Sie sind auf Seite 1von 11

International Journal of Coal Geology 112 (2013) 114–124

Contents lists available at SciVerse ScienceDirect

International Journal of Coal Geology


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

Comparison of three geostatistical approaches to quantify the impact of drill spacing


on resource confidence for a coal seam (with a case example from Moranbah North,
Queensland, Australia)
Alastair Cornah a,⁎, John Vann a, b, c, d, Ian Driver e
a
Quantitative Group, PO Box 1304, Fremantle, Western Australia 6959, Australia
b
Centre for Exploration Targeting, The University of Western Australia, Crawley, Western Australia 6009, Australia
c
WH Bryan Mining and Geology Research Centre, The University of Queensland, Brisbane, Queensland 4072, Australia
d
School of Civil Environmental and Mining Engineering, The University of Adelaide, Adelaide, South Australia 5000, Australia
e
Anglo American Metallurgical Coal, Resource Development and Operational Excellence, Level 10 201, Charlotte Street, Brisbane, Queensland 4000, Australia

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

Article history: Three approaches to characterising the uncertainty associated with coal resource estimates are presented and
Received 27 July 2012 compared: global estimation variance (GEV); local confidence intervals via the discrete Gaussian model
Received in revised form 15 November 2012 (DGM); and the conditional simulation (CS) approach. The methods are applied and compared for three
Accepted 15 November 2012
variables (Thickness, Yield and Sulphur) in a coal deposit at Moranbah North.
Available online 27 November 2012
All three approaches result in a broadly similar characterisation of uncertainty, but each has associated
Keywords:
strengths and weaknesses. GEV is appropriate for 2D situations (like coal seams), is fundamentally robust
Coal (if its assumptions are respected), it is straightforward and is quick to apply. However, being global the
Geostatistics results are somewhat limited (although in some instances a global result may still be ‘fit for purpose’), and
Drill spacing importantly, it does not properly account for skewness and proportional effect. DGM is more sophisticated
Estimation variance and accounts properly for skewness and proportional effect. It allows assessment of local uncertainty at
Conditional simulation the block scale whilst also being relatively computationally efficient but requires increased expertise to
Discrete Gaussian model implement. CS also provides local results and conceptually is the most rigorous solution. However CS is the
Proportional effect
most computationally intensive solution and requires significant amounts of user input and validation. A
short study on the influence of the number of realisations on the reliability of uncertainty assessments
made from CS models is also documented.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction The interested reader is referred to Morgan and Henrion (1990) for
more complete discussion of uncertainty concepts.
Coal seam resources are conventionally estimated and classified on Because coal resource estimates are intrinsically uncertain, charac-
the basis of input data derived from sampling of drillholes that are terisation of this uncertainty is vitally important. Resource classification
spaced at very large distances in comparison to the diameter of the schemes used internally by most mining companies and mandated by
drillholes themselves. Drillhole samples are typically several tens of many stock exchanges rely, at least in part, on expert assessment of
millimetres in diameter but inter-hole distances typically range from uncertainty. The uncertainty related to estimates of coal chemical and
hundreds of metres to over a kilometre. The ratio of sampled volume physical attributes is clearly in part related to drill hole spacing — closer
to the unsampled volume of in situ coal blocks that are estimated is spaced drilling increases the ratio of sampled to unsampled volume and
characteristically in the order of 1:105 to b1:10 7. The Concise Oxford therefore inevitably reduces the uncertainty associated with the esti-
Dictionary defines uncertain as “not known, reliable, or definite” and mated values.
in this sense, resource estimates are inherently uncertain because of Three different geostatistical methods to characterise uncertainty
this relative lack of data compared to the volumes being estimated. associated with estimation of key coal attributes are compared in
this paper. The first method is ‘global estimation variance’ (GEV),
which is a longstanding approach in linear geostatistics (David, 1977)
⁎ Corresponding author. Tel.: +61 8 9433 3511. and is well suited to coal deposits which can be modelled in two dimen-
E-mail address: ac@qgroup.net.au (A. Cornah). sions (2D), that is where the thickness is insignificant relative to the

0166-5162/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.coal.2012.11.006
A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124 115

