\documentclass[twocolumn,letterpaper,sort]{article}
\pagestyle{myheadings} \usepackage{natbib}
\usepackage{../mcfnsSep2012}
%\usepackage[OpenPeerReview]{McfnsDraftCopy}
\usepackage{amsmath,bm,url,amssymb} \usepackage{graphicx,epstopdf,textcomp} \usepackage{setspace,lineno}
\usepackage{color} \usepackage[breaklinks]{hyperref}
\usepackage{rotating} % begin {sidewaystable} instead of begin {table}
%\usepackage[breaklinks,pdfstartview= {FitH -32768},pdfborder= {0 0 0},bookmarksopen,bookmarksnumbered]{hyperref}
%\usepackage {ogonek} %\linenumbers %\pagewiselinenumbers
\widowpenalty=10000 \clubpenalty=10000
%______________________________________________________________________
% Define MCFNS variables:
% Moved to mcfns.sty like the \issueyear \publish {Feb.~00} %Still not sure if Issue Manuscript publication date
\setcounter {page} {115} \def\issueno {2}
\def\editors {{\href{mailto:editors@mcfns.com}{Editor:~Greg~C~Liknes}}}
\def\submit {Mar.~29,~2013} %Submission date can be different than the issue year \issueyear
\def\accept {Jul.~19,~2013} %The works should be Accepted & Published in the year of the Current_Issue \issueyear
\def\lasterrata {Sep.~9,~2013} %Last Errata date can be different than the Issue-Year \issueyear
\def\citename {Clough} %"Author"
\def\citeemail {bclough84@gmail.com} % Use later: {\href{mailto://\citeemail} {FirstName \citename}}
\def\citeetal {~and Green} % {} %or {} %for a single author; or
\author{
{\href{mailto:\citeemail}{Brian J~\citename}}, %Change only 1st name of 1st author
{\href{mailto:green@crssa.rutgers.edu} {Edwin~J~Green}}
}
\affiliation {
%---------------
\large\scshape {
} \\
%---------------
%
% the lines above are for multi-line multiple authors - braking \author causes problems
%
%---------------
\small \it{\href{http://crssa.rutgers.edu/}{Center for Remote Sensing and Spatial Analysis, Department of Ecology, Evolution \& Natural Resources}} \\
\small \it{\href{http://crssa.rutgers.edu/}{Rutgers University, New Brunswick, NJ, USA}}
}
\def\yourtitle
{{
Comparing spatial and non-spatial approaches for predicting forest soil organic carbon at unsampled locations
}} %need double {{for \\ e.g.: {{Title \\ Subtitle}}
\def\yourkwords
{
Forest soil carbon; Geostatistics; Regression; Co-kriging; New Jersey; Coastal Plain.
}
\def\yourabstract
{
Prediction of soil organic carbon (SOC) at unsampled locations is central to statistical modeling of regional SOC stocks. This is often accomplished by applying geostatistical techniques to plot inventory data. However, in many cases inventory data is sparsely sampled ($<$0.1 plots/km$^{2}$) relative to the region of interest, and it is unknown if geostatistics provides any advantage. Our objective was to test whether modeling spatial autocorrelation, in multivariate and univariate predictive models, improved estimates of SOC at prediction locations based on sparsely-sampled inventory data. We conducted our study using a dataset sampled across all forested land in the Coastal Plain physiographic province of New Jersey, USA. We considered five models for predicting SOC, two linear regression models (intercept only and multiple regression with predictor variables), ordinary kriging (a univariate spatial approach), and two multivariate spatial methods (regression kriging and co-kriging). We conducted a simulation study in which we compared the predictive performance (in terms of root mean squared error) of all five models. Our results suggest that our sparsely-sampled SOC data exhibits no spatial structure (Moranâ€™s \textit{I}=0.05, \textit{p}=0.39), though several of the covariates are spatially autocorrelated. Multiple linear regression had the best performance in the simulation study, while co-kriging performed the worst. Our results suggest that when inventory data is dispersed across the region of interest, modeling spatial autocorrelation does not provide significant advantage for predicting SOC at unsampled locations. However, it is unknown whether this autocorrelation does not exist at broad scales, or if sparse sampling strategies are unable to detect it. We conclude that in these situations, multiple regression provides a straightforward alternative to predicting SOC for mapping studies, but that more work on the spatial structure of soil carbon across multiple scales is needed.
} %----------------------------------------------------------------------
% \usepackage[english,russian,polish]{babel}
% THE REST SHOULD BE AUTHOMATIC Go To the first Section
\title {\Large\bf\uppercase\yourtitle}
\begin{document} \markright {\hfil {{{\href {mailto://\citeemail} {\citename} \citeetal}} ~(\issueyear)/\mcfnshead}}
\twocolumn[
\begin{@twocolumnfalse}
\hypersetup {pdftitle= {\mcfnshead},pdfauthor= {\citename~(\issueyear)},pdfsubject= {\yourtitle},pdfkeywords= {\yourkwords}}
\maketitle
\hrule\begin{abstract} \yourabstract\\\\ {\bf Keywords:} \yourkwords
%\\\\ {{\bf S\l {} owa kluczowe:} Polskie slowa kluczowe.}
\end{abstract} \hrule\vspace{.3truein}
\end{@twocolumnfalse}
]
\section{Introduction}
Globally, forests are thought to store approximately 861 petagrams of carbon (Pg
C), with about 44\% of this mass found in forest soils (Pan et al., 2011). The large
capacity of the forest soil pool to sequester carbon makes its management a viable
option for mitigating the effects of atmospheric carbon emissions (Goodale et al.,
2002; Lal, 2008). Naturally, there is considerable interest in the quantification
of forest soil organic carbon (SOC) pools for carbon monitoring projects and the
development of market-based carbon accounting schemes. There is a need for methodologies
that produce consistent results with a degree of accuracy acceptable to policymakers
(Chen et al., 2000a; Houghton, 2003; Shvidenko et al., 2010).
\textbf{ }Forest carbon stocks are typically measured using forest inventories, and
areal estimates are gained by ``scaling up'' these measurements across the region of
interest (Birdsey, 1992; Goodale et al., 2002). However, these data are often sparse
relative to the extent of the stock estimate. In the case of forests, soil carbon
sampling is often excluded from large inventory efforts due to the additional time
and cost needed to collect and process samples relative to aboveground forest measurements.
As a result, regional estimates of forest soil carbon storage are often highly uncertain,
leading to wide disparity among the literature. For example, estimates of carbon
stocks for European forest soils have ranged from 3 Pg C to as high as 79 Pg C; a
difference that constitutes approximately 9\% of the global forest soil carbon stock (Cannell
et al., 1992; Goodale et al., 2002; Liski et al., 2002; Jones et al., 2005).
