International Journal of Coal Geology 48 (2001) 1 – 22

www.elsevier.com/locate/ijcoalgeo

A geostatistical approach to predicting sulfur content

in the Pittsburgh coal bed
William D. Watson a,*, Leslie F. Ruppert b, Linda J. Bragg b, Susan J. Tewalt b
a
US Department of Energy, Energy Information Administration, 950 L’Enfant Plaza South Building EI-52, 1000 Independence Avenue SW,
Washington, DC 20585, USA
b
US Geological Survey, National Center, MS 956, Reston, VA 20192, USA
Received 16 October 2000; accepted 25 June 2001

Abstract

The US Geological Survey (USGS) is completing a national assessment of coal resources in the five top coal-producing
regions in the US. Point-located data provide measurements on coal thickness and sulfur content. The sample data and their
geologic interpretation represent the most regionally complete and up-to-date assessment of what is known about top-producing
US coal beds. The sample data are analyzed using a combination of geologic and Geographic Information System (GIS) models
to estimate tonnages and qualities of the coal beds. Traditionally, GIS practitioners use contouring to represent geographical
patterns of ‘‘similar’’ data values. The tonnage and grade of coal resources are then assessed by using the contour lines as
references for interpolation. An assessment taken to this point is only indicative of resource quantity and quality. Data users may
benefit from a statistical approach that would allow them to better understand the uncertainty and limitations of the sample data.
To develop a quantitative approach, geostatistics were applied to the data on coal sulfur content from samples taken in the
Pittsburgh coal bed (located in the eastern US, in the southwestern part of the state of Pennsylvania, and in adjoining areas in the
states of Ohio and West Virginia). Geostatistical methods that account for regional and local trends were applied to blocks 2.7
mi (4.3 km) on a side. The data and geostatistics support conclusions concerning the average sulfur content and its degree of
reliability at regional- and economic-block scale over the large, contiguous part of the Pittsburgh outcrop, but not to a mine
scale. To validate the method, a comparison was made with the sulfur contents in sample data taken from 53 coal mines located
in the study area. The comparison showed a high degree of similarity between the sulfur content in the mine samples and the
sulfur content represented by the geostatistically derived contours. Published by Elsevier Science B.V.

Keywords: Geostatistical analysis; Pittsburgh coal bed; Comparison with mine data; Kriging; Statistical precision

1. Introduction

The US Geological Survey (USGS) recently com-

*
pleted a digital coal resource assessment model of the
Corresponding author. Tel.: +1-202-287-1971; fax: +1-202- Upper Pennsylvanian Pittsburgh coal bed, which
287-1934.
E-mail addresses: [email protected] (W.D. Watson),
indicates that out of the original 34 Gt (short tons)
[email protected] (L.F. Ruppert), [email protected] (L.J. Bragg), [31 Gt] in the bed, 16 Gt (short tons) [14 Gt] remain
[email protected] (S.J. Tewalt). after subtracting mined-out coal and coal lost in

0166-5162/01/$ - see front matter. Published by Elsevier Science B.V.

PII: S 0 1 6 6 - 5 1 6 2 ( 0 1 ) 0 0 0 3 5 - 0
2 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

mining (Ruppert et al., 1999). After technical, envi- our sample data. Each set of ‘‘new draws’’ is used to
ronmental, and safety restrictions are applied to the repeat the estimation of sulfur content by block. This
remaining Pittsburgh coal, only 12 Gt (short tons) [11 procedure allows us to report prediction error bars for
Gt] are available for mining (Watson et al., 2000). the cumulative distribution of tons of remaining coal
This analysis is an example of regional coal avail- in the Pittsburgh bed by sulfur content. While these
ability studies currently underway to estimate mine- results provide an informative summary of what we
able resources for major coal beds in the US. can currently infer about sulfur in the remaining
This study introduces methodology and prelimi- Pittsburgh coal, it must be emphasized that additional
nary results of USGS’s follow-on analyses of coal data could change the modeled structure and alter the
quality using the Pittsburgh coal bed as an example. estimates of the cumulative distribution significantly.
Using Geographic Information System (GIS) methods Previous applications of geostatistical methods to
and geostatistical modeling, the sulfur content of the the estimation of the sulfur and ash contents in coal
coal resources remaining in the Pittsburgh bed was include studies by Gomez and Hazen (1970), Hohn et
estimated. The analysis was conducted by (1) con- al. (1988), Murphy and Brown (1993) and Cressie
structing a database of sample point data on total (1993). Interestingly, all but Murphy and Brown
sulfur content (wt.%) on an as-received basis (Tewalt (1993) applied geostatistics to sections of the Pitts-
et al., in press), (2) removing a small number (11 out burgh coal bed. Gomez and Hazen (1970) evaluated
of 738) of samples found to be extreme outliers from coal ash and sulfur for the Robena mine on the
an examination of spatial correlation, (3) applying Pittsburgh coal bed in the southwestern corner of
exploratory data analysis to identify a regional trend the state of Pennsylvania, US. Gomez and Hazen
in sulfur with predominantly lower sulfur values (1970) applied a multivariate statistical model that
occurring at eastern locations, (4) constructing an related coal sulfur to ash, sample location, roof type,
empirical variogram that demonstrates coal sulfur and bed thickness. Their objective was to predict
content as spatially correlated for sampling locations organic and pyritic sulfur content to guide mining
up to 24 mi (38.6 km) apart, (5) applying universal decisions, because mainly only pyritic sulfur can be
kriging to the sample data to account simultaneously removed in coal washing plants. In their analysis, they
for the regional east – west trend and local spatial constructed contour maps showing the percentage of
correlation, and (6) applying block kriging to estimate pyritic sulfur that is expected to be removed by coal
expected coal sulfur content and its standard error. washing. Their study broke new ground in terms of
This last step, applied to areas of about 7 mi2 (18.1 data exploration techniques and the creative applica-
km2) with coal adequate to support several large tion of geostatistics to a challenging problem. Later,
prospective coal mines, adjusts the analysis to a scale Cressie (1993) used the same data from the Robena
appropriate for regional assessment. Therefore, these mine to illustrate the power of kriging (a parametric
estimates are not good indicators of coal sulfur con- statistical method described below) as a technique to
tent at a mine scale. Our sample data set is too small predict coal ash content. Cressie identified an east-to-
and of insufficient density to evaluate sulfur distribu- west regional trend in the coal ash data (lower in the
tions and trends at a mine scale. Only later, as mining east, higher in the west), removed this trend from the
companies undertake closely spaced, gridded, in-fill data, and applied kriging to the de-trended data to
sampling, will it be possible to determine, more predict coal ash content at non-sampled points. Mur-
precisely, the location of significant blocks of coal phy and Brown (1993) applied kriging to cores taken
within a target sulfur range at mine scale. However, in a targeted mining area, in order to determine
the large block estimates do provide information whether the coal would meet contract sulfur specifi-
about the likely availability of coal within sulfur cations. Hohn et al. (1988) applied kriging to widely
ranges, and these estimates are generalized indicators spaced observations on coal total sulfur content and
at a regional level of where additional sampling is computed cumulative frequency sulfur distributions
needed to delineate coal within a target sulfur range. for coal in blocks with areas of 1 mi2 (2.6 km2).
Finally, we have simulated random fields for coal The objective of the current study is similar to that
sulfur content based upon the covariance structure in of Hohn et al. (1988), namely, to gain a better under-
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 3

