See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/303687650

Geospatial estimation of the impact of

Deepwater Horizon oil spill on plant oiling
along the Louisiana shorelines

Article in Journal of Environmental Management · May 2016

DOI: 10.1016/j.jenvman.2016.05.041

CITATIONS READS

3 162

4 authors, including:

Pierre Goovaerts Cameron W. Wobus

BioMedware, Inc Abt Associates
238 PUBLICATIONS 11,940 CITATIONS 63 PUBLICATIONS 1,937 CITATIONS

SEE PROFILE SEE PROFILE

Russell Jones
Abt Environmental Research, Boulder, Colora…
32 PUBLICATIONS 509 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

METRIC -- measurement of environmental metrics for cancer research View project

Spatial study of Lung, Liver, Colon and Cervical Cancer mortality in the American West, Particularly
Utah View project

All content following this page was uploaded by Pierre Goovaerts on 07 June 2016.

The user has requested enhancement of the downloaded file.

 Journal of Environmental Management 180 (2016) 264e271

Contents lists available at ScienceDirect

Journal of Environmental Management

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

Research article

Geospatial estimation of the impact of Deepwater Horizon oil spill on

plant oiling along the Louisiana shorelines
Pierre Goovaerts a, *, Cameron Wobus b, Russell Jones b, Matthew Rissing b
a
PGeostat LLC, Jerome, MI 49249, USA
b
Abt Associates, Boulder, CO 80302, USA

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

Article history: Stranded oil covering soil and plant stems in fragile Louisiana marshes was one of the most visible
Received 9 February 2016 impacts of the 2010 Deepwater Horizon (DWH) oil spill. As part of the assessment of marsh injury after
Received in revised form the DWH spill, plant stem oiling was broken into five categories (0%, 0e10%, 10e50%, 50e90%, 90e100%)
13 April 2016
and used as the independent variable for estimating death of vegetation, accelerated erosion, and other
Accepted 17 May 2016
metrics of injury. The length of shoreline falling into each of these stem oiling categories was therefore a
key measure of the total extent of marsh injury, and its accurate estimation is the focus of this paper.
First, we used geographically-weighted logistic regression (GWR) to explore and model spatially varying
Keywords:
Marsh
relationships between stem oiling field data and secondary information (oiling exposure category)
Geographically-weighted regression collected during shoreline surveys. We then combined GWR probability estimates with field data using
Cokriging indicator cokriging to predict the probability of exceeding four stem oiling thresholds (0, 10, 50, and 90%)
Indicator kriging at 50 m intervals along the Louisiana shoreline. Cross-validation using Receiver Operating Characteristic
False positives (ROC) Curves demonstrate the greater prediction accuracy of the multivariate geostatistical approach
relative to either aspatial regression or indicator kriging that ignores secondary information.
© 2016 Published by Elsevier Ltd.

1. Introduction categories within the shoreline exposure database are qualitative,

Studies of vegetation death and accelerated marsh erosion
following Deepwater Horizon (DWH) have shown that both of these
impacts (or “injuries,” when assessing natural resource damages)
can be related to the percent of oiling on the stems of marsh
vegetation (e.g., Hester et al., 2015; Silliman et al., 2015). Spatial
quantification of these injuries thus relies on estimates of how
many kilometers of shoreline fell into each of the four stem oiling
categories on which these injury determinations were based
(0e10%, 10e50%, 50e90%, 90e100%). Vegetation oiling from the
DWH spill was unevenly distributed across Louisiana marsh envi-
ronments, however, and quantitative measurements of stem oiling
were collected only at discrete points (Deepwater Horizon NRDA,
2010a). Spatially continuous observations of shoreline oiling were
collected as part of response activities and the natural resource
damage assessment (NRDA), and these data were combined into a
“shoreline exposure” database for the NRDA (Deepwater Horizon
NRDA, 2010b; Nixon, 2015; NOAA, 2013). However, the oiling
* Corresponding author. PGeostat LLC, 11487 Highland Hills Drive, MI 49249, USA.
E-mail address: goovaerts.pierre@gmail.com (P. Goovaerts).
and do not contain direct information on stem oiling. Furthermore,
due to the scope of the DWH spill and the difficulty of finding oil in
marshes, these qualitative shoreline surveys sometimes docu-
mented segments as “no oil observed” (NOO) in places where more
detailed surveys documented oiling at other points in time.
Recognizing the relative strengths and limitations of both of these
oiling datasets, the goal of this study was to test and apply geo-
spatial methods for combining quantitative point observations of
stem oiling with continuous, qualitative observations in the
shoreline exposure database to estimate the length of shoreline
falling into each of the five stem oiling categories.
One way of quantifying the length of shoreline falling into each
stem oiling category would be to assume that measured stem oiling
values were evenly distributed within each exposure category, and
to calculate the length of each stem oiling category based on pro-
portional assignments within the shoreline exposure framework.
However, stem oiling data were often clustered in space, violating
the assumption of equal distribution within exposure categories.
This is a particular concern for apportionment of stem oiling data
that were recorded within the NOO category (i.e. false negatives),
since there were thousands of kilometers of shoreline within this