Developing a regional soil carbon stock from inventory data involves prediction
of the response variable at many unsampled locations (i.e., all squares of a grid
covering the region of interest). Spatial autocorrelation, where nearby points are
on average more similar than points that are further apart, is a common property
in environmental datasets and, when present in inventory data, may be leveraged to
increase prediction accuracy (Simbahan et al., 2006). Spatial autocorrelation may
be summarized by computing the semivariance, a measure of spatial similarity, and
plotting these values for all pair-wise combinations of the sampling points as a
function of distance (Goovaerts, 1997). These plots, typically referred to as the
empirical semivariogram, may be fitted with ``theoretical'' semivariogram models, such
as Mat\'{e}rn class or spherical functions. Kriging methods, a widely used class of spatial
interpolators, incorporate such theoretical semivariogram models to weight predictions
at unsampled locations (Isaaks and Srivastava 1989). This feature, combined with
the fact that these methods may be extended to model spatial covariance between the
response and predictor variables, makes kriging a logical approach for the prediction
of soil carbon.
Geostatistical techniques have been successfully applied to predict soil organic
carbon at unsampled locations, based on plot inventory data, at a variety of spatial
scales. Several studies are available for agricultural fields, where very dense sampling
regimes ($>$400 plots/km$^{2}$) can be achieved, and clear patterns of spatial
variation are often elucidated (Chen et al., 2000a; Lark, 2000; Mueller and Pierce,
2003; Simbahan et al., 2006). In such situations, geostatistical models have been
shown to offer considerable improvement in prediction results when compared to non-spatial
regression models (Simbahan et al., 2006).
Fewer examples are available for regional soil carbon mapping, where reduced sampling
density may make spatial autocorrelation more difficult to detect. Still, several
studies have shown an increase in prediction accuracy when incorporating geostatistical
techniques. Liski and Westman (1997) used block kriging to interpolate measurements
of soil organic carbon taken as part of the national forest inventory (NFI) in Finland,
and detected significant spatial structure in these clustered, but densely sampled
($\sim$5 plots/km$^{2}$) data. More recently Mishra et al. (2010) compared the performance
of several geostatistical methods, including geographically weighted regression and
regression kriging, to multiple regression models for predicting SOC across a heterogeneous
region in the northern Midwestern United States. Their results suggest a significant
increase in prediction accuracy ($\sim$22\% relative improvement) when incorporating
spatial error into the model. Other examples where significant spatial
structure was detected and used to model SOC are available for grasslands in Ireland (McGrath
and Zhang 2003, Zhang et al., 2011) and agricultural landscapes in the karst region
of China (Liu et al., 2006; Zhang et al., 2012).
While the aforementioned studies modeled SOC across large spatial extents, most
took advantage of reasonably dense plot inventory data ($\geq$0.1 plots/km$^{2}$),
and in the case of Liski and Westman (1997) approximately 5 plots/km$^{2}$. The
exception is the study by Mishra et al. (2010), which utilized sparsely sampled data ($\sim$0.003 plots/km$^{2}$), but modeled SOC across a heterogeneous landscape with several
major cover types and a pronounced latitudinal gradient (from the upper Peninsula
of Michigan south to Kentucky, USA), both of which may exert strong controls on SOC
distribution. When the region of interest comprises a single cover class, as it would
in forestry applications, or does not span many degrees of latitude, it is less clear
that modeling spatial autocorrelation presents any advantage for predicting forest
SOC. In fact, a few studies provide evidence suggesting this is the case. Studies
in tropical forests that examined forest SOC across multiple scales, in tropical
dry forests in the West Indies (Gonzalez and Zak, 1994) and in the Brazilian Amazon (Cerri
et al., 2000; Bernoux et al., 2006), suggest that the spatial structure is limited
to fine scales only.
In this study, our primary objective was to assess whether incorporating spatial
autocorrelation into models for predicting forest soil organic carbon at unsampled
locations improved results for sparsely sampled ($<$0.1 plots/km$^{2}$) inventory
data. To meet our objective, we compared the performance of both univariate and multivariate
spatial models to similar linear regression models. We predicted that the spatial
models would perform the best when predicting forest SOC at independent validation
locations, despite the sparsity of our sampling locations relative to the region
of interest. To test this prediction, we developed a simple simulation experiment
to directly compare the predictive accuracy of all models considered in the experiment.
\section{Methods}
\subsection{Study region}
This study was conducted on the Coastal Plain physiographic province of New Jersey,
USA (Fig. 1). This region is largely forested, and the remaining landcover mainly
consists of residential and agricultural development. Three major upland forest communities
dominate the region: (1) pitch pine (\textit{Pinus rigida}) forest,
(2) oak (\textit{Quercus} spp.) forest, and (3) mixed
communities that span a gradient between these two classes (Hasse and Lathrop, 2010).
On the inner coastal plain, these communities mix with other hardwood species such
as American beech (\textit{Fagus grandifolia}) and hickory (\textit{Carya}) spp.
Forested wetlands are common along river courses or in low areas. Most of these are
hardwood swamps dominated by red maple (\textit{Acer rubrum}), sweetgum (\textit{Liquidambar
styraciflua}), and blackgum (\textit{Nyssa sylvatica}). However, forested
peat bogs with pure stands of Atlantic white cedar (\textit{Chamaecyparis thyoides})
are also present across the landscape. Soils in the region are largely typic Hapludults
and Quartzisappamments of marine or alluvial origin (Tedrow, 1986). Soils range from
very poorly to excessively drained, and are primarily sandy in texture. However,
clayey and mucky soils are frequent in wet areas. The total area of the study region
(i.e., all forested land in New Jersey's Coastal Plain) is approximately 4,522 km$^{2}$.
\begin{figure*}[hbt!]
\centering\includegraphics*[width=.65\textwidth,angle=0]{Fig1.eps}
%\centering\includegraphics*[width=.75\textwidth,trim = 33 50 33 33,angle=270]{Fig1.eps}
%\centering\includegraphics*[width=.48\textwidth, trim = 5 5 5 5, clip]{Fig1.eps}
\caption{Distribution of sampling locations and primary forest cover types
across the study region in New Jersey, USA.}\vspace{.1in}
\label{fig1}
\end{figure*}
\subsection{The datasets}
We considered two plot inventory datasets for this study. The primary dataset,
hereafter referred to as the ``small'' dataset consists of 62 plots, and possesses
measurements for forest soil organic carbon and all of the covariates used in the
model experiments. This corresponds to a sampling density of approximately 0.013
plots/km$^{2}$. The ``large'' dataset consists of 120 plots and contains measurements
of the model covariates only, and was used for the co-kriging analysis. The small
dataset is a subset of the large dataset, so those 62 plots are co-located and present
in each. The plots were sampled in a stratified random design across the landscape,
based on both forest community type and soil drainage class (Fig. 1).