standing of the effects of statistical variability, and the content) was squared and used to calculate the average
role of sample spacing in choosing a scale (for of the squared differences in coal sulfur content for all
example, mine scale or regional scale) so that the the binned pairs at the specified span in distance. The
analysis has support from the data. In addition, this equation for this calculation, with a division by 2, is
study extends the scope of the earlier studies by
1 X
applying kriging methods, to estimate the statistical cðhÞ ¼ ðzi  zj Þ2 ð1Þ
precision of the cumulative frequency distribution for 2jN ðhÞj N ðhÞ
coal sulfur content.
As an aid to the discussion, Table 1 contains where N(h) is the set of all pairwise Euclidean
definitions of geostatistical terms used in the study distances i  j = h, | N(h)| is the number of distinct
to describe the methods and results. pairs in N(h), and zi and zj are data values at spatial
locations i and j, respectively. Fig. 1 is a boxplot of
the square-root-differences for the point sulfur data,
2. Data obtained by taking the square root of the difference
(in absolute value) of the coal sulfur content between
The Pittsburgh coal bed database on sulfur content paired sample points, rather than squaring the differ-
(wt.%) consists of data from core, mine, and outcrop ence, as in Eq. (1). The boxplot of the square root
samples on an as-received basis and was derived from differences has been shown to be a particularly good
a variety of sources including the US Geological tool for identifying atypical observations as opposed
Survey, US Bureau of Mines, the Pennsylvania State to observations that occur because distributions are
University, as well as other federal and state sources. skewed (Kaluzny et al., 1997). The boxplot shows
From a total of 3377 analyses, only 1017 samples had the spread of the square root of the difference in
locations specified by latitude and longitude coordi- sulfur content values (vertical axis) between pairs of
nates. These geo-referenced samples were used in this sulfur samples for different separation distances
study. Out of the 1017 samples, 919 samples were (horizontal axis). The open circles at low separation
found to be located within the large resource area distances (Fig. 1, upper left side) represent sulfur
(labeled A on Fig. 2) of the Pittsburgh coal bed, which content differences that lie well outside the spread of
is the target study area for this analysis. These 919 the rest of the data. Because our objective was to
sample data points have a high degree of clustering examine the data for spatial correlation at a regional
and many points are located in areas where the coal scale, the sulfur samples responsible for these
has already been mined. A number of the samples extreme differences were removed from the sample
have nearly the same location, and elimination of set. In this case, 11 sulfur samples, which are
samples within 50 ft of other samples further reduced scattered throughout the sampled area, account for
the sample set to 738 records. In these cases, a random the extreme outlying differences. Six of these sam-
selection was made from the near co-located samples ples range in value from 5.0% to 6.75% sulfur. The
and a single record at each location was kept in the remaining five samples range in value from 0.7% to
data set. In most cases, the samples at closely spaced 1.5% sulfur. In the areas where they are located,
points had similar values. these samples have values that fall far outside the
Multiple values at near-duplicate locations can range of the sulfur values for surrounding samples.
make it impossible to solve for the geostatistical After their removal, 727 point samples on sulfur
kriging weights (discussed below). Furthermore, in remain (Fig. 2).
this analysis, the emphasis is on block or regional- While many of the samples occur at locations
scale variation in coal sulfur content, not micro-scale where coal has been mined out, they are valid obser-
variation in densely sampled areas. vations for estimating the sulfur content in the large
The 738 sample data points were grouped or contiguous area with remaining coal. The geostatis-
binned according to a specific span of distance sepa- tical procedure used in the analysis assumes that the
rating every unique pair of sample values. The differ- processes that formed sulfur content in the coal were
ence in each pair’s values (difference in sulfur operating over a continuous spatial extent and, there-
4 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