http://dx.doi.org/10.1016/j.jenvman.2016.05.041
0301-4797/© 2016 Published by Elsevier Ltd.
 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271
category but nonzero stem oiling observations within this category
were generally clustered in space.
An alternative is to use a geostatistical approach that can ac-
count for the geographical location of the environmental data and
their spatial correlation, as modeled by variograms (e.g., Goovaerts
et al., 2008; Kitsiou and Karydis, 2011). The application of geo-
statistics to this particular dataset, however, presented several
challenges. First, field data were collected with different spatial
resolutions and degrees of reliability. Some data represent precise
measurements of percentage of stem oiling at specific locations
(“hard data”). At other specific locations, we only know whether
the vegetation was oiled or not; we do not know the percentage of
stem oiling (“soft data”). Finally, although the shoreline exposure
database provides a more comprehensive spatial coverage, the
oiling descriptors in this dataset are more qualitative with respect
to stem oiling (“secondary information”).
The second major challenge arises from the complex geometry
of the site. Louisiana has a deeply dissected and crenulated marsh
coastline, and oil was transported into this region via the bays and
inlets that dissect it. As a result, spatial correlation of stem oiling
may be more directly related to over-water distance than straight
line (Euclidian) distance (Baraba
s et al., 2001; Money et al., 2009).
Third, the heterogeneity of the shoreline and the fact that surveys
were conducted by different teams at different times likely impacts
locally the relationship between secondary data (oiling exposure
category) and percentage of stem oiling measured in the field (i.e.
non-stationary relationships). Fourth, the sheer size of the datasets
analyzed (e.g. 1100 field data and 118,151 shoreline grid nodes to
predict) precluded the use of Bayesian methodologies based on
traditional MCMC (Markov chain Monte Carlo) schemes (Gelfand
et al., 2003), while more powerful approaches, e.g. the INLA (inte-
grated nested Laplace approximations) methodology, rely heavily
on numerical methods and computer programming that are
beyond the scope of this study (Martins et al., 2013).
This paper describes the procedure developed to estimate the
expected lengths of mainland herbaceous shoreline in Louisiana
falling into four stem oiling categories: 0e10%, 10e50%, 50e90%,
and 90e100%. The varying reliability of the different pieces of in-
formation was integrated using a soft and hard indicator coding of
the data (Goovaerts, 1997; Hu et al., 2005), whereas geographically-
weighted logistic regression (Fotheringham et al., 2002; Goovaerts
et al., 2015; van Donkelaar et al., 2015) was used to explore and
model spatially varying relationships between stem oiling field
data and secondary information collected during shoreline surveys.
Probabilities of exceeding stem oiling thresholds estimated from
field measurement and survey data were combined using indicator
cokriging (Goovaerts and Journel, 1995). Sensitivity analysis and
cross-validation helped guide the choice of optimal sets of pa-
rameters and investigate the impact of search strategy and distance
metrics on prediction accuracy.
2. Materials and methods
2.1. Datasets
Detailed measurements of stem oiling were collected at 911
discrete points within mainland herbaceous marshes of coastal
Louisiana (Fig. 1). These marshes are located along the edges of
saline to brackish estuaries and bays throughout Louisiana, and are
dominated by the marsh vegetation Spartina alterniflora. Stem oil-
ing measurements were collected in the late summer and early fall
of 2010, as part of a study referred to as the Marsh Pre-Assessment
Study (Deepwater Horizon NRDA, 2010a). At each of these 911 lo-
cations, field data were recorded as the percent of stem height
oiled, and these raw measurements were then condensed into one
265
of the five stem oiling categories described above (0%, 0e10%,
10e50%, 50e90%, or 90e100%). The marsh pre-assessment
dataset also includes 185 additional sites where stem oiling was
simply categorized as “oiled” or “not oiled.” These 185 “soft” data
provide additional information on the spatial distribution of oiled
plant stems (see Fig. 1).
Between 2010 and 2013, spatially continuous descriptions of
oiling were also collected along the Louisiana shoreline as part of
the shoreline cleanup and assessment technique (SCAT) program
(Michel et al., 2013; NOAA, 2013). These observations were
collected primarily to inform response activities, and summarized
oiling along the shoreline using qualitative, categorical descriptors.
As a supplement to the SCAT data, spatially continuous, qualitative
descriptions of shoreline oiling were also collected for the NRDA as
part of the Rapid Assessment program (Deepwater Horizon NRDA,
2010b). Although both of these data sources provide spatially
continuous coverage, they did not include detailed measurements
of stem oiling.
During the injury assessment phase of the NRDA, the SCAT and
Rapid Assessment datasets were combined into a single database
referred to as the shoreline exposure database (Nixon, 2015) dis-
played in Fig. 2. Oiling exposure in this dataset is classified into one
of the following four categories: NOO, lighter oiling, heavier oiling,
and heavier persistent oiling. For the purposes of our study, the
categorical descriptors of shoreline oiling within this dataset are
referred to as “secondary information” on stem oiling. 729 of the
911 locations with hard data on stem oiling were co-located with
secondary information from the shoreline exposure database
(Fig. 3).
The geospatial analysis was thus based on four main types of
data:
(1) Measurement of the percentage of plant stem oiling from the
pre-assessment dataset (911 “hard” data)
(2) Indicators of presence/absence of plant stem oiling from the
pre-assessment dataset (185 “soft” data)
(3) Oiling exposure category (“secondary information”) sur-
veyed along approximately 1600 km of mainland herbaceous
marsh coastline and at 729 of the 911 hard data locations.
(4) Length of shoreline located within 118,151 50
50 m squares
discretizing the Louisiana coastline.
2.2. Methodology
The analysis was conducted using the following software: 1)
SpaceStat 4.0 (Jacquez et al., 2014) for geographically-weighted
regression and variogram modeling, 2) SAS 9.3 (SAS Institute Inc.,
2011) for aspatial logistic regression and the creation of ROC
curves, 3) SGeMS (Remy et al., 2008) and Gslib (Deutsch and
Journel, 1998) for cross-variogram modeling and indicator cokrig-
ing, and 4) code written by Dr. Goovaerts for data manipulation and
computation of expected lengths of shoreline in different cate-
gories of plant stem oiling. The flowchart in Fig. 4 illustrates the
main steps in the analysis, as described below.
2.2.1. Indicator coding of plant stem oiling data
The analysis started with the coding of each percentage of stem
oiling data into a vector of indicators of exceedance of four
thresholds zc ¼ 0, 10, 50, and 90%. Let ua ¼ (xa,ya) be a vector of UTM
coordinates representing the geographical location of a stem oiling
data point, denoted z(ua) for hard data and s(ua) for soft data. The
set of four indicators at any hard data location ua was then con-
structed as:
266 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271