At each sampling location, soil was collected from three depth intervals: 0-10 cm,
10-20 cm, and 20-30 cm. At each depth interval bulk density was sampled using the
core method (Blake and Hartge, 1986), and a second sample was taken for laboratory
analysis. Bulk density samples were dried for 24 hours at 105 ${}^\circ$C and passed
through a 2 mm sieve to remove the coarse fragments (i.e., gravel and litter material)
that are not a component of the soil organic matter pool. The analytical samples
were air dried for at least 48 hours, sieved to 2 mm, then ground into powder with
a mortar and pestle and homogenized.
Percent soil organic carbon was estimated by elemental (CHN) analysis on a subsample
of the air-dried analytical sample. A second subsample was used to measure percent
soil organic matter (SOM) by loss-on-ignition (LOI). These data were recorded for
all 120 plots, and used as a covariate in the multivariate models. SOM typically
has a significant relationship with SOC, and has been used as a predictor for soil
organic carbon in several studies (Konen et al., 2002; De Vos et al., 2005). Samples
were placed in a Lindberg muffle furnace (General Signal, Watertown, WI, USA) at
400 ${}^\circ$C for 24 hours. Both percent SOC and percent SOM were converted to stock
estimates using the following formula:
\begin{equation} \label{GrindEQ__1_}
S=P\times BD\times V\times g
\end{equation}
where S is the stock estimate (Mg$\cdot$ha$^{-1}$), P is a percent measurement of SOC
or SOM, BD is soil bulk density (g$\cdot$cm$^{-3}$), V is the volume of a 1 ha rectangle
with a depth of 30 cm, and g is a unit scaling constant.
\subsection{Model covariates}
In addition to the plot measured soil organic matter data, we utilized four covariates
extracted from remote sensing and GIS datasets: normalized difference vegetation
index (NDVI), band 2 of the ``tasseled cap'' transform (TC2), compound topographic
index (CTI), and elevation (ELEV). These variables represent a reasonable set of
potential predictors for soil organic carbon, and are similar to covariates incorporated
by several recent regional SOC mapping studies (McGrath and Zhang, 2003; Mishra et
al., 2010; Vasques et al., 2012; Zhang et al., 2012). NDVI and TC2, the ``greenness''
band of the tasseled cap transform, are both related to net photosynthetic output,
which has a theoretical relationship to inputs into the soil organic carbon pool (Chapin
et al., 2002). Terrain position can have a strong influence on soil organic carbon
storage, so we included two related variables: elevation and estimates of the compound
topographic index (CTI). CTI is a steady state wetness index designed to model soil
water content based on values of slope and flow direction extracted from a digital
elevation model (Moore et al., 1991). It has been shown to correlate with soil moisture
content, which may exert influence over soil organic carbon formation and storage (Barling
et al., 1994).
To extract the NDVI and TC2 measurements for our sampling locations, cloud-free Landsat
TM scenes (http://glovis.usgs.gov) were downloaded for a single date during the study,
July 14$^{th}$ 2011, and tiled into a mosaic of the study region. We used a Level
1 data product from the Landsat 5 thematic mapper instrument that had been previously
corrected for radiometric error and terrain variability, geo-referenced, and converted
to Universal Transverse Mercator (UTM) projection. The Erdas Imagine software package
(Leica Geosystems, Atlanta, GA, USA) and ArcGIS (ESRI, Redlands, CA) were used to
separately generate rasters for both variables with a 30-m x 30-m grid cell size for
all forested land within the study region, and to extract values of NDVI and TC2
for our sampling locations. Both the elevation and CTI data were derived
from a 10-m digital elevation model (DEM) provided by the Center of Remote Sensing
and Spatial Analysis (CRSSA), Rutgers University. Compound topographic index was
calculated for all cells in the DEM using ArcGIS.
\subsection{Modeling approaches}
Our objective in this study was to test whether explicitly modeling spatial autocorrelation
improved prediction accuracy for our sparsely sampled forest SOC data. To accomplish
this we considered five models that represent spatial and non-spatial approaches
for both univariate (SOC data only) and multivariate (incorporating the predictor
variables) cases (Tab. 1). This design allowed us to both examine the effect of modeling
spatial autocorrelation only, in the case of the univariate spatial model (OK), as
well as the influence of the spatial variance term for two different multivariate
approaches.
\begin{table}
\label{tab1}
\caption{Dimensions and spatial variance assumptions for the five predictive
models considered in this study.}\vspace{3pt}
\begin{tabular}{p{1.45in} c c} \hline\hline
\textbf{Model} & \textbf{Dimensions} & \multicolumn{1}{p{.6in}}{\centering \textbf{Spatial variance term}} \\ \hline
Intercept only \newline regression (IR) & univariate & no \\
Multiple linear \newline regression (MLR) & multivariate & no \\
Ordinary kriging (OK) & univariate & yes \\
Regression kriging (RK) & multivariate & yes \\
Co-kriging (COK) & multivariate & yes \\ \hline
\end{tabular}
\end{table}
Our non-spatial approach was linear regression models (MLR) of the general form:
\begin{equation} \label{GrindEQ_2}
Y=\ \alpha +\ {\beta }_j*X
\end{equation}
Where \textit{Y }is soil organic carbon, \textit{X} is an \textit{n }x \textit{p} matrix
of the predictor variables, \textit{$\alpha $} is the intercept, and ${\beta }_j$ is
a vector of the slope parameters associated with the \textit{j}th covariate. In the case of
the intercept only model (IR), $\alpha $ is the only parameter.