Table 1
Definitions of geostatistical terms
Term Definition Source
Kriging A statistical prediction method that analyzes correlation of sample values in Myers (1997), p. 344
terms of their spatial proximity to each other and uses a mathematical
representation of their spatial correlation to make predictions at locations
where the values are not known. In theory, the predictions are unbiased and
have the smallest prediction error among predictors calculated as the
weighted sum of known values.
Block kriging The use of kriged point estimates of a characteristic to predict the average Olea (1999), p. 188
of a characteristic within a specified area or block. In theory, the predicted
block averages are unbiased and have the smallest prediction error compared
to other predictors that use weighted sums of known values.
Universal kriging Approximately, known values are regressed on location to form a fitted Deutsch and Journel (1997), p. 64
surface, known values are subtracted from their computed value on the fitted
surface, and kriging is applied to the resulting difference. Predictions are the
composite of the predicted differences from the fitted surface (calculated
using the estimated kriging weights) plus the trend prediction from the fitted
surface.
Cumulative frequency The graph of a characteristic for a variable from lowest to highest plotted on
distribution the horizontal axis versus the cumulative associated sum of the variable
plotted on the vertical axis. See Figs. 12 and 13.
Euclidean distance The distance between two points as measured by the length of the
hypotenuse of the right triangle when each of the points is located at
opposite ends of the hypotenuse. Equal to the square root of
(x2  x1)2+( y2  y1)2, where (x2, y2) and (x1, y1) are the locations of points
2 and 1, respectively.
Boxplot A graph of the distribution of values for a variable. The ‘‘box’’ part of Venables and Ripley (1997), p. 172
the boxplot covers the range of values that make up the middle 50% of
the values. Starting from the box, ‘‘whiskers’’ go out to the extremes of the
data. Whiskers are placed at the most extreme values or at the box end, plus
or minus (usually) 1.5 times the 50% middle range. Very extreme points
outside the whiskers are shown by themselves. See Fig. 1.
Kriging weights The mathematical solution to an optimization problem that solves for Myers (1997), p. 344 – 349
coefficients to be used as the weights pre-multiplying known values to
predict values at locations with unknown values. The optimization problem
is set up to mathematically guarantee that the weighted sum will have
zero expected prediction error and the most precision (or lowest variance) for
the prediction.
Variogram A plot of the average value (vertical axis) calculated using Eq. (1) for Kaluzny et al. (1997), p. 68
unique pairs of points versus the average distance (horizontal axis) separating
the pairs. The sample data are grouped within distance ranges (that is,
‘‘binned’’) for the purpose of estimating the variogram. Each bin of sample
data yields one point for the variogram. See Fig. 8.
Directional variogram A variogram estimated using sample points that fall along a specified direction. Kaluzny et al. (1997), p. 74
A tolerance on the direction, such as 11.25°, is set. The variogram is based on
all pairs of points that fall within the direction gradient plus or minus
the tolerance. See Fig. 7.
Nugget effect Represents micro-scale variation or measurement error. It is estimated from Kaluzny et al. (1997), p. 69
the empirical variogram as the value from Eq. (1) when the separation
distance between data points is zero. See Fig. 8.
Sill The upper limit of the empirical variogram representing the Kaluzny et al. (1997), p. 69
variance of the characteristic being analyzed. See Fig. 8.
Range The distance at which the characteristic being analyzed is no Kaluzny et al. (1997), p. 69
longer spatially correlated with values of the characteristic at
other points. See Fig. 8.
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 5

Table 1 (continued)
Term Definition Source
Realization Values observed from a single outcome of a process that is
stochastic. Because the process is stochastic or probabilistic,
any additional outcomes of observed values (i.e., realizations)
will differ from other observed outcomes and will fall within a
range and at a frequency that are determined by the
probability distributions that govern the process.

fore, the values may be spatially correlated relative to the predicted sulfur contents at such distant locations,
the distance separating the sample pairs. Thus, sam- signaling low precision for the sulfur prediction.
ples from areas where mining has occurred are valid Another factor that guided the delineation of the area
for estimating coal sulfur content in remaining adja- for analysis was consistency in geologic conditions.
cent areas. However, when samples in mined areas are The southernmost extent of the coal in West Virginia
a long way away from areas of remaining coal (where may have formed under geologic conditions that were
we want to make predictions), the geostatistical pro- different than the conditions in the other parts of the
cedure will tend to calculate a large standard error for coal bed. Therefore, because spatial continuity and

Fig. 1. Boxplots of the square-root-difference variogram for the sulfur point-sample data (modified from Eq. (1)). The open circles in the boxes
are the mean values for the square root of the sulfur content difference; the horizontal lines are the median values for the square root of the sulfur
content difference (1 mi = 1.609 km).
 6
W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22
Fig. 2. Location of sulfur samples used to analyze the sulfur content of the Pittsburgh coal bed.
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 7