Fig. 1. Geographical location of hard (-) and soft (þ) data on percentage of plant stem oiling that were used in the geospatial analysis. Red color denotes locations where oiling was
observed while blue symbols correspond to zero percentage of plant stem oiling. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)

Fig. 2. Continuous shoreline exposure data (secondary information) compiled on mainland herbaceous marsh shorelines. This linear dataset was discretized into 118,151
computational nodes on a regular, 50-m grid.

At locations of soft data (i.e. presence/absence of oil), an indi-

1 if zðua Þ > zc cator can only be constructed for the first threshold z1 ¼ 0% since
iðua ; zc Þ ¼ c ¼ 1; /; 4 (1)
0 otherwise the exact percentage of plant stem oiling is unknown:
 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271 267

Fig. 3. Oiling exposure category recorded at 729 pre-assessment sites, superimposed on shoreline exposure data shown in Fig. 2.

Fig. 4. Flowchart describing the different steps of the geospatial analysis for the computation of expected length of shoreline falling into specific classes of percentage of plant stem
oiling. Grey boxes represent the input data, light blue boxes correspond to outputs, and green boxes denote key (geo)statistical methods. (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)
268 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271

is combined with indicators of exceedance at n(u) neighboring

1 if sðua Þ ¼ 1 ðoil present Þ
iðua ; z1 Þ ¼
0 otherwise
2.2.2. Predicting probability of plant stem oiling from oiling
exposure category
Because the percentage of stem oiling was recorded only at 911
discrete locations, the extrapolation of this sole information to the
entire shoreline length is challenging. It is thus beneficial to
incorporate the secondary information provided by oiling exposure
category since it is available on a continuous basis over large sec-
tions of the shoreline. Logistic regression was used to create a
predictive model of the likelihood that a threshold zc ¼ 0, 10, 50,
and 90% of stem oiling is exceeded at any location u (i.e. sampled
location or coastline grid node) on the basis of the oiling exposure
category surveyed at that location. In other words, the dependent
variable is the indicator defined in Eq. (1) while the covariate is the
oiling exposure category. Logistic regression was conducted in two
different settings.
First, an aspatial logistic regression was performed whereby the
predictive model is created using the entire data set (i.e. 729 lo-
cations where both hard data and oiling exposure category were
recorded) and without accounting for the geographical location of
the data. The underlying assumption is that the relationship be-
tween oiling exposure category and percentage of plant stem oiling
does not change along the Louisiana coastline (assumption of sta-
tionarity). This strong assumption was relaxed in a second analysis
where logistic regression was conducted at each location u using
only neighboring data (e.g. N closest data or data falling within a
window of size S centered on u). Because each observation in this
regression is weighted according to its proximity to the center of
the window, this type of local regression is known as
geographically-weighted regression (Fotheringham et al., 2002).
The implementation of geographically-weighted regression
(GWR) requires the selection of two main parameters: 1) spatial
weight function, and 2) search strategy for the local regression (i.e.
size of the search window S or number of neighbors N). The choice
of these parameters is described in Note S1 (Supplementary
material).
2.2.3. Updating prior probabilities of plant stem oiling using
indicator cokriging
The predictive models created by aspatial or geographically-
weighted regression do not make full use of all the information
available in that: 1) out of a total of 1096 stem oiling data only 729
data that also included oiling exposure category were used in the