All three of the spatial models we incorporated are variations on the kriging algorithm,
where spatial prediction is accomplished as a function of the theoretical semivariogram;
a model fitted to a plot of semivariance values against distance for each pair-wise
combination of sampling locations in the dataset (i.e., the empirical variogram) (Goovaerts
1997). For the univariate model (OK), the linear estimator used to predict new values
of the response variable, for some set of locations \textit{u}, takes the form:
\begin{equation}
Z^*\left(u\right)=m\left( u\right) + \sum^{n(u)}_{\alpha =1}{{\lambda }_{\alpha }\left(u\right)[(Z
\left(u_{\alpha }\right)-m\left(u\right)]}
\label{GrindEQ_3}
\end{equation}
Where \textit{Z*(u)} is the predicted value of the response variable at
new locations, $Z\left(u_{\alpha }\right)$ are the known values of the response at
sampled locations, $m(u)$ is the mean response, and the ${\lambda }_{\alpha }$'s
are the ``kriging weights'' for each sampled location, that are determined by the
semivariogram model (Goovaerts, 1997; Simbahan et al., 2006)\textit{.} In the case
of ordinary kriging, note that the mean is taken to be a function of the locations \textit{u} so
that it is allowed to vary across the region (Isaaks and Srivastava, 1989).
In addition to the univariate OK model, we considered two different
approaches for incorporating covariates into spatial models. The first of these is
regression kriging (RK), which is very similar to OK in principle. The difference
is that the residuals of the response and predictor variables are interpolated, and
in this way co-varying spatial patterns are indirectly incorporated into the analysis (Odeh
et al., 1994; Hengl et al., 2004; Simbahan et al., 2006). For prediction at new locations,
the spatially predicted residuals must be added back on to the mean trend, resulting
in the following linear estimator for $Z^*\left(u\right)$:
\begin{equation}
Z^*\left(u\right)=\ \sum^p_{k=0}{{\beta }_kq_k\left(u\right)+\ \sum^{n(u)}_{\alpha
=1}{{\lambda }_{\alpha }e(u)}}
\label{GrindEQ_4}
\end{equation}
Where ${\beta }_k\ $are the regression parameters associated with the predictors $q_k$, \textit{p} is
the number of predictors, and $e\left(u\right)$are the residuals between the response
and covariables (Hengl et al., 2003, 2004). The rest of the terms in the model are
as defined above. We wish to note that the technique we outline here is one of several
closely related approaches that have all variously been termed ``regression kriging'',
``kriging with external drift'', and ``kriging with a trend'' (Goovaerts, 1997; Wackernagel,
1998; Chiles and Delfiner, 1999). We follow Hengl et al. (2004) in describing the
method outlined above, where fitting the non-spatial trend and the spatial interpolation
of the residuals are accomplished separately, as regression kriging.
The second multivariate method considered here is co-kriging (COK), which represents a
particularly flexible approach to modeling multivariate geostatistical data. Rather
than interpolating residuals between the response and predictor variables, COK
starts with the fitting of both direct and cross variograms for all variables in
the model, typically with a linear model of coregionalization (Gelfand et al., 2004).
This variogram system is employed to weight predictions of the response variable
at new locations, according to the following linear estimator:
\begin{eqnarray}
Z^*\left(u\right)&= & m\left( u\right) + \sum^{n_1(u)}_{{\alpha }_1=1}{\lambda }_{{\alpha }_1}\left(u
\right)\left[Z_1\left(u_{{\alpha }_1}\right)-m_1\left(u_{{\alpha }_1}\right)\right] \nonumber
\\
& +& \sum^{N_v}_{i=2}{\sum^{n_i(u)}_{{\alpha }_i=1}{{\lambda }_{{\alpha }_i}\left(u
\right)\left[Z_i\left(u_{{\alpha }_i}\right)-m_i\left(u_{{\alpha }_i}\right)\right]}}\quad\quad
\label{GrindEQ_5}
\end{eqnarray}
Where \textit{Z*(u)} is the predicted value of the response variable at
new locations, ${\lambda }_{{\alpha }_1}$is the weight assigned to the response variable \textit{Z$_{1}$ }and ${
\lambda }_{{\alpha }_i}$ represents the weights for the covariates \textit{Z$_{i}$$_{ }$}(Goovaerts,
1997; Simbahan et al., 2006)\textit{. }In this model, the expected values \textit{m$_{i}$ }are
subtracted from the data, indicating we consider the spatial association between
the response and predictor variables to be a multivariate stationary process.
Co-kriging is appropriate for situations in which a response variable that is expensive
to measure is sampled sparsely, while several ``cheap'' covariates have been sampled
in the same as well as additional locations. In our case, we have the ``large'' dataset
available, which contains 120 measurements of all of the model covariates. These
additional values are used to fit the direct and cross variograms during co-kriging,
along with the 62 observations which also contain measurements of the response variable.
This situation lends itself well to the co-kriging approach.
\subsection{Model comparison simulation}
To compare the performance of the five models for predicting soil organic carbon,
we devised a simulation that compared predicted vs. actual results for independent
validation data. We first randomly divided the ``small'' dataset into fitting and
validation datasets. We reserved 25\% of the data for validation (N=15) and reserved
the remaining 47 plots to fit the models. We split the data this way, rather than
using an even split, because initial runs of several kriging models resulted in undefined
covariance functions when N=31 for the model fitting data. A fitting set of 47 plots
translates to a density of approximately 0.01 plots/km$^{2}$ across the study region.
In the case of the COK model, the additional covariate observations in the
``large'' dataset were included in the model fitting, as the structure of the coregionalization
model permits this design. To increase normality, all variables were log-transformed
prior to fitting the models. Each model was used to predict SOC for the validation
dataset, and we computed root mean squared error (RMSE) to assess model performance.
Prior to computing RMSE, predicted values of log(SOC) were back-transformed into
their original units. To avoid biasing results by selecting a single, favorable fitting
dataset we ran this simulation for 10,000 iterations and tracked mean RMSE for the
entire study. This is especially relevant for the geostatistical models, as relatively
sparse datasets such as ours may possess spatial autocorrelation with some configurations
but not with others.
To initialize the OK and COK models, we supplied values for the sill, range, and
nugget parameters derived by fitting a Mat\'{e}rn class covariance function to the empirical
variograms for soil organic carbon in the full dataset. In the co-kriging model,
these values were used to initialize the parameters for all direct and cross variograms.
In the case of RK, we supplied initial parameter values from a theoretical
variogram fitted to the residuals of SOC and the model covariates. All model fitting
was accomplished with ordinary least squares. A Mat\'{e}rn covariance function was selected
because it is a particularly flexible model for spatial autocorrelation, and is a
popular choice in current geostatistical research (Stein, 1999; Finley et al., 2011).
The simulation was conducted using the R statistical computing environment. Variogram
fitting, OK, and RK were conducted using the geoR package (Ribeiro
and Diggle, 2001), and co-kriging was accomplished in the gstat package (Pebesma,
2004).