geologic conditions are similar only in the large 3.1. Why it is important to account for systematic
resource area A (Fig. 2), the estimation is not regional and local trends
extended to the separate coal areas in Fig. 2.
The spatial variation in the sulfur content of coal
most likely reflects systematic variation at both
3. Analysis regional and local scales. The regional variation can
have its source in large-scale geologic processes that
The geostatistical methods of Cressie (1984, 1993) could have concentrated sulfur in the tops and
and Kaluzny et al. (1997) were used to predict average bottoms of coal beds (Williams and Keith, 1979;
sulfur content and its standard error for the remaining Yancy and Faser, 1921), suggesting that thicker coal
coal in the Pittsburgh coal bed. Cressie analyzed coal may have a lower sulfur content than thinner coal.
ash data for a section of the Pittsburgh bed (Cressie, For the Pittsburgh coal bed, this appears to be true, as
1984) and later (Cressie, 1993) used the analysis of there is a general trend of thicker, lower sulfur coal in
coal ash data to illustrate spatial data analytic techni- the east, and thinner, higher sulfur coal in the west.
ques. Many of the same techniques are employed in This regional trend in sulfur values, which has been
this study. known for many years, requires the application of
A regional-scale interpolator, block kriging (Math- universal block kriging so that predictions of sulfur
Soft, 2000), is the method used in this analysis to content are unbiased and have the smallest variance
characterize the sulfur content of the remaining con- or most precision.
tiguous Pittsburgh coal bed (resource area A, Fig. 2). Geostatistical analysis seeks to estimate weights
Block kriging uses any regional trends and local which, when applied to sample values to predict sulfur
spatial correlation to develop estimates of coal sulfur content at unsampled locations, minimize mean square
content at a block level, and in the case of the current prediction error and produce unbiased predictions. For
analysis, for cells 2.7 mi (4.3 km) on a side super- the prediction error to be minimum and predictions
imposed on resource area A. The estimates from block unbiased, two conditions must be satisfied. First, if a
kriging are estimates of the average sulfur content regional trend is present, the method of determining
(wt.%) and the standard error for the average, for each kriging weights has to account for the regional spatial
separate block as a whole. trend, in which case, the procedure generally known as
Block analysis was applied for several reasons. universal kriging, which incorporates the regional trend
First, the sulfur-content sample data available to the into the kriging model, should be applied (Kaluzny et
analysis exhibited a high degree of clustering and al., 1997). The second condition is that the variances of
did not have sufficient density to support meaningful the random variables (for example, sulfur content)
analysis at mine scale. Blocks with areas of about 7 across space should depend only upon the distance
mi2 (19 km2) on the Pittsburgh coal bed usually separating the realizations. The estimation of a vario-
contain enough coal to support several large coal gram (essentially a plot of gamma (c) (obtained by
mines operating for 15 –20 years. The estimates of plugging the sample data into Eq. (1)) against separa-
average sulfur content and its standard error, at tion distance) is the method used to examine whether
block scale, most likely would provide useful infor- the sample data exhibit such a relationship. If (as in Fig.
mation for regional assessment. Indeed, the estimates 7, as explained below) the directional variograms of
of the standard errors were found to cover relatively coal sulfur content values (after their east – west trend
narrow ranges compared to the ranges for point has been removed) are similar in shape, start with a low
estimates and, thus, the block-scale estimators value for c, and increase up to the variance of the
proved to be useful indicators of generalized regio- sample data, this is a demonstration that the second
nal trends in sulfur content for the remaining Pitts- condition for kriging estimates to be unbiased and
burgh coal bed. Finer scale analysis ordinarily would minimum variance is met.
be undertaken by coal-mining companies and gov- The variogram (or plot of c from Eq. (1) versus
ernment agencies at the time specific mining oper- separation distance) is the map of systematic spatial
ations are delineated. variation at a local scale. The east – west trend in
8 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

sulfur values for the Pittsburgh coal bed is our 3.2. Accounting for a regional trend
model of systematic variation at a regional scale.
Both of these variation models are used to make Exploratory data analysis is a recommended
controlled predictions of sulfur at unsampled loca- method to examine spatially located data for regional
tions, but random, unexplained variation remains. spatial trends. Using exploratory data analysis, Cres-
Thus, the predictions still have error or imprecision, sie (1993, p. 32) found an east – west trend in coal ash
and the quantification of the amount of imprecision data for the Pittsburgh coal bed. Similar to Cressie’s
in the predictions is of high interest and utility. findings, we have found an east – west trend in coal
Fortunately, because it is based upon mathematical sulfur data for the Pittsburgh coal bed. In Cressie’s
statistics, universal block kriging has the capability analysis, the coal ash data were spaced on a uniform
to estimate the precision of the sulfur predictions grid. Consequently, Cressie was able to use row-by-
and such estimates of precision are presented below row and column-by-column plots to uncover the east –
in Section 4. west trend. A similar method is followed for our non-

Fig. 3. Sulfur content values in an east – west direction, conditioned on their north – south location. The lines drawn through the points are non-
parametric local regression lines that indicate a generalized trend (1 mi = 1.609 km).
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 9

uniformly spaced data: examination of paneled sam- variation, in order to satisfy statistical conditions
ple data for east –west and north – south orientations. needed to have unbiased and minimum-variance
The east –west trends in coal sulfur content and predictions of coal sulfur content.
ash for the Pittsburgh coal bed have been known for In Fig. 2, the large dot in the northeast corner is the
a long time (Cross, 1954). For this analysis, as well origin point for determining the location of the sample
as in Cressie’s analysis, the examination of the data points. From the origin point, the sample points span a
for a regional trend is undertaken so that the correct distance of 100 mi (161 km) to the west and 120 mi
procedure can be identified for the geostatistical (193 km) to the south. The top part of Fig. 3 shows
analysis. As discussed previously, if a regional trend that the sample data were split into six north-to-south
exists, then the sample data need to be de-trended (slightly overlapping) panels, each containing equal
before kriging is applied. For implementation, it is numbers of sulfur content sample points (Fig. 4). The
necessary to have a statistical model of the regional bottom part of Fig. 3 contains plots for each panel, of
trend, which is estimated within the universal kriging sulfur content versus east-to-west distance, for all the
program by regressing coal sulfur content values respective points located in each panel. The plots for
against their east-to-west locations. The importance panels 1, 2, 3 and 4 indicate lower sulfur values in the
of the regional trend to this analysis is not the east, and higher sulfur values in the west. Panels 5 and
existence of the trend, which has been known for 6 are for the ‘‘peninsula’’-like area that occurs in the
many years. Rather, it is the systematic statistical southern part of the resource block A (Fig. 4), where
analysis of the trend, separately, from local spatial the narrow span of the coal deposit limits the distance

Fig. 4. Panels used (Fig. 3) to examine sulfur values for regional trends in an east-to-west direction. Each panel contains the same number of
points. Panels overlap in order to see if a generalized regional trend is exhibited by the data (1 mi = 1.609 km).
10 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 11

Fig. 6. Sulfur content values in a north – south direction, conditioned on their east – west location. The lines drawn through the points are non-
parametric local regression lines that indicate a generalized trend (1 mi = 1.609 km).

between sample values and precludes identification of the point predictions of coal sulfur content are not of
possible regional trends. interest to this analysis. Rather, we wish to be able to
It should be noted that even though there is an predict the coal sulfur content (and its statistical
indication of a regional trend, a lot of scatter still precision) for blocks of coal over an areal extent
remains in the observations of coal sulfur content. A capable of supporting several large coal mines for a
part of that variation will be accounted for by the local period of 20 years. Thus, we apply block kriging to
kriging model as applied (below) to the variation estimate the average sulfur content and its standard
remaining after the regional trend is removed from error over a support area that is 2.7 mi (4.3 km) on a
the data, but even after that step, considerable uncer- side. It turns out that the standard errors at block level
tainty remains in predictions of point-located values. are narrow enough to provide predictions that are
However, because the focus is on regional assessment, informative at the block or regional level.