regression; and 2) the geographical coordinates of the data are
ignored in aspatial regression and only indirectly used in GWR,
since only the separation distance from the center of the search
window is accounted for, thereby ignoring any spatial clustering of
the data which could bias results. Indicator cokriging (CK) in-
corporates this missing information into the prediction of the
probability of exceeding a stem oiling threshold zc using the
following estimator:
X
nðuÞ
p*CK ðu; zc Þ ¼ l0
p* ðu; zc Þ þ la
iðua ; zc Þ
a¼1
X
nðuÞ
with l0 þ la ¼ 1
a¼1
where the probability estimated using logistic regression, p*(u;zc),
sampled locations, i(ua; zc), using a set of weights l which are the
(2)
solution of a system of linear equations, known as a cokriging
system (Goovaerts and Journel, 1995). The weights take into ac-
count the proximity of the data to the location u (e.g. closest data
receive more weight), the data configuration (e.g. spatially clus-
tered data receive less weight since they provide redundant infor-
mation), as well as the spatial pattern of the data. Eq. (3) can also be
interpreted as the statistical updating of the probability derived
from the sole oiling exposure category (prior probability p*(u;zc))
using stem oiling data, resulting in a “posterior” probability of
exceeding the threshold zc (Deutsch and Journel, 1998). A simpler
version of Eq. (3) is the indicator kriging estimator, which uses only
the indicators i(ua; zc) in the estimation (i.e. the prior probability
p*(u;zc) receives a zero weight). This estimator was included in the
comparison study to assess the benefit of including soft informa-
tion in the prediction.
Similar to geographically-weighted regression, a key issue in
cokriging is the search strategy to select n(u) neighboring plant
stem oiling data. The number n(u) was set to a maximum of 8 to
avoid smoothing the results by averaging the influence of too many
observations. The search window was restricted to 1 km for the two
highest thresholds (50 and 90%) while it extended to 2 km for the
thresholds of 0 and 10% of stem oiling. We used a different search
window size for the different oiling categories because the physical
processes resulting in the highest percentages of stem oiling were
viewed as more local phenomena than those resulting in the lowest
stem oiling levels (0 and 10% plant stem oiling). When the search
window did not include any stem oiling data, no updating was
conducted and the regression estimate was used as final estimate
(i.e. prior and posterior probabilities are the same).
2.2.4. Validation analysis using Receiver Operating Characteristic
(ROC) curve
The accuracy of the predictive models created by the different
approaches and for different sets of parameters was assessed by
comparing at 729 sampled locations ua the estimated probability of
exceeding a threshold zc with the ground truth. The comparison
was based on ROC curves (Swets, 1988) which are described in note
S2 (Supplementary material). A quantitative measure of the accu-
racy of the prediction is the relative area under the ROC curve (AUC
statistic), which ranges from 0 (worst case) to 1 (best case).
2.2.5. Computing expected length of shoreline in each category of
plant stem oiling
The expected length of shoreline where the percentage of plant
stem oiling exceeds the threshold zc was computed from the pos-
terior probabilities p*CK ðu; zc Þ as:
X
K    
Lðzc Þ ¼ lj
p*CK uj ; zc
iProx uj ; zc (4)
j¼1
where:
 K ¼ 118,151 is the number of grid nodes discretizing the Loui-
siana coastline under mainland herbaceous marsh,
(3)  lj is the length of shoreline located within the 50
50 m square
centered on node uj, and
 iProx(uj;zc) ¼ 1 if the node uj is within 1 km (thresholds 50 and
90%) or 2 km (thresholds 0 and 10%) of a pre-assessment (PA)
site, and zero otherwise.
In other words, the expected length of shoreline with plant stem
oiling above zc was computed as the sum of the product of the
 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271 269