\section{Results}
\subsection{Exploratory analysis }
Table 2 presents the mean and standard deviation for all variables, as well as the
regression parameters for the MLR model and the correlation coefficients between
log(SOC) and each of the covariates for the full dataset (N=62). Soil organic matter
is highly correlated with SOC ($\rho $=0.708), while the remaining variables are
not notably correlated ($\rho $$<$0.2 for each). For the intercept only model, $\alpha $ =
3.59.
\begin{table}
\label{tab2}
\caption{Mean (\textit{$\mu $}), standard deviation (\textit{$\sigma $$^{2}$}),
and slope parameters (${\beta }_j$) and correlation coefficients ($\rho $) for the
five covariates and SOC.}\vspace{3pt}
\begin{tabular}{p{1in}rrrr} \hline\hline
& \textit{$\mu $} & \textit{$\sigma $$^{2}$} & \textit{$\beta $j} & \textit{$\rho $} \\ \hline
SOC (Mg$\cdot$ha$^{-1}$) & 65.93 & 65.67 & ** & ** \\
SOM (Mg$\cdot$ha$^{-1}$) & 113.17 & 153.66 & 0.678 & 0.708 \\
NDVI & 0.61 & 0.05 & 0.395 & 0.103 \\
TC2 & 29.54 & 9.13 & -0.507 & 0.046 \\
CTI & 9.99 & 2.48 & 0.332 & 0.106 \\
ELEV (m) & 26.35 & 12.25 & 0.157 & 0.098 \\ \hline
\end{tabular}
\end{table}
Examination of the spatial structure in the SOC dataset does not reveal any spatial
autocorrelation among the 62 sample sites (Moran's \textit{I}=-0.05, \textit{p}=0.39).
However, slight positive spatial autocorrelation was noted for the following covariates:
TC2 (Moran's \textit{I}=0.06, \textit{p}=0.04), CTI (Moran's \textit{I}=0.05, \textit{p}=0.09),
and ELEV (Moran's \textit{I}=0.036, \textit{p}$<$0.001). Both SOM and NDVI do not possess
significant spatial autocorrelation (\textit{p}$>$0.10). The empirical variograms, as well
as the fitted Mat\'{e}rn covariance functions (i.e., the theoretical variograms), agree
with these results (Fig. 2). TC2, CTI, and ELEV show an increase in semivariance
across distance, each with an asymptotic range $>$ 120,000 m. However, note that
there is considerable residual error between the empirical semivariance values and
the fitted covariance model. NDVI suggests an increase in semivariance, but the scale
of the y-axis for this plot indicates a minute change across the effective range.
Both SOC and SOM do not show spatial structure in either the empirical or theoretical
semivariograms.
\begin{figure*}[hbt!]\vspace{-.1in}
\centering\includegraphics*[width=.75\textwidth,angle=270]{Fig2.eps}\vspace{-.2in}
\caption{Empirical (open circles) and theoretical (solid lines) variograms
for the response (SOC) and the five covariates.}
\label{fig2}
\end{figure*}
\subsection{Model comparison experiment }
The results of the simulation experiment show that the MLR model
provided the most accurate predictions for the validation data, in terms of mean
RMSE over the 10,000 trials in the simulation (Tab. 3). OK was the
worst performing model, followed by COK. These results correspond with the
general lack of structure in the SOC data described above. RK had
a similar, though slightly inferior, performance relative to MLR.
This is not surprising, given that the RK estimator is simply an
extension of that of MLR. Comparing the two univariate methods also suggests a disadvantage
to modeling spatial autocorrelation, as the IR model reduced
error when compared to the OK model.
\begin{table}
\label{tab3}
\caption{Results of the simulation experiment. Note that this table presents
back-transformed values of forest SOC. Mean RMSE refers to the mean root mean squared
error over the 10,000 trials in the simulation experiment. RI refers to the relative
improvement in predictive performance of each model, when compared to the worst performing
method (Ordinary Kriging).}\vspace{3pt}
\begin{tabular}{p{1.75in}c c} \hline\hline
\textbf{Model} & \multicolumn{1}{p{.6in}}{\centering\textbf{Mean RMSE (Mg$\cdot$ha$^{-1}$)}} &
\multicolumn{1}{p{.5in}}{\centering\textbf{RI \newline \newline (\%)}} \\ \hline
Intercept only regression (IR) & 59.65 & 12.1 \\
Multiple regression (MLR) & 51.52 & 24.1 \\
Ordinary kriging (OK) & 67.9 & -- \\
Regression kriging (RK) & 53.43 & 21.3 \\
Co-kriging (CK) & 61.01 & 10.1 \\ \hline
\end{tabular}
\end{table}
\section{Discussion}
In contrast to studies where prediction of SOC is accomplished with relatively dense
plot inventories ($>$0.1 plots/km$^{2}$) (e.g., Liski and Westman, 1997; Lark, 2000;
McGrath and Zhang, 2003; Simbahan et al., 2006; Zhang et al,. 2012), we found that
modeling spatial autocorrelation did not improve prediction accuracy at unsampled
locations for our sparse inventory data. Both variogram analysis and Moran's \textit{I} statistics
suggest a lack of spatial autocorrelation in our soil carbon data. While spatial
structure was noted in some of the covariates, the lack of spatial structure in SOC
resulted in inferior performance of the spatial models relative to MLR.
However, note that the RMSE of all models is large relative to the mean of soil carbon
for the whole dataset (65.9 Mg$\cdot$ha$^{-1}$), suggesting that there is a high degree
of uncertainty in all five models.
Generally, these results highlight the difficulties of spatial prediction of forest
soil carbon. A number of studies have identified spatial structure at local scales
in a variety of forest types, with variogram range parameters from 4-500 meters (e.g., Robertson et al., 1993; Lister et al., 2000; Wang et al., 2002; Garten Jr.
et al., 2007; Worsham et al., 2010). It is not known, however, if these fine-scale
spatial dynamics are meaningful to predictions for regional datasets, where distances
between plots may range from one to hundreds of kilometers. Further, it remains unclear
whether spatial autocorrelation at broad scales is an important factor in understanding
regional forest carbon dynamics. Our results suggest otherwise, as do those of the
few other studies that have looked at this question (Liski and Westman, 1997; Cerri
et al., 2000; Bernoux et al., 2006).
In our data, determining whether regional forest SOC data truly exhibits no spatial
structure, or if this is the result of a detectability issue caused by low sampling
densities, remains unclear. Assuming spatial structure exists at the regional scale,
describing it may require a large number of observations relative to the region of
interest. While national forest inventories, such as the US Forest Service's Forest
Inventory and Analysis (FIA) program, may achieve the requisite densities for aboveground
measurements (Finley et al., 2007), collection of data on soil variables is often
only completed at a fraction of these plots. Further, the ability to detect spatial
autocorrelation is influenced by the sampling design (Fortin et al., 1989). Thus,
surveys may need to be specifically designed to detect broad-scale spatial structure
in forest soils.
