EP A/600/R-19/196
ERASC-018
Summary Report
December 2019
SUMMARY REPORT
SEPARATING ANTHROPOGENIC METALS CONTAMINATION FROM
BACKGROUND: A CRITICAL REVIEW OF GEOCHEMICAL EVALUATIONS AND
PROPOSAL OF ALTERNATIVE METHODOLOGY
Ecological Risk Assessment Support Center
Office of Research and Development
U.S. Environmental Protection Agency
Cincinnati, OH
-------
NOTICE
Mention of trade names or commercial products does not constitute endorsement or
recommendation for use.
Preferred Citation:
U.S. EPA (U.S. Environmental Protection Agency). 2019. Summary Report. Separating Anthropogenic
Metals Contamination from Background: A Critical Review of Geochemical Evaluations and Proposal of
Alternative Methodology. Center for Environmental Solutions and Emergency Response, Ecological Risk
Assessment Support Center, Cincinnati, OH. EPA/600/R-19/196
11
-------
TABLE OF CONTENTS
Page
LIST OF TABLES iv
LIST OF FIGURES v
LIST OF ABBREVIATIONS vi
AUTHORS, CONTRIBUTORS, AND REVIEWERS vii
PREFACE ix
EXECUTIVE SUMMARY x
1. EVALUATION OF GEOCHEMICAL ASSOCIATIONS AS A SCREENING TOOL FOR
IDENTIFYING ANTHROPOGENIC CHEMICAL CONTAMINATION 1
1.1 INTRODUCTION 1
1.2 STATISTICAL ANALYSES 2
1.3 RESULTS AND DISCUSSION 3
1.4 RATIONALE FOR ALTERNATIVE METHODOLOGY 5
2. APPLICATION OF DISCRIMINANT ANALYSIS WITH CLUSTER ANALYSIS TO
DETERMINE ANTHROPOGENIC CHEMICAL CONTAMINATION 6
2.1 INTRODUCTION 6
2.2 MATERIALS AND METHODS 7
2.3 RESULTS AND DISCUSSION 9
2.3.1 Cluster Analysis 9
2.3.2 Discriminant Analysis 9
2.3.3 Potential Applications 11
CONCLUSION 11
REFERENCES 12
in
-------
LIST OF TABLES
No. Title Page
1. Non-parametric Test Results Showing Differences in Selected Geochemical Ratios
Across Predominant U.S. Soil Orders for Selected Trace Metals 17
2. Summarized Total Chemical Concentrations Among Reference and Site-related Sample
Soil Populations 18
IV
-------
LIST OF FIGURES
No. Title Page
1. Log-linear Relationships Between Selected Trace Metal and Fe Concentrations Among
Predominant USD A Soil Orders 19
2. Box Plots Depicting Log Distributions of Geochemical Ratios Among Predominant
USD A Soil Orders 20
3. Site Soils Map 21
4. Geochemical Associations Between Total Soil Aluminum and Iron Concentrations
Among Reference and Site-related Sample Soil Populations Illustrating Incompatible
Matrices 22
5. Dendrogram Depicting the Hierarchical Clustering of Chemical Concentrations 23
6. Results from Discriminant Analysis of the Three Significant Chemical Signatures
Illustrating the Multivariate Separation of Chemical Concentrations 24
7. Results from Discriminant Analysis Illustrating, in Order of Relative Magnitude, the
Contaminants of Potential Concern for Site-related Soils 25
v
-------
LIST OF ABBREVIATIONS
A1 aluminum
CA cluster analysis
Can 1 first canonical variable
Can 2 second canonical variable
Cd cadmium
COPCs Contaminants of Potential Concern
Cr chromium
Cu copper
DA discriminant analysis
EPA U.S. Environmental Protection Agency
Fe iron
Mn manganese
NRCS Natural Resource Conservation Service
Pb lead
QDA quadratic discriminant analysis
USD A U.S. Department of Agriculture
V vanadium
Zn zinc
VI
-------
AUTHORS, CONTRIBUTORS, AND REVIEWERS
(Reviewers of this Summary Report are marked with an asterisk)
AUTHORS
Richard H. Anderson1
U.S. Environmental Protection Agency
Office of Research and Development
National Center for Environmental Assessment
Cincinnati, OH 45268
CONTRIBUTORS
David B. Farrar
U.S. Environmental Protection Agency
Office of Research and Development
National Center for Environmental Assessment
Cincinnati, OH 45268
Sharon R. Thorns
U.S. Environmental Protection Agency
Region 4 Superfund Division
Atlanta, GA 30303
REVIEWERS
Brad Venner
U.S. Environmental Protection Agency
National Enforcement Investigations Center
Denver, CO 80225
Evan Englund
U.S. Environmental Protection Agency
Environmental Response Team - West
Las Vegas, NV 89119
1 Permanent Address: Air Force Civil Engineer Center (AFCEC), Technical Support Branch, Lackland AFB, TX.
Vll
-------
AUTHORS, CONTRIBUTORS, AND REVIEWERS (continued)
Robert G. Garrett
Geological Survey of Canada
Natural Resources Canada
Ottawa, Ontario K1A 0E8
Lena Q. Ma
University of Florida
Gainesville, FL 32611
Fangjie J. Zhao
Rothamsted Research
Hertfordshire AL5 2JQ, UK
Michael Griffith*
U.S. Environmental Protection Agency
Office of Research and Development
National Center for Environmental Assessment
Cincinnati, OH 45268
Tim Frederick*
U.S. Environmental Protection Agency
Region 4 Superfund Division
Atlanta, GA 30303
ACKNOWLEDGMENTS
This research was supported in part by an appointment to the Research Participation
Program at the U.S. Environmental Protection Agency (EPA) National Center for Environmental
Assessment (NCEA) administered by the Oak Ridge Institute for Science and Education through
an interagency agreement between the U.S. Department of Energy and EPA. Special thanks to
the Fort McClellan Base Closure Team, J.E. Bentkowski (EPA Superfund Division, Region 4)
for his help with the geological descriptions, and James Bond (a senior GIS analyst with Shaw
Environmental) for his help with Site soils map.