Fig. 5. Panels used (Fig. 6) to examine sulfur values for regional trends in a north-to-south direction. Each panel contains the same number of
points. Panels overlap in order to see if a generalized regional trend is exhibited by the data (1 mi = 1.609 km).
12 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

Resource area A was split into six panels with a kriging as the appropriate geostatistical method (Cres-
north-to-south orientation to examine the coal sulfur sie, 1993). Universal kriging accounted for the east-
content sample data for the presence of a regional to-west trend by fitting a surface for coal sulfur
north-to-south trend (Fig. 5). There is no apparent content as a function of east-to-west location. These
north-to-south directional trend in sulfur values except regionally de-trended data were then analyzed for
for panel 6 (Fig. 6), which indicates a trend from lower local spatial correlation using kriging methods.
sulfur values in the north to higher sulfur values in the
south. Panel 6 includes the western-most part of block 3.3. Examining local spatial correlation in different
A (Fig. 5) of the Pittsburgh coal bed. Most of the directions
sample data in panel 6 are located in areas where the
coal has been mined. The area with remaining coal, Directional empirical variograms, developed after
which has few sample data points, occurs mainly to the the influence of an east – west trend is removed, are
south of most of the sample data points in panel 6. similar to each other indicating that spatial correlation
Consequently, estimates of sulfur for the remaining is isotropic or the same, irrespective of direction. For
coal in the west will tend to have a high amount of example, Fig. 7 shows variograms (using Eq. (1))
uncertainty. plotted for sulfur content (after their east – west trend
The finding of an east-to-west trend in coal sulfur was removed) in four directions: north to south (0°),
values (Fig. 3) supported the choice of universal northeast to southwest (45°), east to west (90°), and

Fig. 7. Directional variograms after the east – west trend was removed from coal sulfur content sample values (1 mi = 1.609 km).
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 13

southeast to northwest (135°). The variograms all strates, the sulfur content of the Pittsburgh coal bed
have the same appearance in a range of about 25 mi exhibits a pattern of local spatial statistical correlation.
(40.2 km). The variogram along the 45° direction Sulfur samples within close proximity to each other
contains low sulfur values in the northeast section of tend to be similar, thus, variance (or gamma) is small
the Pittsburgh coal bed separated by a long distance at short separation distances. As the geographic dis-
from low sulfur values in the Pittsburgh ‘‘peninsula’’. tance between samples increases, variance increases
Because the Pittsburgh ‘‘peninsula’’ has many low and statistical spatial correlation decreases.
sulfur points, as does the northeast section, gamma (c, In the context of a kriging model, a local spatial
calculated using Eq. (1)) is relatively low at high trend is said to exist if values separated by short
separation distances in that case. distance are more similar than values separated by
This outcome demonstrates that the variance across long distances. A plot of Eq. (1) is the exploratory
the sulfur values depends only upon the distance data method used to examine the sample data for
separating the values, and not upon the directions in such local spatial trends. To ensure that local trends
which the values are arrayed. Thus, our data satisfy are not confounded or masked by regional trends, the
the second condition outlined in Section 3.1 and the technique is to examine the local relationship along a
predictions from the final kriging model will have direction gradient that does not exhibit a regional
minimum variance and be unbiased. trend. In the previous section, it was found that
along a north-to-south direction gradient, coal sulfur
3.4. Estimation of an empirical variogram content does not exhibit a regional trend. Therefore,
coal sulfur content sample values located in the
Local-scale variability of coal bed characteristics, north-to-south direction gradient were plugged into
such as thickness and quality, can equal or exceed Eq. (1) to calculate c and create the empirical
regional variability. Nonetheless, as Fig. 7 demon- variogram (Fig. 8).

Fig. 8. Empirical variogram based upon 727 sulfur values for the Pittsburgh coal bed. The empirical variogram is developed from data pairs
arrayed in a north – south direction, where there is no regional spatial trend. A spherical variogram model was fit to the empirical variogram (1
mi = 1.609 km).
14 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