Those studies which have detected regional spatial autocorrelation in the soil organic
pool have typically done so over heterogeneous landscapes, spanning multiple cover
classes (McGrath and Zhang, 2003; Mishra et al., 2010; Vasques et al., 2010; Zhang
et al., 2011). In these contexts, the spatial structure of soil carbon is influenced
by other spatially-explicit dynamics, such as patterns in land use and land cover,
which may make regional patterns easier to define (Vasques et al., 2012). Modeling
soil carbon over very large regions, such as the northern portion of the Midwestern
United States (Mishra et al., 2010), also incorporates the effect of latitudinal
climate gradients which are well known to influence soil organic carbon (Chapin et
al., 2002). In this way, Mishra et al. detected an advantage to spatial approaches
(geographically weighted regression and regression kriging) over multiple regression
despite a very low plot density ($<$0.001 plots/km$^{2}$). In forestry applications,
particularly across comparatively small regions such as the Coastal Plain of New
Jersey, there may be fewer influences on the regional spatial structure of soil organic
carbon. However, additional studies in different forest types will be necessary to
determine if this is in fact the case.
Incorporating covariates of soil carbon into predictive models is a typical strategy,
and one employed by almost all of the studies outlined here. In our case, four of
the five covariates we considered were not strongly correlated with forest SOC. These
patterns may be unique to our region in some ways. For instance, one would expect
a strong relationship between soil organic carbon and elevation. However, the Coastal
Plain of New Jersey is a fairly low-relief landscape, and fully capturing the covariance
between SOC and elevation in an inventory dataset may be particularly challenging.
Field
measured soil organic matter was the one covariate that was reasonably correlated
with SOC, but given that this variable was also sampled as part of our forest inventory
it has limited usefulness for predicting soil carbon at unsampled locations. For
regression models, it is generally necessary to have values for the covariates at
the prediction locations (i.e., for all cells of a sampling grid in mapping applications).
Methods based on fitting coregionalization models, such as co-kriging, are attractive
in that they do not share this prerequisite (Goovaerts, 1997; Banerjee et al., 2004;
Gelfand et al., 2004). However, in the absence of spatial structure in the response
variable, these methods will likely yield poor results, as was the case with our
data. An alternative strategy is to model the spatial dynamics of the covariates
themselves. For example, soil organic matter may be interpolated based on ancillary
variables in order to inform a sampling grid for soil carbon. However, this introduces
additional sources of uncertainty which may propagate through to the final estimate
of the response variable.
The results of our study have applications for forest SOC mapping projects, particularly
where new inventories are being established to accomplish these goals. This may be
especially relevant in developing countries, where international funding mechanisms
such as the United Nations' Reducing Emissions from Deforestation and Degradation
(REDD+) program has motivated increased interest in managing forests to offset carbon
emissions (Edwards et al., 2010). Newly established forest inventories will be important
for both gathering baseline data on forest carbon stocks in these regions, and for
verifying gains in carbon sequestration (Maniatis and Mollicone, 2010). Our results
suggest that when plot inventories are sparsely distributed ($<$0.1 plots/km$^{2}$),
there is no spatial autocorrelation present in forest SOC data, and as a result modeling
spatial structure does not result in increased prediction accuracy. In these cases,
multiple linear regression presents a straightforward alternative, providing a set
of reasonable covariates can be identified for all prediction locations.
Taken in context with the existing literature on the spatial dynamics of forest SOC,
our work highlights the need for more studies that explicitly model soil carbon across
a range of spatial scales. Without these data, it remains unknown whether regional
spatial autocorrelation does not exist or requires more dense sampling schemes to
detect. Further, our results are from but one forest type, and it is not clear that
the dynamics we describe are generalizable to other forest ecosystems. That all of
our models provide a fairly poor fit to our SOC data demonstrates just how challenging
characterizing uncertainty in regional soil carbon stocks can be. Advanced statistical
modeling techniques such as geostatistics present many appealing methods for the
prediction of forest soil carbon, but their utility is premised on a set of assumptions
that the available data may not meet. We advocate that these methods are considered
for the prediction of forest carbon, and for SOC mapping studies, but only when their
use is warranted by the data.
\section{Conclusions}
When predicting soil organic carbon at unsampled locations based on sparse inventory
datasets, it may be difficult to detect a significant degree of spatial autocorrelation.
This is especially true on homogeneous landscapes, or for studies that only consider
one cover type, as there may be spatial structure associated with the correlation
between SOC density and land cover type. In such cases, geostatistical models may
be inappropriate, and multiple linear regression offers an appealing and straightforward
alternative. Including covariates can increase the predictive performance of statistical
models. The best predictors will not only be closely correlated with soil organic
carbon, but will be available for the full extent of the study region. The results
of our study have implications for SOC mapping approaches using existing inventories,
where analytical efforts are constrained by data quality and availability, as well
as for new sampling efforts where resources are limited. Future work should look
to model spatial autocorrelation of soil carbon across multiple scales, to fully
characterize the relationship of well-described local spatial structure to broad-scale,
regional patterns.
\section*{Acknowledgements}
We thank the guest editors and the three anonymous reviewers for their helpful comments.
Additionally, we wish to acknowledge the funding we have received from the United
States Forest Service and Rutgers, The State University of New Jersey.
\begin{thebibliography} {99}
\bibitem[a(2012)]{a}
Banerjee, S., B. P. Carlin, and A. E. Gelfand. 2004. Hierarchical Modeling
and Analysis for Spatial Data. Chapman {\&} Hall, Boca Raton, FL. 472 p.
\bibitem[a(2012)]{a}
Barling, R., I. D. Moore, and R. B. Grayson. 1994. A quasi-dynamic wetness
index for characterizing the spatial distribution of zones of surface
saturation and soil water content. Water Resour.\ Res.\ 30:1029-1044.
\bibitem[a(2012)]{a}
Bernoux, M., D. Arrouays, C. E., P. Cerri, and C. C. Cerri. 2006. Regional
organic carbon storage maps of the western Brazilian Amazon based on prior
soil maps and geostatistical interpolation. Dev.\ Soil Sci.\ 31:497-506.
\bibitem[a(2012)]{a}
Birdsey, R. A. 1992. Carbon storage and accumulation in United States forest
ecosystems. USDA For.\ Serv.\ Gen.\ Tech.\ Rep.\ WO-59. 55 p.