Vlll
-------
PREFACE
Meaningful estimates of background contaminant levels are a critical component of site
assessments. A request was submitted by the Ecological Risk Assessment Forum (ERAF) to the
Office of Research and Development's Ecological Risk Assessment Support Center (ERASC)
relating to the issue of background soil trace metal and metalloid (hereafter referred to as
"chemical") demarcation at metals contaminated sites. Specifically, the request pertained to the
validity of an empirical methodology (geochemical association plots) that utilizes covariation
between chemical concentrations and concentrations of major soil elemental constituents (i.e.,
reference metals) to identify samples that deviate from "natural" variation.2 Consequently, a
comprehensive investigation of this methodology was conducted and assumes assessments are
conducted with chemical and reference metal data collected from reference sites (i.e.,
background data) and site-related locations. This document summarizes the results of this
investigation as described in the ERASC draft response and two peer-reviewed articles
(Anderson and Kravitz, 2010, and Anderson et al., 2009). Part 1 of the document tests
chemical/reference metal associations among uncontaminated soils of contrasting mineralogy
and chemical/physical composition to help determine the extent of compatible background data
sets. Chemical/reference metal associations are shown to vary significantly among background
data sets. Thus, geochemical association plots are a useful screening tool for environmental site
assessments, but ubiquitous application of generic background data sets could result in erroneous
conclusions. Additional methodologies are needed as objective lines of evidence to conclude that
a chemical occurs as site-related contamination. Part 2 of the document proposes a novel
application to environmental site assessments. This application uses a multivariate-analysis
methodology utilizing discriminant analysis with clustered chemical concentrations to determine,
in relative order of magnitude, contaminated chemicals.
2Trace metals and metalloids (i.e., chemicals) are minor constituents of many geologic materials. In general,
cationic metals can be found in a variety of silicate and aluminosilicate minerals such as olivines, amphiboles,
micas, and feldspars (Wilson et al., 2008). In contrast, high levels of metalloids can be found in sedimentary
deposits associated with pyritic sulfur-containing materials (Chen et al., 2002; Strawn et al., 2002). Metamorphic
rocks comprised of serpentine minerals are also associated with high background chemical levels (Burt et al., 2001;
Lee et al., 2001). Because of the variety of sources, regional background chemical levels can vary by up to
three orders of magnitude (Gustavsson et al., 2001; Wilson et al., 2008).
ix
-------
EXECUTIVE SUMMARY
Empirical associations among trace metals and a major (i.e., reference) soil elemental
constituent, such as iron (Fe), are used during environmental site assessments to screen for
contaminants of potential concern (COPCs). These "geochemical association plots" use
empirical log-log relationships to discern sites with naturally elevated chemical levels from sites
with anthropogenic contamination. Point of fact, log-log relationships have been consistently
observed between chemicals and reference metal concentrations and are often implicitly assumed
to be constant. This investigation tests that assumption by using a regional geochemistry data set
to evaluate background chemical/Fe log-log associations across soils with highly diverse
composition. The results indicate that although geochemical associations may be proportional,
they differ statistically across predominant U.S. Department of Agriculture (USD A) soil orders.
Also, intraorder variability in geochemical ratios generally ranged multiple orders of magnitude,
which suggests that the order level of the USDA soil taxonomic system is insufficient to
reasonably classify background chemical concentrations. Geochemical association plots are a
useful screening tool for environmental site assessments, but ubiquitous application of generic
background data sets could result in erroneous conclusions (Anderson and Kravitz, 2010).
Reference soils are used to define baseline chemical values during remedial
investigations and are selected based on professional judgment, usually predicated on factors
such as soil type, proximity to site, topoedaphic landscape position, and habitat. Often, however,
representative reference soils are difficult to delineate. One could argue that a preferred
methodology for determining COPCs should be independent of the assessor and capable of
meaningful interpretation despite potential bias from incompatible reference soils or when
reference soils cannot be collected at all (e.g., logistic or monetary constraints). Additional
methodologies are needed as objective lines of evidence for concluding a chemical occurs as
site-related contamination—especially for sites with spatially heterogeneous soil composition
where soil matrix composition can be highly variable.
A multivariate-analysis methodology utilizing discriminant analysis with clustered
chemical concentrations is proposed as a novel application to environmental site assessments that
determine, in relative order of magnitude, contaminated chemicals. Finite mixture models are
x
-------
presented as a means to assess latent chemical clusters with some basis in statistical inference.
The methodology is illustrated with a typical localized data set containing total chemical
concentrations, extracted from bulk soil collected from reference (i.e., background data) and
site-related locations, obtained from a former military installation in the southeast United States.
The illustration is particularly applicable because site-related soils inherently possessed higher
background chemical levels relative to reference soils, which biases conventional analyses.
However, two distinct chemical signatures were observed within site-related samples illustrating
the versatility of the proposed methodology. Using these results along with known information
regarding the history of contamination at the site, a qualitative and quantitative assessment of
contaminated chemicals was made. Results are intended for illustration purposes only and are
discussed within the context of environmental site assessment.
XI
-------
1. EVALUATION OF GEOCHEMICAL ASSOCIATIONS AS A SCREENING TOOL
FOR IDENTIFYING ANTHROPOGENIC CHEMICAL CONTAMINATION
1.1 INTRODUCTION
The focus of this investigation is the bivariate analysis between soil chemical
concentrations and concentrations of major soil elemental constituents (i.e., reference metals),
collectively referred to as a geochemical evaluation (NAVFAC, 2003). Geochemical evaluations
are currently used to help determine contaminants of potential concern (COPCs) at some
metals-contaminated sites (NAVFAC, 2003; U.S. EPA, 2007) and are predicated on the
covariation between chemicals and reference metals, which have been repeatedly documented
(Hamon et al., 2004; Myers and Thorbjornsen, 2004; Thorbjornsen and Myers, 2007a,b). In this
approach, chemicals that produce trends with anomalous observations are assumed to occur as
site-related contamination. All uncontaminated soil samples are implicitly assumed to follow
similar, if not consistent, trends (Hamon et al., 2004; Myers and Thorbjornsen, 2004;
Thorbjornsen and Myers, 2007a,b). Although the validity of this assumption has been
empirically supported (Hamon et al., 2004), it has not been formally tested.
Soil matrix composition may modify geochemical associations. Although formal
causality cannot be inferred from scatter plots alone, covariation between chemicals and
reference metals reflect soil mineral composition (Lopez et al., 2005) and soil chemical
sequestration due to chemical adsorption to soil solid phases (Sparks, 2003). However, total soil
chemical concentrations are ambiguous in that they provide no information regarding the solid
phase from which the analyte was extracted.
Variable geochemical associations could occur across soils with contrasting siliceous and
hydrous oxide clay minerals. For example, chemical reactivity (i.e., binding capacity) of
hydrous oxides is influenced by the degree of crystallinity, which is influenced by various
pedogenic (i.e., soil forming) factors (Gerth, 2005). Chemicals also have variable affinities for
different hydrous oxide minerals (Huelin et al., 2006; Sparks, 2003). Similarly, cationic trace
metal affinity for the exchange complex of negatively charged siliceous clay minerals increases
with surface area, which can be regulated by clay mineral structure and expandability rather than
total clay content (Meier and Kahr, 1999). Thus, anomalous observations among geochemical
1
-------
association plots may not necessarily indicate anthropogenic contamination. Further
investigation is required to determine applicable background data sets for geochemical
evaluations of sites based on log-log trace metal/reference metal associations.