Gamma (as calculated by Eq. (1) and plotted in Fig. forms have been found to have this property. The
8) is a measure of similarity between sample values as spherical variogram model used in this analysis is
influenced by the distance that separates the values. among the models that guarantee a solution (Cres-
When the pair of data samples has coal sulfur content sie, 1993). In addition, the spherical model, rather
values that are very close in numerical value, c will be than another model, was chosen, because it appears
a relatively small number. The plotted value for c was to fit the empirical variogram better than other
found to be relatively small when a short distance models.
separates the coal sulfur values (near the origin on the The spherical variogram model fit to the empirical
horizontal axis in Fig. 8). As the separation distance variogram has a range estimated to be 24.1 mi (38.8
between sample values increases (moving to the right km) (Fig. 8). Sulfur estimates separated by less than
along the horizontal axis in Fig. 8), the computed value 24.1 mi (38.8 km) are spatially correlated. Spatial
for c increases, indicating that the coal sulfur values in correlation of sulfur estimates is inversely related to
each pair are not as close in value to each other as they gamma, and approaches zero as separation distance
are when the distance separating the samples is shorter. approaches 24.1 mi (38.8 km).
When such a pattern is observed, the values are said to In principle, at extremely close distances, sample
be spatially correlated. sulfur values should be identical, and, therefore,
The empirical variogram (Fig. 8) shows a clear variation between values should be equal to zero.
pattern of local spatial continuity. At short separa- When variance at zero distance is non-zero, as
tion distances, the magnitudes of the coal sulfur measured by the empirical variogram, it is termed
contents in the sample are similar. As the distances the nugget effect (Fig. 8). The presence of a nugget
separating the locations of data pairs increase, the effect indicates the presence of measurement error
degree of spatial continuity tapers off and disap- or unexplained short-scale variability in pyrite con-
pears at about 25 mi (40.2 km). Above a separation centrations. The ‘‘sill’’ ( = 1.37% squared, Fig. 8) is
distance of 25 mi (40.2 km), the variance in coal approximately the variance of the sulfur content
sulfur content for data pairs is practically equivalent values, as estimated from the sample data. The
to the variance for coal sulfur content calculated difference between the sill and the nugget effect
across all the samples (727) in the data set ( = 0.91% squared) is an indicator of the important
( = 1.36% squared). role that spatial correlation plays in making esti-
mates of sulfur content at other points. At extremely
3.5. Model variogram to represent local spatial short distances, the spatial correlation reduces vari-
continuity in coal sulfur content ability in coal sulfur content estimates by almost
0.91% squared. As separation distance approaches
Once it has been established that sample data 24.1 mi (38.8 km), spatial correlation approaches
exhibit local spatial continuity or correlation, then zero and the variability of sulfur content values
the next step is to fit a functional form to the data approaches 1.37% squared, or approximately the
that make up the plot of c against separation distance. variance of the sample data.
In the kriging literature, this step is referred to as
‘‘estimating the model variogram’’ (Kaluzny et al., 3.6. Kriging model of sulfur in the Pittsburgh coal bed
1997).
The solution for kriging weights requires esti- Regional trends and local spatial trends are com-
mates of spatial correlation between sample points bined to estimate a sulfur kriging model for the
and prediction points. These estimates of spatial Pittsburgh coal bed. In this case, the statistically
correlation are derived from the model variogram, efficient procedure is universal kriging, which deter-
which is a functional form fitted to the empirical mines kriging weights that minimize mean square
variogram. Importantly, the model variogram has to prediction error (Cressie, 1993, p. 173; Goldberger,
have a form that guarantees a mathematical solution 1962). Fig. 9, a wireframe diagram of the kriged
for the system of equations that determine kriging sulfur surface, clearly shows that the model captures
weights (Cressie, 1993). Only certain functional both regional and local variation in sulfur.
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22
Fig. 9. Kriged sulfur surface for the Pittsburgh coal bed. The wireframe surface floats above the sulfur content sample point values shown as triangles. The large square at the origin
(0,0) corresponds to the large origin square in Fig. 2 (1 mi = 1.609 km).

15
16 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

4. Sulfur predictions near 2% have an intermediate level of uncertainty;

their standard errors range from 0.42% to 0.69% S. As
The data available for our model were far below an example of interpretation, a block with an esti-
the level that would support detailed assessment. mated average of 2.2% S and a standard error of
Therefore, it is only appropriate to estimate coal sulfur estimation of 0.5% S would have a 95% chance of
content at a large-area scale. We have applied univer- having its true average sulfur content in the range
sal block kriging (Isaaks and Srivastava, 1989; Math- between 1.2% and 3.2% S (assuming a normal dis-
Soft, 2000) to estimate the average sulfur content for tribution and a standard normal variate = 1.96).
square areas, 2.7 mi (4.3 km) on a side. Watson et al. Fig. 12 combines the estimate of sulfur content
(2000) suggested that an area of this size has enough with an estimate of coal tons to create a cumulative
coal for several large-scale mines operating over a 20- distribution of remaining tonnage by sulfur content
year period. The block estimates can be used as a and upper and lower confidence limits at the 90%
basis for identifying areas where additional drilling level for average sulfur content, block by block.
could be targeted to refine sulfur estimates. Also, the Blocks at every level of predicted average sulfur
block estimates and the procedures used to estimate content have wide variation in estimation precision
sulfur by block were the basis for estimating a (Fig. 12). Fig. 11 could provide a basis for prioritizing
cumulative distribution of remaining Pittsburgh coal areas for additional core drilling to refine estimates.
by sulfur content. As discussed in the Section 1, the For example, if it is thought that coal mining costs are
estimates are not suitable for supporting inferences relatively low for a particular block, and the block has
about sulfur content at a mine scale. relatively low average sulfur content (Fig. 10) but
wide confidence limits for sulfur content (Fig. 11),
4.1. Sulfur predictions by coal block then mining companies or land brokers may find it
worthwhile to undertake additional sampling to fine-
Results are summarized in Figs. 10– 12. Average tune these preliminary regional-scale estimates of
sulfur values by block are the basis for the sulfur sulfur content.
contours shown in Fig. 10. The sulfur contours are a
prediction of the average sulfur content at block scale. 4.2. Error bars for the cumulative sulfur distribution
Based upon a statistical model of point values, new
cores drilled in a given block can have sulfur values The average sulfur cumulative curve plotted in Fig.
that cover a range of around 2% S above and below 12 is a single observed realization of a stochastic
predicted point values. In the case of the block process. The sample data allow us to make inferences
estimates, high extreme and low extreme values are about the stochastic process that gave rise to these
averaged with other more-numerous intermediate val- estimated values. Our model of that process is repre-
ues to form a block average. Thus, block averages sented by an east – west trend and by the local varia-
ordinarily will have prediction errors for the average tion captured in the spherical variogram model. The
sulfur value of the block that are narrower than statistical model was used to simulate alternate real-
prediction errors for point estimates. izations from the ‘‘known estimated’’ processes. This
Fig. 11 is a map of the standard errors for the procedure has the advantage of retaining the interac-
estimates of average sulfur content at the block level. tion between trend and local variability. Additional
The most precise estimates (those with the smallest realizations will generate values in a geographical
standard error) occur mainly in blocks with a rela- pattern that will pair blocks (that have given amounts
tively large number of data samples. The blocks with of coal) with a different ordered set of sulfur values,
the smallest standard error (0.11– 0.14% S), shown in thereby generating different cumulative distributions
dark blue, are located near the bottom of the Pitts- for remaining Pittsburgh coal by sulfur content. A
burgh ‘‘peninsula’’, where there is a concentration of large set of these alternate realizations of the cumu-
samples. The relatively high precision is due to the lative curve would provide an estimate of the mean,
dense sampling pattern. In the northeast section of the upper, and lower limits of the cumulative curve by
coal, the blocks with an estimate of average sulfur average sulfur content.
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22
Fig. 10. Contours of estimated average sulfur content (wt.% as-received basis) for blocks, 2.7 mi on a side (1 mi = 1.609 km).