shoreline lengths within each 50
50 m square and the probability
that oiling is above zc within that square. Only locations in the vi-
cinity of a PA site were included in the computation to avoid
extrapolating results to sparsely sampled segments of shoreline.
The distances of 1 and 2 km correspond to the size of the cokriging
search windows.
3. Results and discussion
3.1. Data configuration and summary statistics
Fig. 1 shows where hard and soft data on percentage of plant
stem oiling were collected at pre-assessment (PA) sites. Summary
statistics indicate that stem oiling was observed at 39.6% of hard
data sites and only 6% of soft data sites (Table 1). Similarly, the NOO
exposure category is more prevalent at soft data sites (81.4%)
relative to hard data sites (47.9%); see Table 2. This table also
highlights the fact that secondary information on oil exposure was
not recorded at a number of sites: 20% (182) of the hard data sites
and 36.2% (67) of the soft data sites did not have oil exposure data.
In other words, the PA survey data extend beyond the boundaries of
the shoreline exposure dataset.
The oiling exposure category surveyed along the coastline,
which was discretized using a 50 m spacing grid, is mapped in
Fig. 2. The entire grid includes 118,151 nodes under mainland her-
baceous marsh, and their repartition among four categories of oil-
ing exposure is listed in the final column of Table 2. The percentage
of observations in the NOO category is much larger in the shoreline
exposure dataset compared to the hard dataset: 85% of the shore-
line exposure data are NOO, compared to 47.9% of the PA points,
which reflects the preferential sampling of oiled locations during
the PA survey.
3.2. Continuous versus categorical secondary data
Table S3 (supplementary material) summarizes the results for
the aspatial logistic regression between the oiling exposure cate-
gories (4 levels) and the indicator of exceedance of one of the four
thresholds zc for the percentage of plant stem oiling. Incorporating
the exposure categories as a categorical vs continuous variable has
a moderate impact on the estimated probabilities of exceedance,
yet the AUC statistic indicates that the predictive power of the two
types of model is identical. This result justifies the use of a
Table 1
Distribution of hard and soft data between the five classes of percentage of plant
stem oiling. Numbers in parenthesis are percentages of the total number of non-
missing data.
% Stem oiling Hard data (n ¼ 911) Soft data (n ¼ 185)
0% 551 (60.4%) 174 (94.0%)
0e10% 59 (6.5%)
10e50% 169 (18.6%) 11 (6.0%)
50e90% 80 (8.8%)
>90% 52 (5.7%)
continuous exposure variable in geographically-weighted regres-
sion (GWR), which allowed using smaller search windows.
3.3. Aspatial versus geographically-weighted regression
A sensitivity analysis (described in Note S3, supplementary
material) was conducted to support the choice of a search strat-
egy for GWR and the use of a probability threshold T for merging
aspatial and GWR estimates. The best predictive power, as
measured by the AUC statistic, was achieved when using the 50
closest observations and T ¼ 0.0075. The results listed in Table 3
(top 3 rows) demonstrate the greater accuracy of merged proba-
bilities relative to the separate application of aspatial regression or
GWR.
3.4. Cross-validation analysis
The stem oiling probabilities derived by merging aspatial and
GWR estimates are referred to as “prior” probabilities, as they are
based solely on the oiling exposure category recorded at each grid
node. Cokriging was applied to update these “prior” probabilities
using the additional information provided by indicator coding of
hard and soft data (Eqs. (1) and (2)). For each threshold, direct and
cross indicator variograms (Goovaerts, 1997; Goovaerts and Journel,
1995) were computed and modeled using a combination of one
exponential variogram model with range of 600 m, and another
exponential model with a range of 15 km for thresholds zc ¼ 0 and
10%, 10 km for threshold zc ¼ 50%, and 5 km for threshold zc ¼ 90%.
The benefit of this updating was assessed using a cross-validation
similar to the one conducted during the sensitivity analysis.
Table 3 indicates that indicator cokriging increases the accuracy of
the prediction compared to aspatial regression and indicator krig-
ing which uses only PA data. However, indicator cokriging fails to
improve over GWR results. Because of the geographical clustering
of vegetation oiling data and the fact that the soft data could not be
cross-validated, cokriging is still expected to benefit the prediction
and was retained as the interpolation technique of choice.
3.5. Euclidean distance versus over-water distance
For both geographically-weighted regression and cokriging, the
selection of neighbors and computation of spatial weight functions
were based on Euclidean distances. Due to the complex shape of
the Louisiana marsh edge, and the mechanisms by which oil was
transported over water during the spill, we might expect over-
water distance to be a better metric of spatial correlation than
Euclidian distance. However, translating from Euclidian to over
water distance for all of 118,151 nodes in the domain was prohibi-
tively expensive computationally.
To explore the sensitivity of the results to the use of Euclidian
distances, we instead performed a coordinate transformation for
the set of 729 locations where both the percentage of stem oiling
and oiling exposure categories were recorded, following the pro-
cedure described by Løland and Høst (2003). This approach, which

Table 2
Distribution of hard and soft data between the four categories of oiling exposure. The last column reports the number of shoreline nodes within each oiling exposure category.
Numbers in parenthesis are percentages of the total number of data points (excluding missing data).