Central to the issue of identifying contaminated samples is identifying appropriate
background chemical concentrations fundamental to geochemical evaluations. Co-contamination
of reference metals and spatial heterogeneity could impede accurate conclusions and result in
statistical uncertainty. Without representative background data or with a predominance of
suspect site-related data, geochemical evaluations could result in erroneous conclusions if, on
average, site-related data were collected at a scale impacted by contamination (Molinaroli et al.,
1999; Kazemi et al., 2008; Koptsik et al., 2003; Vanderlinden et al., 2006). Sufficient
representative background chemical data are fundamentally critical to geochemical evaluations.
Consequently, the objective of this investigation was to test chemical/reference metal
associations among uncontaminated soils of contrasting mineralogy and chemical/physical
composition to help determine the extent of compatible background data sets. A regional
geochemistry data set, obtained from the USDA Natural Resource Conservation Service (NRCS)
Cooperative Soil Survey Program, was used to test the implicit assumption that all
uncontaminated soil samples follow similar, if not consistent, trends (Hamon et al., 2004; Myers
and Thorbjornsen, 2004; Thorbjornsen and Myers, 2007a,b). In this context, the central
tendencies of the bivariate relationships are evaluated among dissimilar soils in terms of their
estimated Y-intercept and slope values. Log-linear relationships between selected chemicals and
Fe, a surrogate soil reference metal, were analyzed for selected chemicals.
1.2 STATISTICAL ANALYSES
Log-log trace metal/reference metal associations were evaluated across individual pedons
of the same soil series (i.e., the finest level of US taxonomic classification) and analyzed for
differences across nine predominant soil orders (i.e., the broadest level of US taxonomic
classification). Further evaluations on sub-order classifications were operationally difficult
because most soils were not classified to sub-order levels. Proportionality of geochemical log-
log associations among soil orders was evaluated by analysis of covariance (ANCOVA) with a
2
-------
model described in Anderson and Kravitz, 2010. See that publication for further detail on
statistical analyses.
1.3 RESULTS AND DISCUSSION
Soils derived from geologic materials with appreciable trace metal concentrations retain a
similar chemical composition throughout early genesis. However, as soils weather and undergo
pedogenesis (i.e., chemical/physical transformations), there is a redistribution of analytes,
organic matter accrual, and secondary (siliceous and hydrous oxide) clay mineral precipitation
resulting in soil chemical/physical properties with variable similarities to that of the original
parent material (Minasny et al., 2008). Organic matter and secondary clay minerals both affect
trace metal retention through ion exchange reactions and semi- to highly stable covalent binding
(Sparks 2003). Hence, a highly diverse population of reference soils comprised of contrasting
chemical/physical properties was essential to comprehensively evaluate geochemical trace
metal/reference metal log-log associations.
Geochemical trace metal/reference metal log-log associations are usually illustrated with
Fe, Al, or Mn as the representative soil reference metal. Log-linear relationships between Fe and
Al concentrations and Fe and Mn concentrations were plotted by Anderson and Kravitz, 2010
(see Figure 1 in that paper). Because significant associations were observed, trace
metal/reference associations were evaluated with Fe as the surrogate reference metal.
Log-log plots are used to compress metal distributions in order to obtain linear
relationships and assume log-normality for statistical inference. Log-log plots between selected
trace metals and total soil Fe concentrations are presented in Figure 1. Obvious trends are readily
apparent with the exception of Cd (and to a lesser extent Pb) where relationships are highly
variable. Significant interaction was observed for all metals except Cd (Table 2 in Anderson and
Kravitz, 2010). Interaction indicates a departure from parallelism suggesting disproportional
relationships among predominant soil orders. However, as discussed in Anderson and Kravitz,
2010, evidence of disproportionality is marginal at best and can be seen by evaluating trends in
Figure 1. Thus, further evaluation using a method robust to the influence of sample size was
necessary.
3
-------
Log trace metal/Fe ratios were calculated according to Eq. 3 in Anderson and Kravitz,
2010. Figure 2 shows box plots illustrating distributions of ratios among predominant soil orders.
Interestingly, ratios generally ranged multiple orders of magnitude within individual soil orders
except for V, which only ranged two orders of magnitude reflecting the lack of variability
observed in Figure 1. Numerous extreme data points were observed (i.e., shaded circles) and
were determined as data below and above the 5th and 95th percentiles, respectively. Distributions
of all ratios generally fell on the same scale—except for Cd, which was considerably lower
(Figure 2).
Geochemical ratios were evaluated at the order level of USD A soil taxonomy to
specifically test the broadest hierarchical level of soil classification. Similar ratios across
predominant soil orders would support ubiquitous application of generic background data sets for
site assessments based on geochemical evaluations. However, nonconstant ratios were observed
as evidenced by statistically significant differences among soil orders for all chemicals evaluated
(Table 1). These results suggest that soil composition affects the ratio of selected chemicals to
Fe. Thus, although geochemical associations may be proportional (Figure 1), the application of
generic background data sets could result in Type II or Type I statistical errors. Alternatively
stated, geochemical evaluations can result in the identification of anthropogenic contamination as
background (contaminated observations at or below the trend line) or identification of
background as anthropogenic contamination (background observations above the trend line).
The variability observed in trace metal/Fe ratios within soil orders suggests that the order
level of soil classification may be too broad of a category to determine applicable background
data sets for site assessments based on geochemical association. The USDA Soil Taxonomy
system sub-classifies soils on the basis of mineralogy only at the family (i.e., sub-order) level
and may provide a better hierarchical scale at which to evaluate trends in geochemical ratios
(NRCS, 2003). Hence, further study is required to explicitly determine the taxonomic level of
classification where constant trace metal/reference metal ratios occur.
Geochemical surveys have repeatedly shown that background chemical concentrations
strongly depend on geologic and pedogenic processes (Burt et al., 2003; Gustavsson et al., 2001;
Klassen, 1998; Miretzky et al., 2001; Wilson et al., 2008). Overall, this investigation suggests
that parent material composition and pedogenic factors may all influence geochemical ratios
given the diversity of soils evaluated (Burt et al., 2003; Wilson et al., 2008). Assuming consistent
4
-------
effects of geologic and pedogenic processes on geochemical ratios during environmental site
assessments is an over simplification and may result in erroneous identification of anthropogenic
contamination. Site-specific background chemical data should be utilized if ample reference
sites with similar (i.e., suborder) soil composition can be identified and sampled.