17
 18
W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22
Fig. 11. Uncertainty in the estimates of block average sulfur content. Blocks color-coded from lowest to highest in order of the estimated standard errors for the average sulfur content
by block.
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 19

Fig. 12. Remaining coal in order by block average sulfur content, including upper and lower 90% confidence limits (1 short ton = 0.9072 metric
ton).

Our procedure, similar to Isaaks and Srivastava range that has about a 95% chance of covering the
(1989), p. 512), is to translate the sample data loca- true distribution. Each set of new realizations is used
tions. The translation is designed to span two block to estimate average sulfur values at the centroids of
widths around original locations, which is approxi- the cells in the 2.7 mi (4.3 km) block grid. A new
mately equivalent to spanning by two standard normal cumulative distribution of average sulfur content is
variates at the block level. This procedure generates a estimated by ordering the sulfur values from low to

Fig. 13. Cumulative distribution of remaining Pittsburgh coal by average sulfur content, blocks 2.7 mi on a side, expected, upper 95%
confidence level, and lower 95% confidence level cumulative distribution (1 mi = 1.609 km; 1 short ton = 0.9072 metric ton).
 20
W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22
Fig. 14. Comparison of sulfur (wt.%, as-received basis, washed) in coal shipments with sulfur content values (wt.%, as-received basis) predicted by the statistical model.
 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22 21

high and pairing the sulfur value with cumulative 4.3. Comparison of model estimates with sulfur values
tonnage built up from the linked block tonnages. in shipped Pittsburgh coal
The outer hull of the distributions forms an estimate
of the upper and lower confidence limits for the By combining data from EIA Form 7A (Energy
cumulative distribution of remaining Pittsburgh coal Information Administration, 2000) and a commercial
by sulfur content (Fig. 13). Because the simulations database of coal shipments (Resource Data Interna-
capture trend and local variability, the confidence tional, 1998), shipment data were assembled from
range for the cumulative distribution is expected to mines extracting coal only from the Pittsburgh bed
be more narrow than the average of the confidence from 1989 to 1997. The shipment data report the
range of the individual block values illustrated in Fig. sulfur content of the coal as-received by power plants.
12. Most of the coals were washed prior to shipment and,
Fig. 13 is useful for predicting the future avail- thus, the sulfur values for these data will be less than
ability of Pittsburgh coal by sulfur content. Assuming the raw coal values reported thus far in this paper. The
that the market will only accept coal with a sulfur data set has shipment records for 53 different mines.
content of 2.15% or less, and Pittsburgh coal is Because the average sulfur values for the shipments
washed prior to delivery, raw coal up to about are representative of large mined blocks, they can be
2.75% would be marketable. Removal of approxi- compared (on more or less equal statistical footing)
mately 0.6% S agrees with coal sulfur values reported with the estimates from our model. The shipment data
for shipments of cleaned Pittsburgh coal (Resources are used to make a new set of contour lines showing
Data International, 1998). Fig. 13 indicates that there sulfur content for the northern extent of the unmined
are between 1.9 and 5.5 Gt (short tons) [1.7 and 5 Gt, Pittsburgh coal bed (Fig. 14).
respectively] still remaining at sulfur contents below Bearing in mind that washing removes about 0.6%
2.75%. The best estimate of availability is 3.5 Gt S, there is similarity between the red contour values
(short tons) [3.2 Gt]. The current rate of annual for the sulfur content of shipped (washed) coal and the
production from the Pittsburgh coal bed is around black contour values of raw coal as estimated by the
80 Mt (short tons) [73 Mt], recovery rates are about statistical model (Fig. 14). For example, the 2% sulfur
70%, wash recovery rates for the coal component are contour for shipped coal mimics the 2% sulfur contour
about 90%, and about 25% is unavailable due to for raw coal but encompasses a slightly larger area
technical, social, and environmental restrictions (Wat- extending to the west. The larger area includes raw
son et al., 2000). Therefore, about 170 Mt (short tons coal above 2% in sulfur cleaned down to the 2% level
[154 Mt] of resources are needed for coal deliveries of for delivery. The 2% shipped-sulfur contour arches
80 Mt (short tons) [73 Mt]. At the current rate of around the Bailey and Enlow Fork mines, two of the
production of 80 Mt (short tons) [73 Mt] per year, the largest producers of Pittsburgh coal. These two mines
remaining life for production from the Pittsburgh coal deliver washed Pittsburgh coal containing about 1.6%
bed is bracketed between 11 and 32 years, with the sulfur and appear to mine raw coal a little bit in excess
best estimate being 21 years. In combination with of 2% sulfur. The data set for shipments contains
similar cumulative curves for other eastern coal beds, significantly fewer observations than the data set used
questions concerning the availability of low sulfur for statistical analysis. Therefore, the comparison is
coal in eastern coal fields could be addressed quanti- only indicative of consistency in the statistical model.
tatively. Although sulfur content is indicative of
marketability, many other factors such as mining cost,
transport cost, and growth of customer base also 5. Conclusions
would have to be considered. Nonetheless, coal avail-
ability by sulfur content would be a key input to The US Geological Survey is assessing regional
comprehensive analysis. New sample data should be coal resources and providing regional-scale informa-
added as it becomes available and the analysis tion about coal tonnages and coal quality, including
repeated. Newer data would refine the results and, sulfur content. The data available for such assess-
most likely, would narrow uncertainty ranges. ments are limited. The survey’s interpretation inten-
22 W.D. Watson et al. / International Journal of Coal Geology 48 (2001) 1–22