planar extent; approaches for three dimensional modelling of heteroge- in the southern part of the case study area may be partly related to the
neous facies are provided by Faliven et al. (2007). depositional environment and the occurrence of siderite (Diessel, 1999).
GEV relies on calculation of an ‘extension variance’ which mea-
sures the average squared error incurred when we assume that a 1.2. Case study: data
block can be estimated by using a single sample value located at its
centre as the estimate. The extension variance is the term given to es- As for most essentially flat dipping coal seams in the Bowen Basin,
timation variance in the specific case where a single block is informed the Moranbah North deposit is drilled using vertical drill holes. The
by a single sample. This extension variance for a block of specified size variables considered were:
and geometry (or ‘block support’ in geostatistical parlance) can be cal-
1. Vertical seam thickness in metres (herein referred to as ‘Thickness’);
culated from the variogram of the relevant attribute. The variogram,
2. Washed yield (%) targeting an 8.5% target ash content (herein re-
in turn, is calculated from the sample data and is a model for the spatial
ferred to as ‘Yield’); and
variance; i.e., the average variance between samples as a function of
3. Raw (i.e., unwashed) total (%) sulphur content (herein referred to
separation vector between the sample pair. From the unitary extension
as ‘Sulphur’).
variance, we can easily calculate the global estimation variance associ-
ated with a volume (or area) comprised of N unit blocks. In geostatistical (spatial) analysis, an important consideration is
The second method considered is a non-linear geostatistical ap- spatial stationarity: in general a prior decision must be made about
proach based upon the ‘Discrete Gaussian Model’ (DGM) framework volumes or areas that can be considered as ‘sufficiently statistically
(see Riviorard, 1994). This framework is widely used to model global homogeneous’ for the purposes of the geostatistical tools at hand. In
or local grade and tonnage relationships; in this case it is used to pro- most cases, the operative assumption is ‘intrinsic stationarity’ which
vide the local probability of meeting a given grade criteria and thus implies that a single variogram can characterise spatial variability with-
local block-by-block confidence intervals. in a specified ‘domain’ (Wackernagel, 2003). In the 2D case, given that
The third approach is conditional simulation (CS); this involves the top and bottom of the seam have been stratigraphically defined,
the generation of multiple images of the deposit which reproduce the the main domaining considerations relate to lateral geochemical or
global declustered histogram and variogram at the scale of the sample facies variations.
data, these results can then be ‘averaged’ to generate block-scale results. For this study, two domains are of interest: the Main (MAIN) do-
Additionally the simulations honour sample data locations. From the main was defined on the basis of spatial variations in the Thickness
full set of realisations, local probability distributions can be constructed, and Yield, and the non-contiguous Low (LOWTS) domain is based
which provide a route to calculation of local (or global) confidence in- upon spatial variation of Sulphur. Base maps for Thickness and Yield
tervals. Recent studies involving applications of conditional simulation within the MAIN domain and Sulphur within LOWTS are shown in
in coal resource assessments are provided by Heriawan and Koike Fig. 3. The figure shows that the attributes are not sampled at all com-
(2008a, 2008b), and by Olea et al. (2011); another recent study which mon locations and also that the three data sets overlap. The test zone
applies of conditional simulation and in gas emission zone modelling within the overlapping area comprises eight planned longwalls which
is provided by Karacan et al. (2012). each amount to approximately one year of production. Within the
This paper explores the benefits, limitations and assumptions of test area, a total of 45 samples exist for Thickness (equating approxi-
these three methods and aims to inform the industry geologist in mately to a 375 m × 250 m drill spacing), whilst 15 and 16 samples
choosing a ‘fit for purpose’ methodology for assessing uncertainty in respectively exist for Yield and Sulphur, equating approximately to
coal resource estimates. Moranbah North Coal Mine in Queensland, a 500 m × 625 m drill spacing. Fig. 3 reveals a trend of increasing
Australia provided the data used to facilitate this comparison. Thickness from west to east, although this variable is relatively con-
sistent within the test area. The domains defined are considered to
be acceptable for the three geostatistical methods which are tested.
1.1. Case study: location and geology In earth science data sets, variables often have skewed distributions
and are heteroscedastic (i.e., subsets of the data show differences in var-
The Moranbah North coal deposit is located in the Central Highlands iability). Where variability (and consequently also estimation error
area of Queensland, Australia (see Fig. 1) and is currently being mined by variance) is related to the local mean this is known as the proportional
long-wall underground methods. The deposit is situated near the town effect (Journel and Huijbregts, 1978). Testing for the presence or ab-
of Moranbah which is approximately 180 km inland from the major sence of skewness and proportional effect is a key prerequisite to the
regional center and coastal port of Mackay. The Moranbah North Coal choice of uncertainty characterization method as they are not properly
Mine is operated by Anglo American Metallurgical Coal (formerly accounted for by all of the techniques which are discussed in this paper.
Anglo Coal). The study area is located in the northern portion of the Histograms and statistics for the three variables within entire do-
Permo-Triassic Bowen Basin and is contained in the late Permian mains are shown in Fig. 4; statistics for each within the entire domain
Moranbah Coal Measures. These Coal Measures are considered to be flu- and within the test area are shown in Table 1. The histograms indicate
vial and fluvio-deltaic in origin and formed in an upper delta plain envi- that the three variables are skewed to different extents, although not
ronment with attendant shallow marine incursions (Michaelsen and markedly so in comparison to other mineral deposits. Yield and Thick-
Henderson, 2000). ness are slightly negatively skewed whereas Sulphur has a more signif-
Overall the Moranbah Coal Measures are approximately 300 m in icant positive skew resulting from localised ‘spikes’ in pyritic Sulphur
thickness and contain nine coal seams, which may split locally (Laws and a significantly higher coefficient of variation (CV). Table 1 indicates
and Bos, 2000). Only one of these seams, the Goonyella Middle (GM) that the mean of Thickness in the test area exceeds the entire domain,
seam (which dips at about 3° to the east, see Fig. 2), currently offers but the means of Yield and particularly Sulphur within the test area
potential for economic underground extraction. are slightly lower than the domain means. The standard deviation of
The GM seam has traditionally been subdivided into five plies for coal Sulphur within the test area is also significantly lower than that of the
quality analysis. This subdivision is mainly based on the recognition of full LOWTS domain.
consistent, identifiable stone bands. A working section has been defined A simple test for proportional effect involves analyzing moving win-
(Laws and Bos, 2000) to reflect the maximum economic thickness of coal dow statistics for the key variables (a 3 km× 3 km moving window was
that could be extracted by the longwall mining method which is desig- used in this case). The average value within each window is plotted
nated the GMI (GM — Ideal Section). The GMI working section is the sub- against the standard deviation of the data within the window for each
ject of this study. A zone with poor washability characteristics occurring of the three variables in Fig. 5. Windows which contained less than
116 A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124