\bibitem[a(2012)]{a}
Blake, G. R., and K. H. Hartge. 1986. Bulk Density. P.377-382 in Methods of
Soil Analysis Part 1. Physical and Mineralogical Methods, A. Klute (ed.).
American Society of Agronomy, Inc.; Soil Science Society of America Inc.,
Madison, WI, USA.
\bibitem[a(2012)]{a}
Cannell, M. G. R., R. C. Dewar, and J. H. M. Thornley. 1992. Carbon flux and
storage in European forest. P. 256-271 in Responses of Forest Ecosystems to
Environmental Changes, A. Teller, P. Mathy, and J. N. R. Jeffers (eds.).
Elsevier, London.
\bibitem[a(2012)]{a}
Cerri, C. C., M. Bernoux, D. Arrouays, B. J. Feigl, and M. C. Piccolo. 2000.
Carbon Stocks in Soils of the Brazilian Amazon. in Global climate change and
tropical ecosystems, R. Lal, M. Kimble, and B. A. Stewart (eds.). CRC Press,
Boca Raton, FL.
\bibitem[a(2012)]{a}
Chapin, F. S., P. A. Matson, and H. A. Mooney. 2002. Principles of
terrestrial ecosystem ecology. Springer Science {\&} Business Media Inc.,
New York, NY. 350 p.
\bibitem[a(2012)]{a}
Chen, F., D. E. Kissel, L. T. West, and W. Adkins. 2000a. Field-scale
mapping of surface soil organic carbon using remotely sensed imagery. Soil
Sci.\ Soc.\ Am.\ J. 64:746-753.
\bibitem[a(2012)]{a}
Chen, W., J. Chen, J. Liu, and J. Cihlar. 2000b. Approaches for reducing
uncertainties in regional forest carbon balance. Global Biogeochem. Cycles
14:827-838.
\bibitem[a(2012)]{a}
Chiles, J., and P. Delfiner. 1999. Geostatistics: Modeling Spatial
Uncertainty. Wiley, New York. 734 p.
\bibitem[a(2012)]{a}
De Vos, B., B. Vandecasteele, J. Deckers, and B. Muys. 2005. Capability of
loss-on-ignition as a predictor of total organic carbon in non-calcaerous
forest soils. Commun.\ Soil Sci.\ Plant Analysis 36:2899-2921.
\bibitem[a(2012)]{a}
Edwards, D. P., B. Fisher, and E. Boyd. 2010. Protecting degraded
rainforests: enhancement of forest carbon stocks under REDD+. Conserv.\
Letters 3:313-316.
\bibitem[a(2007)]{a}
Finley, A. O., S. Banerjee, and R. E. McRoberts. 2007. A Bayesian approach
to multi-source forest area estimation. Environ.\ Ecol.\ Stat. 15:241-258.
\bibitem[a(2011)]{a}
Finley, A. O., S. Banerjee, and D. W. Macfarlane. 2011. A hierarchical model
for quantifying forest variables over large heterogeneous landscapes with
uncertain forest areas. J. Am.\ Stat.\ Assoc. 106:31-48.
\bibitem[a(2012)]{a}
Fortin, M.J., P. Drapeau, and P. Legendre. 1989. Spatial autocorrelation and
sampling design in plant ecology. Vegetatio. 83:209-222.
\bibitem[a(2012)]{a}
Garten J., C. T., S. Kang, D. J. Brice, C. W. Schadt, and J. Zhou. 2007.
Variability in soil properties at different spatial scales (1m-1km) in a
deciduous forest ecosystem. Soil Biol.\ Biochem. 39:2621-2627.
\bibitem[a(2012)]{a}
Gelfand, A. E., A. M. Schmidt, S. Banerjee, and C. F. Sirmans. 2004.
Nonstationary Multivariate Process Modeling Through Spatially Varying
Coregionalization. Test. 13:263-312.
\bibitem[a(2012)]{a}
Gonzalez, O. J., and D. R. Zak. 1994. Geostatistical analysis of soil
properties in a secondary tropical dry forest, St. Lucia, West Indies. Plant
and Soil. 163:45-54.
\bibitem[a(2012)]{a}
Goodale, C. L., M. J. Apps, R. A. Birdsey, C. B. Field, L. S. Heath, R. A.
Houghton, J. C. Jenkins, et al. 2002. Forest Carbon Sinks in
the Northern Hemisphere. Ecol.\ App.\ 12:891-899.
\bibitem[a(2012)]{a}
Goovaerts, P. 1997. Geostatistics for Natural Resources Evaluation. Oxford
University Press, New York. 496 p.
\bibitem[a(2012)]{a}
Hasse, J. E., and R. G. Lathrop. 2010. Changing Landscapes in the Garden
State: Urban Growth and Open Space Loss in NJ 1986 thru 2007. Center for
Remote Sensing and Spatial Analysis, Rutgers University, New Brunswick, NJ,
USA.
\bibitem[a(2012)]{a}
Hengl, T., G. Heuvelink, and A. Stein. 2003. Comparison of kriging with
external drift and regression kriging. Technical Report, International
Institute for Geo-Information Science and Earth Observation, Enschede. 18 p.
\bibitem[a(2012)]{a}
Hengl, T., G. Heuvelink, and A. Stein. 2004. A generic framework for spatial
prediction of soil variables based on regression-kriging. Geoderma
120:75-93.
\bibitem[a(2012)]{a}
Houghton, R. a. 2003. Why are estimates of the terrestrial carbon balance so
different? Glob.\ Change Biol. 9:500-509.
\bibitem[a(2012)]{a}
Isaaks, E. H., and M. Srivastava. 1989. An Introduction to Applied
Geostatistics. Oxford University Press, Oxford. 592 p.
\bibitem[a(2012)]{a}
Jones, R. J. A., R. Hiederer, E. Rusco, P. J. Loveland, and L. Montanarella.
2005. Estimating soil organic carbon in the soils of Europe for policy
support. Eur.\ J. Soil Sci. 56:655-671.
\bibitem[a(2012)]{a}
Konen, M. E., P. M. Jacobs, C. L. Burras, B. J. Talaga, and J. A. Mason.
2002. Equations for predicting soil organic carbon using loss-on-ignition
for North Central U.S. soils. Soil Sci.\ Am.\ J. 66:1878-1881.
\bibitem[a(2012)]{a}
Lal, R. 2008. Carbon sequestration. Philos. T. Roy. B. 363:815-30.
\bibitem[a(2012)]{a}
Lark, R. M. 2000. Regression analysis with spatially autocorrelated error:
simulation studies and application to mapping of soil organic matter. J.
