Use of kalman filter and particle filter in a one
Analysis of intrinsic bioremediation of trichloroethene-contaminated ground
dimensional leachate transport model. In: Proceedings of the 2007 National
water at Eielson Air Force Base, Alaska. In: Symposiom on Natural Attenuation
Conference on Environmental Science and Technology. Springer, pp. 157e163.
of Chlorinated Organics in Ground Water. Dallas, Texas., pp. 104e109.
Chang, S.Y., Latif, S.M.I., 2010. Extended kalman filtering to improve the accuracy of
Gorder, K.A., Dupont, R.R., Sorensen, D.L., Kemblowski, M.W., McLean, J.E., 1997.
a subsurface contaminant transport model. Journal of Environmental Engi-
Field evaluation of intrinsic bioremediation of chlorinated solvents. In Situ and
neering-Asce 136 (5), 466e474.
On-Site Bioremediation 4 (3), 189531.
Chen, Z., 2003. Bayesian Filtering: From Kalman Filters to Particle Filters, and
GRASS Development Team, 2009. Geographic Resources Analysis Support System
Beyond. Tech. rep.. Communications Research Laboratory, McMaster University,
(GRASS GIS) Software. Open Source Geospatial Foundation, USA. http://grass.
Hamilton, Ontario, Canada.
osgeo.org.
Chu, W.S., Strecker, E.W., Lettenmaier, D.P., 1987. An evaluation of data requirements
Harbough, A., McDonald, M.G., 1996. User’s documentation for modflow-96, an
for groundwater contaminant transport modeling. Water Resources Research
update to the U.S. geological survey modular finite-difference ground-water
23 (3), 408e424.
flow model. Tech. Rep. Open-File Report 96e485, U.S. Geological Survey.
Cooper, R.M., Istok, J.D., 1988. Geostatistics applied to groundwater contamination.
Holm, O., 2007. Untersuchung von Bohrkernen aus Baumstämmen am Standort
i: methodology. Journal of Environmental Engineering 114 (2), 270e286.
Krampnitzt zum Nachweis von Schadstoffen mittels HS-SPME/GC-MS. Tech.
Critto, A., Torresan, S., Semenzin, E., Giove, S., Mesman, M., Schouten, A.J.,
rep., Technische Universität Berlin, Fachgebiet Umweltchemie, Interim project
Rutgers, M., Marcomini, A., 2007. Development of a site-specific ecological risk
1780 M. Morio et al. / Environmental Modelling & Software 25 (2010) 1769e1780

report for BMBF-project REFINA e SINBRA, contract no. 0330757D, http://www. Neupauer, R.M., Wilson, J.L., 2004. Numerical implementation of a backward proba-
sinbra.de. bilistic model of ground water contamination. Ground Water 42 (2), 175e189.
Holm, O., 2009. Teilvorhaben 5 “Demonstrations- und Modellstandort Potsdam- O’Reilly, M., Brink, R., 2006. Initial risk-based screening of potential brownfield
Krampnitz: Parametrisierung, Validierung und Visualisierung der Methoden”. Tech. development sites. Soil & Sediment Contamination 15 (5), 463e470.
rep., Technische Universität Berlin, Fachgebiet Umweltchemie, Final project report Pollock, D.W., 1994. User’s guide for modpath/modpath-plot, version 3: A particle
for BMBF-project REFINA e SINBRA, contract no. 0330757D, http://www.sinbra.de. tracking post-processing package for modflow. Tech. Rep. Open-File Report
Huang, C., Hu, B., Li, X., Ye, M., 2009. Using data assimilation method to calibrate 94e464, U.S. Geological Survey.
a heterogeneous conductivity field and improve solute transport prediction Porter, D.W., Gibbs, B.P., Jones, W.F., Huyakorn, P.S., Hamm, L.L., Flach, G.P., 2000.
with an unknown contamination source. Stochastic Environmental Research Data fusion modeling for groundwater systems. Journal of Contaminant
and Risk Assessment 23 (8), 1155e1167. Hydrology 42 (2e4), 303e335.