1.4 RATIONALE FOR ALTERNATIVE METHODOLOGY
Although geochemical evaluations provide a tool for assessing potential soil
contamination, this investigation has quantitatively illustrated that results can be biased if
site-related soils and/or reference soils are of dissimilar composition, although further study is
required to explicitly determine the appropriate taxonomic level of classification for geochemical
evaluations. Reference soils for background data sets are selected at the discretion of the
assessor and based on professional judgment usually predicated on factors such as site geology,
soil type/parent material, proximity to site, topoedaphic landscape position, and habitat. One
could argue that a preferred methodology should be independent of the assessor and capable of
meaningful interpretation despite potential bias from incompatible reference soils or when
reference soils cannot be collected at all (e.g., logistic or monetary constraints). Additional
methodologies are needed as objective lines of evidence for concluding a chemical occurs as
site-related contamination—especially for sites with spatially heterogeneous soils where soil
matrix composition can be highly variable.
Multivariate analysis techniques can simultaneously evaluate multiple chemicals and
allow the development of a site-specific chemical profile that has been referred to as a signature
(Ridgway et al., 2003). Multivariate chemical signatures have been used to identify
contaminated sites through techniques such as robust multivariate outlier detection (Filzmoser et
al., 2005), principal component analysis (Imrie et al., 2008; Korre, 1999a), and factor analysis
(Dragovic et al., 2008; Korre, 1999a; Reimann et al., 2002; Vaccaro et al., 2007). Formal
inference among these procedures assumes multivariate lognormality out of convenience because
it leads to chi-square distributed distance values for significance tests (Filzmoser et al., 2005).
Univariate and bivariate applications are perfectly acceptable exploratory analyses for
some situations, but multivariate methods are relatively more comprehensive and can reveal
features of environmental data that univariate and bivariate methods cannot. The methodology
5
-------
summarized here, from Anderson et al., 2009, is emphasized for sampling schemes that collect a
large percentage of contaminated samples making it difficult to separate outliers from population
mixtures. In this context, a method capable of separating samples into categories that represent
sites with relatively homogeneous multivariate signatures seems logical for reliable inference.
2 APPLICATION OF DISCRIMINANT ANALYSIS WITH CLUSTER ANALYSIS TO
DETERMINE ANTHROPOGENIC CHEMICAL CONTAMINATION
2.1 INTRODUCTION
Cluster analysis (CA) is an assortment of multivariate data analysis methods that aim to
discover groups of similar observations. A considerable variety of CA techniques have been
proposed, as reviewed in specialized texts (Everitt, 1993; Gordon, 1999) as well as general texts
on multivariate data analysis (Rencher, 2002; Seber, 1984; Venables and Ripley, 1998). Many
CA methods summarize information in the form of a matrix giving similarities (or distances)
relating pairs of observations, as determined using some appropriate index, and generate
classifications such that observations in a given class are more similar to one another (as
measured using the chosen index) than to observations in different classes. Such methodology
has been driven particularly by the needs of biological taxonomy and ecological community
analysis. However, applications to environmental site assessments have increased in recent years
(Dragovic et al., 2008; Martinez et al., 2007; Mico et al., 2006; Molinaroli et al., 1999).
In addition to similarity-based procedures, there are recent methods of model-based CA,
which are based on distributional models (Fraley and Raftery, 2002; Fruhwirth-Schnatter, 2006).
Basing CA on a probability model leads to likelihood-based statistical inferences, jointly for
classifying observations and estimating parameters of class-specific distributions. In practice,
model-based CA is usually based on the theory of finite mixtures (Fruhwirth-Schnatter, 2006;
Ter Braak et al., 2003; Yang and Chang, 2005). A model-based approach assumes that sites can
be grouped into classes that can be described with distinct multivariate models (Banfield and
Raftery, 1993). Results from model-based clustering may be directly useful in the context of
environmental site assessments possibly leading to discovery of heterogeneity within and/or
among reference and/or site-related sample soil populations that could cast doubt on standard
6
-------
statistical comparisons. Multivariate application of model-based CA to environmental site
assessment remains uninvestigated.
Discriminant analysis (DA) is commonly applied with samples grouped using CA, in
order to evaluate the variables that account for cluster differences. DA methods are based on
assumptions similar to those of a finite mixture of multivariate normal distributions. The
primary advantage of DA is that the classification of samples is used to derive a linear
combination of the original variables (i.e., the discriminant function) that can serve to
discriminate among clusters. Canonical DA allows the identification of a few, orthogonal
discriminant functions based on the number of classifications, which can be determined by CA
(see Lambrakis et al., 2004; Mariner et al., 1997; Petalas and Anagnostopoulos, 2006; Sielaff and
Einax, 2007). Despite an explosion of data-mining methodologies, basic DA methods continue
to perform a critical role in many disciplines, particularly for small data sets (Hastie et al., 2009).
The remainder of this paper illustrates the application of CA and DA to environmental
site assessments. The methods are illustrated with a typical localized data set containing total
soil chemical concentrations, collected from reference and site-related locations, obtained from a
small arms munitions firing range in the southeast United States. Specifically, objectives are to
(1) illustrate the application of CA to define multivariate chemical signatures and (2) illustrate
the use of DA to characterize differences in chemical signatures among clusters identified with
CA. Results are discussed within the context of environmental site assessment.
2.2 MATERIALS AND METHODS
The proposed methodology is illustrated with field data collected during the remedial
investigation of Ft. McClellan, a former military installation located in Alabama, USA. Ft.
McClellan was closed under the Base Realignment and Closure program in 1999. The Fort
consisted of the Main Post, Pelham Range, and Choccolocco Corridor covering more than
18,000 ha (Figure 3). Firing ranges located in the Choccolocco Corridor (hereafter referred to as
"Site") were used for military small arms training during World War II, the Korean War, and the
Vietnam War, and the potentially affected area was subject to this investigation. The range was
abandoned in 1974 and mostly restricted to small arms training.
7
-------
The major soil complexes at the Fort are described in Anderson et al., 2009. Most of the
reference samples associated with the existing data set came from the Main Post. Other reference
samples came from the Pelham Range (Figure 3).
Soil samples were collected and stratified by depth using a stainless steel hollow stem
hand auger in accordance with ASTM method D1452 (ASTM, 2009). Only samples collected
from 0-30 cm were used for statistical analyses. Site-related soil samples were analyzed for the
Target Analyte List (TAL) by U.S. EPA SW-846 analytical methods (U.S. EPA, 2018). All
chemicals, except mercury, were analyzed with Inductively Coupled Plasma-Atomic Emission
Spectroscopy (ICP-AES). Reference soil samples were also analyzed for TAL chemicals using
ICP-AES. Further information on soil collection and analytical measurements techniques are
described in Anderson et al., 2009.