tionally is focused on regional trends. Statistical in coal seams by statistical response surface regression analysis.
methods offer a valuable tool to aid interpretation. U.S. Bureau of Mines Report of Investigations 7377, 120 pp.
Hohn, M.E., Smith, C.J., Ashton, K.C., McColloch Jr., G.H., 1988.
In this research, geostatistical methods were applied to Mapping coal quality parameters for economic assessment (Ab-
estimate the cumulative distribution of remaining coal stract). AAPG Bull. 72, 965.
by sulfur content for the Pittsburgh coal bed. This type Isaaks, E., Srivastava, R., 1989. An Introduction to Applied Geo-
of information, accompanied by a quantitative esti- statistics. Oxford Univ. Press, New York, 561 pp.
Kaluzny, S., Vega, S., Cardoso, T., Shelly, A., 1997. S+ Spatial Stats
mate of its uncertainty, can be an important input to
User’s Manual for Windows and UNIX. Springer, New York,
public and private decision-making. Whenever possi- 327 pp.
ble, the predictions from kriging models should be MathSoft, 2000. S + Spatial Stats Version 1.5 Supplement. Math-
compared against real data. Because there are many Soft, Seattle, WA, 84 pp.
active mines on the Pittsburgh coal bed, it was Murphy, T.D., Brown, K.E., 1993. Combining geostatistics and
possible to compile a data file of actual sulfur content simulation to predict sulfur at a central Illinois coal mine.
Min. Eng. 45, 284 – 287.
values from mines located throughout the area of the Myers, J.C., 1997. Geostatistical Error Management: Quantifying
Pittsburgh coal bed. We found good agreement Uncertainty for Environmental Sampling and Mapping. Van
between our modeled predictions of coal sulfur con- Nostrand Reinhold, New York, 571 pp.
tent and these actual values, indicating consistency in Olea, R.A., 1999. Geostatistics for Engineers and Earth Scientists.
the geostatistical model. Kluwer Academic Publishing, Boston, MA, 328 pp.
Resource Data International, 1998. COALDAT Comprehensive. Re-
source Data International, Golden, CO, CD-ROM.
Ruppert, L., Tewalt, S., Bragg, L., Wallack, R., 1999. A digital
References resource model of the Upper Pennsylvanian Pittsburgh coal
bed, Monongahela Group, northern Appalachian basin coal re-
Cressie, N., 1984. Towards resistant geostatistics. In: Verly, G., gion, USA. Int. J. Coal Geol. 41, 3 – 24.
David, M., Journel, A., Marechal, A. (Eds.), Geostatistics for Tewalt, S.J., Ruppert, L.F., Bragg, L.J., Carlton, R.W., Brezinski,
Natural Resources Characterization, Part 1. Reidel, Dordrecht, D., Wallack, R.N., and Butler, D.T., in press. A digital resource
pp. 21 – 44. model of the upper Pennsylvanian Pittsburgh coal bed, Mono-
Cressie, N., 1993. Statistics for Spatial Data. Wiley, New York, ngahela group, Northern Appalachian Basin Coal Region, USA.
900 pp. US Geological Survey Professional Paper, 1625C, Chap. C,
Cross, A.T., 1954. The geology of the Pittsburgh coal [Appalachian CD-ROM.
Basin]. Proceedings, Conference on the Origin and Constitution Venables, W.N., Ripley, B.D., 1997. Modern Applied Statistics with
of Coal, June 1, 1952, (Nova Scotia Department of Mines), pp. S-Plus. Springer-Verlag, New York, 548 pp.
32 – 111. Reprinted: The geology of the Pittsburgh coal — Watson, W., Ruppert, L., Tewalt, S., Bragg, L., 2000. The upper
stratigraphy, petrology, origin and composition, and geo- Pennsylvanian Pittsburgh coal bed: geology and mine models.
logic interpretation of mining problems. West Virginia Geol. Proceedings of the 2000 SME Annual Meeting, Salt Lake City,
and Econ. Surv., Rep. of Investigations 10, pp. 31 – 99. UT, CD-ROM. Preprint 00-26, Society for Mining, Metallurgy,
Deutsch, C.V., Journel, A.G., 1997. GSLIB: Geostatistical Software and Exploration, Inc., Littleton, CO, 12 pp.
Library and User’s Guide. Oxford Univ. Press, New York, 380 pp. Williams, E.G., Keith, M.L., 1979. Relationship between sulfur in
Energy Information Administration, 2000. Coal Industry Annual coals and the occurrence of marine roof beds. In: Ferm, J.C.,
1998. US Department of Energy, Washington, DC, DOE/EIA- Horne, J.C., Weisenfluh, G.A., Staub, J.R. (Eds.), Carboniferous
0584(98), 308 pp. Depositional Environments in the Appalachian Region. Univer-
Goldberger, A., 1962. Best linear unbiased prediction in the sity of South Carolina, Columbia, SC, pp. 102 – 109.
generalized linear regression model. J. Am. Stat. Assoc. 57, Yancy, A.F., Faser, T., 1921. The Distribution of the Forms of Sulfur
369 – 375. in the Coal Bed. University of Illinois Engineering Experiment
Gomez, M., Hazen, K., 1970. Evaluating sulfur and ash distribution Station Bulletin vol. 125, 94 pp.