Oiling exposure category Hard data (n ¼ 911) Soft data (n ¼ 185) Shoreline (n ¼ 118,151)

No Oil Observed (NOO) 349 (47.9%) 96 (81.4%) 100,418 (85.0%)

Light Oiling 172 (23.6%) 17 (14.4%) 12,234 (10.3%)
Heavier Oiling 153 (21.0%) 5 (4.2%) 4096 (3.5%)
Heavier persistent Oil 55 (7.5%) 0 (0.0%) 1403 (1.2%)
Missing 182 67
270 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271

Table 3 categories. In other words, the location-specific posterior proba-

Results of cross-validation analysis for different interpolation techniques and each
stem oiling threshold zc, as quantified using the Area Under the ROC Curve (AUC)
statistic. The best models are those with an AUC statistic closest to 1.
Interpolation technique Estimated probability of exceeding a stem oiling
threshold zc
zc ¼ 0% zc ¼ 10% zc ¼ 50% zc ¼ 90%
Aspatial regression 0.766 0.774 0.765 0.769
GWR 0.866 0.868 0.857 0.896
GWR & aspatial regression 0.876 0.877 0.872 0.913
Indicator kriging 0.754 0.753 0.614 0.584
Indicator cokriging 0.837 0.842 0.854 0.832
is based on a multidimensional scaling (MDS) of the 729
729
matrix of over-water distances between each pair of observations,
creates a new data configuration where Euclidean distances be-
tween observations approximate the original over-water distances
(Figure S2, supplementary material). A geographically-weighted
regression of data using the 50 closest neighbors before and after
projection by MDS yielded probabilities p*(ua;zc) that were
strongly correlated: r ¼ 0.977 to 0.989 depending on the threshold
zc. Based on this result and the similarity of variograms computed
on both datasets, the additional complexity associated with pro-
jecting the entire coastline into a new system of coordinates was
deemed not justified.
3.6. Geostatistical model for Louisiana shoreline
Following the cross-validation analysis, geographically
weighted regression followed by indicator cokriging was con-
ducted at 118,151 nodes discretizing the Louisiana coastline under
mainland herbaceous marsh. The posterior probabilities were
multiplied by the corresponding length of shoreline according to
Eq. (4) to compute the expected lengths of shoreline where the
percentage of plant stem oiling exceeds each of the four thresholds
zc ¼ 0, 10, 50 and 90%. The lengths of shoreline falling into each of
the four classes of stem oiling were then obtained as the differences
L(zcþ1)L(zc). The resulting lengths are listed in Table 4 (last col-
umn). The total length of oiled mainland herbaceous shoreline is
estimated to be 1161 km, with 29% (334 km) falling into the two
upper classes of stem oiling (i.e. > 50% stem oiling).
Scientifically it would have been interesting to propagate the
uncertainty attached to the predicted percentages of stem oiling
through the computation of lengths of shoreline, yielding a confi-
dence interval for each of the four lengths. However, the imple-
mentation of stochastic simulation (Goovaerts and Glass, 2014;
Goovaerts et al., 2008, 2009; van Donkelaar et al., 2015) instead
of (co)kriging would have provided only a measure of uncertainty
intrinsic to this particular geospatial approach. In the absence of
such uncertainty assessment, it is noteworthy to compare the
cokriging-based results to the lengths computed through a simple
proportional assignment of stem oiling data by shoreline exposure
Table 4
Expected length (km) of Louisiana shoreline under mainland herbaceous marsh
falling into different classes of percentage of plant stem oiling. Lengths were
computed using a simple apportionment of stem oiling data and indicator cokriging
results.
% Stem oiling Aspatial apportionment Indicator cokriging
bilities p*CK ðu; zc Þ in Eq. (4) are replaced by the exposure category-
specific probabilities listed in Table S3 (supplementary material).
Table 4 indicates that the aspatial approach leads to an additional
309 km of oiled shoreline compared to cokriging: 1470 vs 1161 km,
which confirms the risk of ignoring the geographical clustering of
the observations in particular for the apportionment of stem oiling
data that were recorded within the NOO category. Because the
probability of oiling is fairly small for the NOO category, the relative
difference between the two approaches is the largest for the 0e10%
stem oiling category: 305/199 ¼ 1.53. The second largest difference
is recorded for the 90e100% stem oiling category (156/109 ¼ 1.43)
which is the most critical for vegetation injuries. Regardless of the
category of stem oiling, incorporating the location of observations
in the analysis (cokriging approach) reduces the expected length of
oiled shoreline.
4. Conclusions
The characterization of contaminated sites following an envi-
ronmental crisis such as the Deepwater Horizon oil spill faces many
technical and scientific challenges (Peterson et al., 2012). In a crisis,
time is of the essence and there is little room for prolonged debate
or hesitation (Anastas et al., 2010). Such an emergency led here to
the collection of a wide variety of datasets describing shoreline and
vegetation oiling, with different spatial resolutions and degrees of
reliability.
In the natural resource damage assessment (NRDA) phase of the
Deepwater Horizon incident, various metrics of vegetation health
were found to be a function of the percent of plant stems oiled (e.g.,
Hester et al., 2015). Estimating the total injury to vegetation along
the Louisiana shoreline therefore required a way of synthesizing
the diverse shoreline oiling datasets to quantify the number of
shoreline miles within each stem oiling category. The main objec-
tive of the present analysis was to use these shoreline oiling data-
sets to provide accurate, reproducible and defensible estimates of
the spatial extent of stem oiling along the Louisiana shorelines.
The geospatial methodology described in this paper has been
tailored to the question and data at hand; it is the result of
exploratory data analysis, sensitivity analysis, cross-validation, and
scientific reviews conducted over the course of several years. Cross-
validation demonstrated the benefit of incorporating the
geographic location of the samples, as well as any secondary data
available, into the prediction of stem oiling along the Louisiana
shoreline. It also illustrated the need to rely on local calibration
techniques, such as geographically-weighted regression, for the
characterization of complex and heterogeneous coastal environ-
ments that span thousands of kilometers.
Ideally, data collection strategies for an NRDA would be tailored
to ensure that the metrics of contamination are perfectly aligned
with the quantification of injury to natural resources. However,
recognizing the inherent difficulties in establishing a perfectly
tailored data collection strategy in a time of crisis, the methodology
and findings described here can be extended to other sites where
inconsistent data quality, heterogeneous field conditions, and
complex geometries complicate the extrapolation of field obser-
vations into continuous measures of natural resource damages.
Acknowledgments