Hierarchical cluster analysis (CA) was used to assess patterns of geochemical association.
Results of this approach are displayed by a tree-like structure called a dendrogram. A
classification with a specified number of clusters was obtained from the dendrogram. Model-
based CA was implemented using the R package mclust Version 3 (Fraley and Raftery 2002,
2006) with options set to defaults.
Multivariate DA was used to summarize between-cluster variability as a function of all
measured chemicals. Because results from model-based CA suggested three components, DA
was performed with the three clustered chemical groups (i.e., signatures) allowing estimation of
within-cluster covariances. Assuming equal covariance matrices for each cluster leads to the
method of linear discriminant analysis, while quadratic discriminant analysis (QDA) allows for
different covariance matrices. However, discriminant plots for QDA are currently a topic of
statistical research (Khattree and Naik, 2000; Pardoe et al., 2007). Discriminant plotting based
on linear DA gave highly interpretable results, despite results from model-based clustering
suggesting differences among covariance matrices. As a result, the structure of the standardized
linear discriminant function of interest (defined for each chemical as the linear correlation
coefficient between the canonical scores and actual values) was evaluated in order to determine,
in relative order of magnitude, the chemicals that best distinguished between inferred
background and contaminated signatures. Multivariate DA was conducted using PROC
DISCRIM in SAS Version 9.2 for Windows.
8
-------
2.3 RESULTS AND DISCUSSION
Summarized total chemical concentrations for both reference and Site-related sample soil
populations are presented in Table 2. All mean cationic metal levels are elevated in Site-related
soils relative to reference soils, while the opposite is true for the anionic species (arsenic,
chromium, and vanadium). This suggests dissimilarity in the chemical/physical matrices of the
two sample populations, which can be assessed by evaluating relationships between
concentrations of dominant clay mineral constituents such as aluminum (Al) and iron (Fe).
Figure 4 shows elevated Al concentrations in Site-related soils relative to Fe concentrations.
Significantly (p < 0.0001) different intercepts, as determined by Analysis of Covariance, verify
that differences in soil matrices are in part responsible for elevated cationic trace metal
concentrations in Site-related soils. This is because 1) Al is a dominant constituent of the clay
minerals involved in cationic metal sequestration (Sparks, 2003) and 2) neither Al nor Fe
contamination has ever occurred at the Fort. Consequently, conventional assessment
methodologies (e.g., univariate tests, tests of location, and bivariate associations) would likely
result in high false-positive error. However, assessment of latent chemical signatures can still be
assessed illustrating the versatility of CA.
2.3.1 Cluster Analysis
All data were subjected to hierarchical CA to assess the empirical pattern of multivariate
association among chemical concentrations. A dendrogram depicting clusters that minimize
within-cluster sum of squared error at all possible cluster configurations is shown in Figure 5.
The three cluster configuration accounted for over half the variability in chemical concentrations
(R-square=0.652). Subsequent cluster configurations accounted for increasingly less variability
(Figure 5).
2.3.2 Discriminant Analysis
Once divergent chemical signatures have been determined and have been associated with
a contamination event, the objective is to determine the chemical or chemicals responsible for
the divergent signatures. Obviously, simple statistics (mean, min, and max) for each chemical
9
-------
within each observed signature are relevant diagnostics. However, for purposes of identifying
multivariate signatures, DA provides quantitative and graphical results that are particularly suited
for evaluating the signatures that best account for cluster differences. While the mclust software
includes DA capabilities, for present, illustrative purposes we prefer the current relatively
standard approach based on use of specialized canonical DA software.
The discriminant plot of the standardized canonical coefficients is presented in Figure 6.
Notably apparent is the vertical separation of all reference and Site-related samples (signatures 1
and 2 vs. 3) characterized in the first canonical variable (Can 1). Thus, Can 1 quantifies the
degree to which reference and Site-related samples differ in chemical concentrations, which we
argue to be the result of differences in soil matrix composition (Figure 2). Can 1 accounted for
the vast majority (92.5%) of the overall variability in chemical concentrations. However, the
second canonical variable (Can 2) was still highly significant (p <0.0001) and resulted in roughly
equal separation of Site-related samples in the horizontal space. Thus, although minor compared
to the matrix effect, substantive differences in chemical concentrations were also evident
between Site-related signatures quantified by Can 2 (Figure 6).
Of primary importance to environmental assessors is the ability to determine
contaminants among a suite of measured chemicals, with the most important differences between
site-related and reference subsets. Typical assessments involve an independent evaluation of
each chemical using the previously discussed methodologies (NAVFAC, 2003). Such
procedures generate large quantities of output, which can be labor intensive to evaluate.
Multivariate procedures, in general, and our proposed CA with subsequent DA approach provide
a simultaneous assessment based on information from multiple chemicals. Using the structure of
the relevant canonical variable(s) along with known information regarding the history of
contamination events at a site, both a qualitative and quantitative assessment of contaminated
chemicals can be made. By evaluating the structure of Can 2, the relative order of elevated
chemicals (i.e., chemicals above background) can be determined and is illustrated in Figure 7.
Cu, Pb, and Zn were determined as the three most elevated chemicals, respectively, and are
consistent with contamination associated with small arms firing ranges. Combining site history
information with these results, one could conclude that Cu, Pb, and Zn are contaminated
chemicals, while the other chemicals are within background ranges.
10
-------
2.3.3 Potential Applications
A common deliverable from studies designed to investigate the nature and extent of
contamination are geographic maps of chemical concentrations. Spatial interpolation methods
such as kriging are commonplace in Geographic Information Systems and are used to estimate
spatially continuous chemical distributions within a discretely sampled area. Usually, however,
chemical distributions are mapped individually without regard to co-occurring chemicals
(Johnson and Ander, 2008). On the other hand, multivariate assessment of complex chemical
signatures can consolidate output necessary for decision making (Korre, 1999b). For example,
classification output from a CA has been used to map multivariate chemical signatures for
predetermined grid cells within a study site (Martinez et al., 2007). Alternatively, spatial
interpolation methods can be applied to discrete variables (Mancho et al., 2006), such as
categories of chemical concentrations for continuously distributed maps. Another objective of
environmental site assessments is the development of site-specific benchmarks of chemical
concentrations that delineate the upper limit of background values. Typically, background
benchmarks are estimated as the upper tolerance limit (90% coverage and 95% confidence) of a
reference population of soil samples (U.S. EPA, 2007). Reference soils are selected at the
discretion of the assessor and based on professional judgment usually predicated on factors such
as proximity to site, topoedaphic landscape position, and habitat. Inherent in their application is
the assumption that reference and site-related soils are of similar composition which is
contradicted in the current study (Figure 4). Thus, methodologies are needed that estimate
background benchmarks where incompatible reference soils are inadvertently collected (e.g., the
current study) or when reference soils can not be collected at all (e.g., monetary or logistic
constraints).