Isaaks, E.H., Srivastasa, R.M., 1989. Applied Geostatistics. Oxford University Press, Prommer, H., Barry, D.A., Davis, G.B., 2002. Modelling of physical and reactive
New York. processes during biodegradation of a hydrocarbon plume under transient
Istok, J.D., Rautman, C.A., 1996. Probabalistic assessment of ground-water contam- groundwater flow conditions. Journal of Contaminant Hydrology 59 (1e2),
ination: 2. results of case study. Ground Water 34 (6), 1050e1064. doi:10.1111/ 113e131.
j.1745-6584.1996.tb02171. Quantum GIS Development Team, 2009. Quantum GIS Geographic Information
Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. Academic Press, London, System. Open Source Geospatial Foundation. http://qgis.osgeo.org.
New York. R Development Core Team, 2009. R: a Language and Environment for Statistical
Journel, A.G., Rossi, M.E., 1989. When do we need a trend model in kriging. Computing. R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-
Mathematical Geology 21 (7), 715e739. 900051-07-0. http://www.R-project.org.
Kalman, R.E., 1960. A new approach to linear filtering and prediction problems. Rautman, C.A., Istok, J.D., 1996. Probabalistic assessment of ground-water contam-
Transactions of the ASMEeJournal of Basic Engineering 82 (Series D), 35e45. ination: 1. geostatistical framework. Ground Water 34 (5), 899e909.
Kerry, R., Oliver, M., 2007. Determining the effect of asymmetric data on the vari- doi:10.1111/j.1745-6584.1996.tb02084.x.
ogramm. i. underlying asymmetry. Computers and Geosciences 33, 1212e1232. Razavi, A.H., 2004. ArcGIS Developer’s Guide for Visual Basic Applications, second
King, M.W.G., Barker, J.F., 1999. Migration and natural fate of a coal tar creosote ed. OnWord Press.
plume 1. overview and plume development. Journal of Contaminant Hydrology Reed, P.M., Ellsworth, T.R., Minsker, B.S., 2004. Spatial interpolation methods for
39 (3e4), 249e279. nonstationary plume data. Ground Water 42 (2), 190e202. doi:10.1111/j.1745-
Kistemann, T., Hundhausen, J., Herbst, S., Claßen, T., Färber, H., 2008. Assessment of 6584.2004.tb02667.x.
a groundwater contamination with vinyl chloride (vc) and precursor volatile Rein, A., Holm, O., Popp, S., Bittens, M., Rotard, W., Dietrich, P., 2008. Revitalisation
compound (voc) by use of a geographical information system (gis). Interna- of rural megasites e adaptive strategies and technologies for risk-based site
tional Journal of Hygiene and Environmental Health 211, 308e317. characterisation and monitoring. In: ConSoil 2008. Milano, pp. C93eC100.
Kitanidis, P.K., Shen, K.F., Dec 1996. Geostatistical interpolation of chemical Rein, A., Popp, S., Leven, C., Bittens, M., Dietrich, P. Comparison of approaches for the
concentration. Advances in Water Resources 19 (6), 369e378. characterization of contamination at rural megasites. Environmental Geology,
Kübert, M., Finkel, M., 2006. Contaminant mass discharge estimation in ground- Submitted to publication.
water based on multi-level point measurements: a numerical evaluation of Schumacher, J., Struckhoff, G., Burken, J., 2004. Assessment of subsurface chlori-
expected errors. Journal of Contaminant Hydrology 84 (1e2), 55e80. nated solvent contamination using tree cores at the front street site and
Larsen, M., Burken, J., Machackova, J., Karlson, U.G., Trapp, S., 2008. Using tree core a former dry cleaning facility at the riverfront superfund site, new haven,
samples to monitor natural attenuation and plume distribution after a pce spill. missouri, 1999e2003. Tech. Rep. Scientific Investigations Report 2004-5049,
Environmental Science & Technology 42 (5), 1711e1717. U.S. Department of the Interior & U.S. Geological Survey.