0e10% 305 199 We are grateful to two anonymous reviewers for their
10e50% 730 628 constructive comments on this manuscript, and to Andy Finley for
50e90% 279 225 comments on an earlier version of the paper. This study was con-
>90% 156 109 ducted within the Deepwater Horizon NRDA investigation and was
Total 1470 1161
funded by the State of Louisiana.
 P. Goovaerts et al. / Journal of Environmental Management 180 (2016) 264e271
Appendix A. Supplementary data
Supplementary data related to this article can be found at http://
dx.doi.org/10.1016/j.jenvman.2016.05.041.
References
Anastas, P.T., Sonich-Mullin, C., Fried, B., 2010. Designing science in a crisis: the
Deepwater horizon oil spill. Environ. Sci. Technol. 44, 9250e9251.
Baraba
s, N., Goovaerts, P., Adriaens, P., 2001. Geostatistical assessment and valida-
tion of uncertainty for three-dimensional dioxin data from sediments in an
estuarine river. Environ. Sci. Technol. 35 (16), 3294e3301.
Deepwater Horizon, N.R.D.A., 2010a. Deepwater Horizon/MC252/BP Shoreline/
Vegetation NRDA Pre-assessment Data Collection Plan. July 12 2010. Available.
https://pub-dwhdatadiver.orr.noaa.gov/dwh-ar-documents/900/DWH-
AR0013202.pdf.
Deepwater Horizon, N.R.D.A., 2010b. Deepwater Horizon/MC252/BP Shoreline
Vegetation Rapid Oiling Survey NRDA Pre-assessment Data Collection Plan.
December 15, 2010. Available. https://pub-dwhdatadiver.orr.noaa.gov/dwh-ar-
documents/900/DWH-AR0013221.pdf.
Deutsch, C.V., Journel, A.G., 1998. GSLIB: Geostatistical Software Library and User
Guide. Oxford University Press, New York.
Fotheringham, A.S., Brunsdon, C., Charlton, M., 2002. Geographically Weighted
Regression: the Analysis of Spatially Varying Relationships. John Wiley & Sons,
Chichester.
Gelfand, A.E., Kim, H.-J., Sirmans, C.F., Barnejee, S., 2003. Spatial modeling with
spatially varying coefficient processes. J. Am. Stat. Assoc. 98, 387e396.
Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation. Oxford Univer-
sity Press, New-York.
Goovaerts, P., Glass, G., 2014. Geostatistical modeling of the spatial distribution of
surface soil arsenic around a smelter. J. Jpn. Soc. Soil Phys. 128, 5e10.
Goovaerts, P., Journel, A.G., 1995. Integrating soil map information in modelling the
spatial variation of continuous soil properties. Eur. J. Soil Sci. 46 (3), 397e414.
Goovaerts, P., Trinh, H.T., Demond, A.H., Franzblau, A., Garabrant, D., Gillespie, B.,
Lepkowski, J., Adriaens, P., 2008. Geostatistical modeling of the spatial distri-
bution of soil dioxin in the vicinity of an incinerator. 1. Theory and application
to midland. Mich. Environ. Sci. Technol. 42 (10), 3648e3654.
Goovaerts, P., Schofield, J., Telech, J., 2009. Geostatistical estimation of contaminated
sediment volumes: review of common challenges and solution. In: Proceedings
of StatGIS 2009, Milos, Greece.
Goovaerts, P., Xiao, H., Adunlin, G., Ali, A., Tan, F., Gwede, C.K., Huang, Y., 2015.
Geographically-weighted regression analysis of percentage of late-stage pros-
tate cancer diagnosis in Florida. Appl. Geogr. 62, 191e200.
Hester, M.W., Willis, J.M., Rouhani, S., Steinhoff, M., Baker, M., 2015. Impacts of the
Deepwater Horizon Oil Spill on the Salt Marsh Vegetation of Louisiana. Avail-