CONCLUSION
Chemical/reference metal associations among uncontaminated soils of contrasting
mineralogy and chemical/physical composition were tested to help determine the extent of
compatible background data sets. Chemical/reference metal associations were shown to vary
significantly among background data sets. Thus, geochemical association plots are a useful
11
-------
screening tool for environmental site assessments, but ubiquitous application of generic
background data sets could result in erroneous conclusions. Additional methodologies are needed
as objective lines of evidence to conclude that a chemical occurs as site-related contamination. A
multivariate-analysis methodology utilizing discriminant analysis with clustered chemical
concentrations is proposed as a novel application to environmental site assessments that
determine, in relative order of magnitude, contaminated chemicals. The methodology is
illustrated with a typical localized data set - from a former military installation in the southeast
United States - containing total chemical concentrations, extracted from bulk soil collected from
reference (i.e., background data) and site-related locations. Site-related soils inherently possessed
higher background chemical levels relative to reference soils, yet two distinct chemical
signatures were observed within site-related samples, illustrating the versatility of the proposed
methodology. Using these results, along with known information regarding the history of
contamination at the site, a qualitative and quantitative assessment of contaminated chemicals
was made. Copper, lead, and zinc were the contaminated chemicals, while the other chemicals
were within background ranges.
REFERENCES
Anderson, R.H., D.B. Farrar, and S.R. Thorns. 2009. Application of discriminant analysis with
clustered data to determine anthropogenic metals contamination. Sci. Total Environ. 408:50-56.
Anderson, R.H. and M.J. Kravitz. 2010. Evaluation of geochemical associations as a screening
tool for identifying anthropogenic trace metal contamination. Environ. Monit. Assess. 167:631-
641.
ASTM. 2009. Standard practice for soil exploration and sampling by auger borings.
Banfield, J.D. and A.E. Raftery. 1993. Model-based Gaussian and non-Gaussian clustering.
Biometrics. 49:803-821.
Burt, R., M. Fillmore, M.A. Wilson, E.R. Gross, R.W. Langridge, and D.A. Lammers. 2001.
Soil properties of selected pedons on ultramafic rocks in Klamath Mountains, Oregon. Comm.
Soil Sci. Plant Anal. 32(13):2145-2175.
Burt, R., M.A. Wilson, M.D. Mays, and C.W. Lee. 2003. Major and trace elements of selected
pedons in the USA. J. Environ. Qual. 32(6):2109-2121.
12
-------
Chen, M., L.Q. Ma, and W.G. Harris. 2002. Arsenic concentrations in Florida surface soils:
Influence of soil type and properties. Soil Sci. Soc. Am. J. 66:632-640.
Dragovic, S., N. Mihailovic, and B Gajic. 2008. Heavy metals in soils: Distribution,
relationship with soil characteristics and radionuclides and multivariate assessment of
contamination sources. Chemosphere. 72(3):491-495.
Everitt, B. 1993. Cluster Analysis. Edward Arnold, London.
Filzmoser, P., R.G. Garrett, and C. Reimann. 2005. Multivariate outlier detection in exploration
geochemistry. Comp. Geosci. 31:579-587.
Fraley, C. and A.E. Raftery. 2002. Model-based clustering, discriminant analysis, and density
estimation. J. Am. Stat. Assoc. 97:611-631. Available at
http://www.stat.washington.edu/rafterv/Research/PDF/fralev2002.pdf.
Fraley, C. and A.E. Raftery. 2006. MCLUST Version 3 for R: Normal Mixture Modeling and
Model-based Clustering, Technical Report No. 504. Department of Statistics, University of
Washington.
Fruhwirth-Schnatter, S. 2006. Finite mixture and Markov switching models. Springer, New
York.
Gerth, J. 2005. Effects of crystal modification on the binding of trace metals and arsenate by
goethite. J. Soil Sed. 5:30-36.
Gordon, A.D. 1999. Classification. Chapman and Hall/CRC, New York.
Gustavsson, N., B. Bolviken, D.B. Smith, and R.C. Severson. 2001. Geochemical Landscapes
of the Conterminous United States: New Map Presentations for 22 Elements. Department of the
Interior, USGS Report # 1648, Denver, CO. Available at http://pubs.usgs.gov/pp/pl648/.
Hamon, R.E., M.J. McLaughlin, R.J. Gilkes et al. 2004. Geochemical indices allow estimation
of heavy metal background concentrations in soils - art. No. GB1014. Global Biogeochem. Cyc.
18:B1014.
Hastie T, R. Tibshirani and J. Friedman. 2009. The Elements of Statistical Learning: Data
Mining, Inference, and Prediction, 2nd. Ed. Springer Series in Statistics. Springer, NY.
Huelin, S.R., H.P. Longerich, D.H.C. Wilton, and B.J. Fryer. 2006. The determination of trace
elements in Fe-Mn oxide coatings on pebbles using LA-ICP-MS. J. Geochem. Explor.
91(1-3): 110-124.
Imrie, C.E., A. Korre, G. Munoz-Melendez, I. Thornton, and S. Durucan. 2008. Application of
factorial kriging analysis to the FOREGS European topsoil geochemistry database. Sci. Total
Environ. 393:96-110.
13
-------
Johnson, C.C. and E.L. Ander. 2008. Urban geochemical mapping studies: How and why we do
them. Environ. Geochem. Health. 30(6):511-530.
Kazemi, H.V., S.H. Anderson, K.W. Goyne, and C.J. Gantzer. 2008. Spatial variability of
bromide and atrazine transport parameters for a Udipsamment. Geoderma. 144(3-4):545-556.
Khattree R. and D. Naik. 2000. Multivariate Data Reduction and Discrimination with SAS
Software. SAS Press and John Wiley Sons Inc, New York.
Klassen, R.A. 1998. Geological factors affecting the distribution of trace metals in glacial
sediments of central Newfoundland. Environ. Geol. 33(2-3): 154-169.
Koptsik, S., G. Koptsik, S. Livantsova, L. Eruslankina, T. Zhmelkova, and Z. Vologdina. 2003.
Heavy metals in soils near the nickel smelter: Chemistry, spatial variation, and impacts on plant
diversity. J. Environ. Monit. 5(3):441-450.