Liu, G., Chen, Y., Zhang, D., 2008. Investigation of flow and transport processes at the made Schwarz, R., Ptak, T., Holder, T., Teutsch, G., 1998. Groundwater risk assessment at
site using ensemble kalman filter. Advances in Water Resources 31 (7), 975e986. contaminated sites: a new investigation approach. Groundwater Quality:
Mack, J., Crumbling, D.M., Ellerbusch, F., 2004. A data integration framework to Remediation and Protection 250, 68e71.
support triad projects. Remediation 15 (1), 21e37. Schwarzenbach, R.P., Gschwend, P.M., Imboden, D.M., 2003. Environmental Organic
MacKay, D.M., 1990. Characterization of the distribution and behaviour of Chemistry, second ed. Wiley, Hoboken, N.J.
contaminants in the subsurface. In: Conway, R.A. (Ed.), Ground Water and Soil Schädler, S., Morio, M., Finkel, M., 2008. Land use related cost estimates for
Contamination Remediation: toward Compatible Science, Policy, and Public contaminated site development: consequences of uncertainty to planning and
Perception: Report on a Colloquium Sponsored by the Water Science and invest-ment decisions. In: Trefry, M. (Ed.), Groundwater Quality 2007: Securing
Technology Board. National Academy Press, Washington, pp. 70e90. Groundwater Quality in Urban and Industrial Environments. IAHS publication
McGrath, W.A., Pinder, G.F., 2003. Search strategy for groundwater contaminant plume no. 324. IAHS Redbook. IAHS-Press, Fremantle, Western Australia, pp. 539e546.
delineation. Water Resources Research 39 (10), 1298. doi:10.1029/2002WR001636. Shlomi, S., Michalak, A.M., 2007. A geostatistical framework for incorporating
Mehran, M., Olsen, R.L., Rector, B.M., 1987. Distribution coefficient of trichloroeth- transport information in estimating the distribution of a groundwater
ylene in soil-water systems. Ground Water 25 (3), 275e282. contaminant plume. Water Resources Research 43, 3412. doi:10.1029/
Mehrjardi, R.T., Jahromi, M.Z., Sh, M., Heidari, A., 2008. Spatial distribution of 2006WR005121.
groundwater quality with geostatistics (case study: Yazd-ardakan plain). World Storck, P., Eheart, J.W., Valocchi, A.J., 1997. A method for the optimal location of
Applied Sciences Journal 4 (1), 9e17. monitoring wells for detection of groundwater contamination in three-
Meyer, P.D., Valocchi, A.J., Eheart, J.W., 1994. Monitoring network design to provide dimensional heterogenous aquifers. Water Resour. Res. 33 (9), 2081e2088.
initial detection of groundwater contamination. Water Resources Research 30 Sudicky, E.A., Cherry, J.A., Frind, E.O., 1983. Migration of contaminants in ground-
(9), 2647e2659. water at a landfill: a case study: 4. a natural-gradient dispersion test. Journal of
Michalak, A.M., Kitanidis, P.K., 2004a. Application of geostatistical inverse modeling Hydrology 63 (1e2), 81e108.
to contaminant source identification at dover afb, delaware. Journal of Warrick, A., Young, M., Wierenga, P., 1998. Probabilistic analysis of monitoring
Hydraulic Research 42 (Extra Issue), 9e18. systems for detection of subsurface contaminant plumes. Ground Water 36 (6),
Michalak, A.M., Kitanidis, P.K., 2004b. Estimation of historical groundwater distri- 894e900.
bution using the adjoint state method applied to geostatistical modeling. Water Zheng, C., Wang, P., 1999. MT3DMS, a Modular Three-dimensional Multi-species
Resources Research 40 (8), W08302. doi:10.1029/2004WR003214. Transport Model for Simulation of Advection, Dispersion and Chemical Reac-
Michalak, A.M., Kitanidis, P.K., 2005. A method for the interpolation fo nonnegative tions of Contaminants in Groundwater Systems: Documentation and User’s
functions with an application to contaminant load estimation. Stoch Environ Guide. Tech. Rep. Contract Report SERDP-99e1. U.S. Army Engineer Research
Res Risk Assess 19, 8e23. doi:10.1007/s00477-004-0189-1. and Development Center.