able. https://pub-dwhdatadiver.orr.noaa.gov/dwh-ar-documents/901/DWH-
AR0270701.pdf.
View publication stats
271
Hu, K., Huang, Y., Li, H., Li, B., Chen, D., White, R.E., 2005. Spatial variability of
shallow groundwater level, electrical conductivity and nitrate concentration,
and risk assessment of nitrate contamination in North China Plain. Environ. Int.
31, 896e903.
Jacquez, G.M., Goovaerts, P., Kaufmann, A., Rommel, R., 2014. SpaceStat 4.0 User
Manual: Software for the Space-time Analysis of Dynamic Complex Systems,
4th, ed. BioMedware, Ann Arbor https://www.biomedware.com/files/SpaceStat_
4.0_Documentation.pdf.
Kitsiou, D., Karydis, M., 2011. Coastal marine eutrophication assessment: a review
on data analysis. Environ. Int. 37, 778e801.
Løland, A., Høst, G., 2003. Spatial covariance modelling in a complex coastal domain
by multidimensional scaling. Environmetrics 14 (3), 307e321.
Martins, T.G., Simpson, D., Lindgren, F., Rue, H., 2013. Bayesian computing with
INLA: new features. Comput. Stat. Data Anal. 67, 68e83.
Michel, J., Owens, E.H., Zengel, S., Graham, A., Nixon, Z., et al., 2013. Extent and
degree of shoreline oiling: Deepwater Horizon oil spill, Gulf of Mexico. PLoS One
8, e65087 doi:65010.61371/journal.pone.0065087.
Money, E.S., Carter, G.P., Serre, M.L., 2009. Modern space/time geostatistics using
river distances: data integration of turbidity and e. coli measurements to assess
fecal contamination along the Raritan river in New Jersey. Environ. Sci. Technol.
43 (10), 3736e3742.
Nixon, Z., 2015. Deepwater Horizon Shoreline Oil Exposure Mapping and Database.
Technical Report. Prepared for National Oceanic and Atmospheric Administra-
tion. Available. https://pub-dwhdatadiver.orr.noaa.gov/dwh-ar-documents/
901/DWH-AR0270684.pdf.
NOAA., 2013. Shoreline Assessment Manual, fourth ed. U.S. Dept. of Commerce,
Seattle, WA. Emergency Response Division, Office of Response and Restoration,
National Oceanic and Atmospheric Administration. 73 pp þ appendices.
Available. http://response.restoration.noaa. gov/sites/default/files/manual_
shore_assess_aug2013.pdf.
Peterson, C.H., Anderson, S.S., Cherr, G.N., Ambrose, R.F., Anghera, S., et al., 2012.
A tale of two spills: novel science and policy implications of an emerging new
oil spill model. Bioscience 62, 461e469. http://dx.doi.org/10.1525/
bio.2012.62.5.7.
Remy, N., Boucher, A., Wu, J., 2008. Applied Geostatistics with SGeMS: a User’s
Guide. Cambridge University Press, Cambridge.
SAS Institute Inc, 2011. SAS/STAT 9.3 User’s Guide. SAS Institute Inc, Cary, NC.
Silliman, B., He, Q., Dixon, P., Wobus, C., Willis, J., Hester, M., 2015. Accelerated
marsh loss following the BP Deepwater Horizon oil spill: a region wide survey.
DWH NRDA shoreline technical working group report. Prepared for the Loui-
siana Coastal Protection and Restoration Authority.
Swets, J.A., 1988. Measuring the accuracy of diagnostic systems. Science 240,
1285e1293.
van Donkelaar, A., Martin, R.V., Spurr, R.J.D., Burnett, R.T., 2015. High-resolution
satellite-derived PM2.5 from optimal estimation and geographically weighted
regression over north America. Environ. Sci. Technol. 49 (17), 10482e10491.
http://dx.doi.org/10.1021/acs.est.5b02076.