Korre, A. 1999a. Statistical and spatial assessment of soil heavy metal contamination in areas of
poorly recorded, complex sources of pollution Part 1: factor analysis for contamination
assessment. Stoch. Environ. Res. Risk Assess. 13:260-287.
Korre, A. 1999b. Statistical and spatial assessment of soil heavy metal contamination in areas
of poorly recorded, complex sources of pollution Part 2: canonical correlation analysis and GIS
for the assessment of contamination sources. Stoch. Environ. Res. Risk Assess. 13:288-316.
Lambrakis, N., A. Antonakos, and G. Panagopoulos. 2004. The use of multicomponent
statistical analysis in hydrogeological environmental research. Water Res. 38(7): 1862-1872.
Lee, B.D., R.C. Graham, T.E. Laurent, C. Amrhein, and R.M. Creasy. 2001. Spatial
distributions of soil chemical conditions in a serpentinitic wetland and surrounding landscape.
Soil Sci. Soc. Am. J. 65:1183-1196.
Lopez, J.M.G., B. Bauluz, C. Fernandez-Nieto, and A.Y. Oliete. 2005. Factors controlling the
trace-element distribution in fine-grained rocks: The Albian kaolinite-rich deposits of the Oliete
Basin (NE Spain). Chem. Geol. 214(1 —2): 1—19.
Mancho, A.M., D. Small, and S. Wiggins. 2006. A comparison of methods for interpolating
chaotic flows from discrete velocity data. Comp. Fluids. 35(4):416-428.
Mariner, P.E., F.J. Holzmer, R.E. Jackson, J.E. White, A.S. Wahl, and F.G. Wolf. 1997.
Fingerprinting arsenic contamination in the sediments of the Hylebos waterway, Commencement
Bay Superfund Site, Tacoma, Washington. Environ. Eng. Geosci. 3(3):359-368.
Martinez, J., J. Llamas, E. de Miguel, J. Rey, and M.C. Hidalgo. 2007. Determination of the
geochemical background in a metal mining site: Example of the mining district of Linares (South
Spain). J. Geochem. Explor. 94(1-3): 19-29.
14
-------
Meier, L.P. and G. Kahr. 1999. Determination of the cation exchange capacity (CEC) of clay
minerals using the complexes of copper (II) ion with triethylenetetramine and
tetraethylenepentamine. Clays Clay Min. 47(3):386-388.
Mico, C., L. Recatala, A. Peris, and J. Sanchez. 2006. Assessing heavy metal sources in
agricultural soils of an European Mediterranean area by multivariate analysis. Chemosphere.
65(5):863-872.
Minasny, B., A.B. McBratney, and S. Salvador-Blanes. 2008. Quantitative models for
pedogenesis - A review. Geoderma. 144(1-2): 140-157.
Miretzky, P., V. Conzonno, and A.F. Cirelli. 2001. Geochemical processes controlling silica
concentrations in groundwaters of the Salado River drainage basin, Argentina. J. Geochem.
Explor. 73(3): 155-166.
Molinaroli, E., M. Pistolato, G, Rampazzo, and S. Guerzoni. 1999. Geochemistry of natural and
anthropogenic fall-out (aerosol and precipitation) collected from the NW Mediterranean: Two
different multivariate statistical approaches. Appl. Geochem. 14(4):423-432.
Myers, J. and K. Thorbjornsen. 2004. Identifying metals contamination in soil: A geochemical
approach. Soil Sed. Contam. 13(1): 1-16.
NAVFAC (Naval Facilities Engineering Command). 2003. Guidance for environmental
background analysis, Vol. II: Sediment. In: NFESC User's Guide UG-2054-ENV. Washington,
DC. Prepared by Battelle Memorial Institute, Earth Tech, Inc., and New Fields, Inc. under
Contract No. N47408-01-D8207.
NRCS (National Resource Conservation Service). 2003. Keys to Soil Taxonomy, 9th ed. U.S.
Department of Agriculture, NRCS, Washington, DC.
Pardoe, I, X.R. Yin, and R.D. Cook. 2007. Graphical tools for quadratic discriminant analysis.
Technometrics. 49(2): 172-183.
Petalas, C. and K. Anagnostopoulos. 2006. Application of stepwise discriminant analysis for
the identification of salinity sources of groundwater. Water Res. Mgmt. 20(5):681-700.
Reimann, C., P. Filzmoser, and R.G. Garrett. 2002. Factor analysis applied to regional
geochemical data: Problems and possibilities. Appl. Geochem. 17(3): 185-206.
Rencher, A.C. 2002. Methods of Multivariate Analysis, 2nd ed. Wiley, New York.
Ridgway, J., N. Breward, W.J. Langston, R. Lister, J.G. Rees, and S.M. Rowlatt. 2003.
Distinguishing between natural and anthropogenic sources of metals entering the Irish Sea.
Appl. Geochem. 18(2):283-309.
Seber, G.A.F. 1984. Multivariate observations, 1st ed. Wiley Series in Probability and
Mathematical Statistics, V Barnett, R.A. Bradley, J.S. Hunter et al., Eds. Wiley, New York.
15
-------
Sielaff, K. and J.W. Einax. 2007. The application of multivariate statistical methods for the
evaluation of soil profiles. J. Soil Sed. 7(l):45-52.
Sparks, D.L. 2003. Environmental Soil Chemistry, 2nd ed. Academic Press, London, UK.
Strawn, D., H. Doner, M. Zavarin, and S. McHugo. 2002. Microscale investigation into the
geochemistry of arsenic, selenium, and iron in soil developed in pyritic shale materials.
Geoderma. 108(3-4):237-257.
Ter Braak, C.J.F., H. Hoijtink, W. Akkermans, and P.F.M. Verdonschot. 2003. Bayesian
model-based cluster analysis for predicting macrofaunal communities. Ecol. Model.
160(3):235-248.
Thorbjornsen, K. and J. Myers. 2007a. Identifying metals contamination in groundwater using
geochemical correlation evaluation. Environ. Foren. 8(1):25—35.
Thorbjornsen, K. and J. Myers. 2007b. Identification of metals contamination in firing-range
soil using geochemical correlation evaluation. Soil Sed. Contam. 16(4):337-349.
U.S. EPA (Environmental Protection Agency). 2007. Framework for metals risk assessment.
Office of the Science Advisor, Risk Assessment Forum, Washington, DC. EPA/120/R-07/001.