Geogr.\ Info.\ Sci.\ 14:247-264.
\bibitem[a(1997)]{a}
Liski, J., and C. J. Westman. 1997. Carbon storage in forest soil of
Finland. Biogeochemistry 36:261-274.
\bibitem[a(2002)]{a}
Liski, J., D. Perruchoud, and T. Karjalainen. 2002. Increasing carbon stocks
in the forest soils of western Europe. Forest Ecol.\ Manag.\ 169:159-175.
\bibitem[a(2012)]{a}
Lister, A. J., P. P. Mou, R. H. Jones, and R. J. Mitchell. 2000. Spatial
patterns of soil and vegetation in a 40-year-old slash pine (Pinus
elliottii) forest in the Coastal Plain of South Carolina, U.S.A. Can.\ J.
For.\ Res.\ 30:145-155.
\addtolength{\textheight}{-2.4truein}
\bibitem[a(2012)]{a}
Liu, D., Z. Wang, B. Zhang, K. Song, X. Li, J. Li, F. Li, and H. Duan. 2006.
Spatial distribution of soil organic carbon and analysis of related factors
in croplands of the black soil region, Northeast China. Agric., Ecosyst.\
Environ.\ 113:73-81.
\bibitem[a(2012)]{a}
Maniatis, D., and D. Mollicone. 2010. Options for sampling and
stratification for national forest inventories to implement REDD+ under the
UNFCC. Carbon Balance Manage. 5:9-16.
\bibitem[a(2012)]{a}
McGrath, D., and C. Zhang. 2003. Spatial distribution of soil organic carbon
concentrations in grassland of Ireland. Appl. Geochem. 18:1629-1639.
\bibitem[a(2012)]{a}
Mishra, U., R. Lal, D. Liu, and M. Van Meirvenne. 2010. Predicting the
Spatial Variation of the Soil Organic Carbon Pool at a Regional Scale. Soil
Sci.\ Am.\ J. 74:906.
\bibitem[a(2012)]{a}
Moore, I. D., R. B. Grayson, and A. R. Ladson. 1991. Digital terrain
modelling: a review of hydrological, geomorphological, and biological
applications. Hydrol.\ Process. 5:3-30.
\bibitem[a(2012)]{a}
Mueller, T. G., and F. J. Pierce. 2003. Soil Carbon Maps: Enhancing Spatial
Estimates with Simple Terrain Attributes at Multiple Scales. Soil Sci.\ Soc.\
Am.\ J. 67:258--267.
\bibitem[a(2012)]{a}
Odeh, I., A. McBratney, and D. Chittleborough. 1994. Further results on
prediction of soil properties from terrain attributes: heterotropic
cokriging and regression-kriging. Geoderma. 63:197-214.
\bibitem[a(2012)]{a}
Pan, Y., R. A. Birdsey, J. Fang, R. A. Houghton, P. E. Kauppi, W. A. Kurz,
O. L. Phillips, et al. 2011. A large and persistent carbon sink in the world's forests.
Science 333:988-993.
\bibitem[a(2012)]{a}
Pebesma, E. J. 2004. Multivariable geostatistics in S: the gstat package.
Comput.\ Geosci. 30:683-691.
\bibitem[a(2012)]{a}
Ribeiro, P. J., and P. J. Diggle. 2001. geoR: a package for geostatistical
analysis. R-NEWS. 1:15-18.
\bibitem[a(2012)]{a}
Robertson, G. P., J. R. Crum, and B. G. Ellis. 1993. The spatial variability
of soil resources following long-term disturbance. Oecologia 96:451-456.
\bibitem[a(2012)]{a}
Shvidenko, A., D. Schepaschenko, I. McCallum, and S. Nilsson. 2010. Can the
uncertainty of full carbon accounting of forest ecosystems be made
acceptable to policymakers? Clim.\ Change 103:137-157.
\bibitem[a(2012)]{a}
Simbahan, G. C., A. Dobermann, P. Goovaerts, J. Ping, and M. L. Haddix.
2006. Fine-resolution mapping of soil organic carbon based on multivariate
secondary data. Geoderma. 132:471-489.
\bibitem[a(2012)]{a}
Stein, M. 1999. Spatial Interpolation, Some Theory for Kriging.
Spring-Verlag, New York. 249 p.
\vspace{3pt}
\bibitem[a(2012)]{a}
Tedrow, J. C. F. 1986. The Soils of New Jersey. New Jersey Agricultural
Experiment Station publication no. A-15134-1-82. 456 p.
\vspace{3pt}
\bibitem[a(2010)]{a}
Vasques, G. M., S. Grunwald, J. O. Sickman, and N. B. Comerford. 2010.
Upscaling of Dynamic Soil Organic Carbon Pools in a North-Central Florida
Watershed. Soil Sci.\ Am.\ J. 74:870.
\vspace{3pt}
\bibitem[a(2012)]{a}
Vasques, G. M., S. Grunwald, and D. B. Myers. 2012. Associations between
soil carbon and ecological landscape variables at escalating spatial scales
in Florida, USA. Landscape Ecol. 27:355-367.
\vspace{3pt}
\bibitem[a(2012)]{a}
Wackernagel, H. 1998. Multivariate Geostatistics: An Introduction With
Applications, 2nd Ed. Springer, Berlin. 403 p.
\vspace{3pt}
\bibitem[a(2012)]{a}
Wang, H., C. S. Hall, J. D. Cornell, and M. P. H. Hall. 2002. Spatial
dependence and the relationship of soil organic carbon and soil moisture in
the Luquillo Experimental Forest, Puerto Rico. Landscape Ecol. 17:671-684.
\vspace{3pt}
\bibitem[a(2012)]{a}
Worsham, L., D. Markewitz, and N. Nibbelink. 2010. Incorporating spatial
dependence into estimates of soil carbon contents under different land
covers. Soil Sci.\ Am.\ J. 74:635-646.
\vspace{3pt}
\bibitem[a(2012)]{a}
Zhang, C., Y. Tang, X. Xu, and G. Kiely. 2011. Towards spatial geochemical
modelling: Use of geographically weighted regression for mapping soil
organic carbon contents in Ireland. Appl.\ Geochem. 26:1239-1248.
\vspace{3pt}
\bibitem[a(2012)]{a}
Zhang, W., K. Wang, H. Chen, X. He, and J. Zhang. 2012. Ancillary
information improves kriging on soil organic carbon data for a typical karst
peak cluster depression landscape. J. Sci.\ Food Agric. 92:1094-1102.
\label{docend}
\end{thebibliography}
\end{document}