U.S. EPA. 2018. SW-846 Update VI On-Line. Chapter Three, Inorganic Analytes. Revision 6.
https://www.epa.eov/sites/production/files/2019-06/dociiments/chapter three lit
Vaccaro, S., E. Sobiecka, S. Contini, G. Locoro, G. Free, and B.M. Gawlik. 2007. The
application of positive matrix factorization in the analysis, characterisation and detection of
contaminated soils. Chemosphere. 69(7):1055-1063.
Vanderlinden, K., R. Ordonez, M.J. Polo, and J.V. Giraldez. 2006. Mapping residual pyrite
after a mine spill using non co-located spatiotemporal observations. J. Environ. Qual.
35(1):21—36.
Venables, W.N. and B.D. Ripley. 1998. Modern Applied Statistics with S-PLUS, 2nd ed.
Springer, New York.
Wilson, M.A., R. Burt, S.J. Indorante et al. 2008. Geochemistry in the modern soil survey
program. Environ. Monit. Assess. 139(1—3): 151—171.
Yang, S.Y. and W.L. Chang. 2005. Use of finite mixture distribution theory to determine the
criteria of cadmium concentrations in Taiwan farmland soils. Soil Sci. 170(l):55-62.
16
-------
Table 1. Non-parametric Test Results Showing Differences in Trace Metal/Fe Ratios"
Across Predominant USDA Soil Orders for Selected Trace Metals.
Statistic
Cd
Cr
Cu
Pb
V
Zn
Chi-Square
28.4
37.6
22.8
79.2
22.2
39.6
Degrees of Freedom
8
8
8
8
8
8
p-Value
0.0004
<0.0001
0.0036
<0.0001
0.005
<0.0001
*See Eq. 3. in Anderson and Kravitz, 2010
17
-------
Table 2. Summarized Total Chemical Concentrations (mg kg Among Reference" and
Site-relatedb Sample Soil Populations.
Contaminant
Reference Soils
Site-Related Soils
Mean
Std. Dev.
Min
Max
Mean
Std. Dev.
Min
Max
Antimony
1.02
1.33
0.04
3.57
6.65
4.40
4.30
50.3
Arsenic
6.59
7.75
1.10
48.5
4.97
2.62
0.57
13.8
Barium
60.2
54.0
11.2
288
136
197
28.9
2290
Beryllium
0.39
0.21
0.06
0.87
0.78
0.39
0.37
2.66
Cadmium
0.17
0.28
0.01
2.01
0.78
1.68
0.50
18.8
Chromium
18.4
20.5
1.99
134
15.4
7.62
3.80
51.2
Cobalt
7.10
11.1
0.39
70.9
10.3
20.4
1.14
214
Copper
6.32
4.30
0.25
23.5
60.0
83.9
4.86
389
Lead
19.4
14.5
2.89
82.8
410
824
8.44
4640
Mercury
0.04
0.05
0.01
0.32
0.08
0.11
0.03
1.18
Nickel
5.08
4.14
0.82
21.8
7.97
6.91
1.07
59.7
Selenium
0.24
0.14
0.13
1.28
1.03
0.62
0.53
3.79
Silver
0.18
0.34
0.01
1.87
1.18
0.10
1.05
1.69
Thallium
0.75
1.32
0.02
3.31
1.23
0.47
0.73
5.68
Vanadium
29.6
26.4
4.66
158
23.3
11.6
5.66
53.1
Zinc
20.4
26.7
2.46
209
182
1060
9.74
9540
aNonaffected by Site-related activity; n = 67.
bPossibly affected by Site-related activity; n = 139.
18
-------
1e+2 1e+3 1e+4 1e+5 1e+6
1e+2 1e+3 1e+4 1e+5 1e+6
Fe (mg kg"1)
• ALFISOL
O ANDISOL
~ ARIDISOL
1e+2 1e+3 1e+4 1e+5 1e+6
1e+2 1e+3 1e+4 1e+5 1e+6
Fe (mg kg"1)
~ MOLLISOL
¦ INCEPTISOL
A ENTISOL
1e+2 1e+3 1e+4 1e+5 1e+6
1e+2 1e+3 1e+4 1e+5 1e+6
Fe (mg kg"1)
A VERTISOL
O ULTISOL
~ SPODOSOL
Figure 1. Log-linear Relationships Between Selected Trace Metal and Fe Concentrations Among Predominant USDA Soil
Orders.
19
-------
-1 1 1 1 1 1 1 1 1
~~i 1 1 1 1 r~
~~i 1 r~
~~i 1 1 1 r~
Figure 2. Box Plots Depicting Log Distributions of Geochemical Ratios Among Predominant USDA Soil Orders. Boxes
represent the interquartile range (i.e., 25th-75th percentiles) and the center horizontal line represents the 50th percentile or
median value. Lower and upper horizontal lines represent threshold values (i.e., 5th and 95th percentiles, respectively)
beyond which constitute extreme observations represented by the shaded circles.
20
-------
-------
CHOCCOLOCCO
CORRIDOR
Legend
Anniston-Allen Soils
[88881 Clarksville-Fullerton Soils
Decatur-Cumberland Soils
iiiiiijiijl Rarden-Montevallo-Lehew
^ Stony Rough Land
I I Other Soil Type
0 2.5 5
1 1 J liles
Figure 3. Site Soils Map. Former Choccolocco Corridor Ranges, Fort McClellan, AL, USA.
21
-------
000 Reference
~ ~~Site
-1 1 r~
~1 i—
10000
-1 1 1—r~
1000
100000
Iron (mg/kg)
Figure 4. Geochemical Associations Between Total Soil Aluminum and Iron
Concentrations Among Reference and Site-related Sample Soil Populations
Illustrating Incompatible Matrices.
22
-------
1X5 0.9 0.8 0.7 0.6 0.5 Q4 0.3 0.2 0.1 0.0
Proportion of Variance Explained
Figure 5. Dendrogram Depicting the Hierarchical Clustering of Chemical
Concentrations.
23
-------
-4-3-2-10 1 2 3 4
Can 2
~ Signature 1 A Signature 2 O Signature 3
Figure 6. Results from Discriminant Analysis of the Three Significant Chemical
Signatures Illustrating the Multivariate Separation of Chemical
Concentrations. Signature 1 contains 51% of the Site-related soil samples,
Signature 2 contains the other 49% of the site-related soil samples, and Signature 3
contains 100% of the reference soil samples.
24
-------
0.81
0.7
0.6
S>
"§ 0.5
-fa
CO
* 0.4
o
E 0.3
(C
O
0.2
0.1
0.0
b u ^
a <& c a '
q a A ¦-> -fc ^
q v ^ o
a
o o o ^ o
*0" "Ss '*s "V "s '-S, ^ o ,St
o * o c ij O ^
X" . <& <• a a,
o q?
------- |