Fig. 1. Regional map showing the Moranbah North Area.

five samples were excluded from the analysis; a higher minimum proportional effect should be accounted for in the characterization of
would be ideal in the characterization of local variability, but was pre- Sulphur uncertainty.
cluded in this case by data scarcity. Fig. 6 shows modelled experimental variograms for Thickness,
Fig. 5 indicates that for Thickness and Yield very little dependency Yield and Sulphur over the study area. Thickness is the most continu-
exists between the mean and the standard deviation, this is a state ous of the three attributes and exhibits zonal anisotropy with greater
which is known as homoscedasticity. For Sulphur, however, a relatively variability in the east–west orientation than the north–south orienta-
strong positive linear dependency exists (Pearson correlation 0.85): tion. It is modelled using a nugget effect, a cubic structure and a
local variability is greatest in higher grade areas and least in low grade spherical structure. Yield is less continuous modelled using a nugget
areas. Sulphur clearly exhibits heteroscedasticity and therefore the effect and two isotropic spherical structures; Sulphur is the least
A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124 117

Fig. 2. Stratigraphic column showing Moranbah Coal Measures.

continuous attribute and is modelled using a nugget effect and single (i.e. assuming additivity of the attributes, given that the GM seam is
spherical structure. well behaved), but the reader is referred to Bertoli et al. (2003) for
more information on accumulation approaches.
1.2.1. Additivity of case study data However, certain types of variables may behave non-additivity
Geostatistical methods typically assume that the variables of interest even if measured under uniform support or following an accumulation
behave additively (i.e. that the mean calculated through a simple linear approach to dealing with variable supports. Coward et al. (2009), pro-
average is unbiased). Sample assays represent a concentration of an pose a framework within which variables may be classified as either
attribute per unit of mass. In coal, attributes are generally measured ‘primary’ or ‘response’ depending upon the degree to which the variable
over varying seam thicknesses, and therefore sample determinations reflects an intrinsic attribute of the rock (‘primary’) or its response to
represent variable masses. As such some attributes of interest may be measurement process (‘response’). The authors argue that response
strictly non-additive and an accumulation type approach may be pref- variables tend to behave non-additively whereas primary variables
erable. The authors consider direct approach is acceptable in this case can usually be considered as additive. Under this framework, Thickness
118 A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124

Fig. 3. Base maps for Thickness and Yield within the MAIN domain and Sulphur within LOWTS. The test area which represents 8 longwall panels is also shown as north-south ori-
entated blocks.

is clearly a primary variable; total Sulphur is a primary variable (although a 2D block of given dimensions and shape. The variogram model
washed Sulphur would be a response variable); Yield, however, is a characterises variance as a function of separation distance (strictly,
‘response’ variable. For the purpose of this paper Yield was treated as vector). In a one sample, one block situation, it is therefore possible to
additive, but for further information on testing additivity of response calculate the expected error variance (or the expected squared error)
variables, and in particular Yield, the reader is referred to Carrasco et al. incurred when we attempt to estimate the block with a sample. This
(2008). result, applying to a single ‘typical’ block, can then be used to infer the
global error variance for a given drillhole spacing (assuming that all
2. Explanation and implementation of methods the blocks have a central hole, i.e. regular coverage of drilling and a rect-
angular grid pattern centred on the blocks). The method used to obtain
2.1. Global estimation variance global estimation variance from the single block estimation variance is
summarised with an example by David (1977).
The GEV approach uses the variogram model to make an assess- The GEV method accounts for the geometry of the sample data and
ment of how precisely a single sample can be expected to estimate the spatial continuity of the attribute of interest but does not account

Fig. 4. Histograms for Yield, Thickness and Sulphur.


A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124 119

Table 1 [G] Plot this relative precision (expressed as a percentage) vs. the
Statistics for Yield, Thickness and Sulphur within their entire domains and within the dimensions of the sampling grid defined in [A] (which can be
test area.
correlated to a drilling cost for the area).
Area Variable Count Min Max Mean SD CV Skew [H] Repeat for a different grid spacing (which will also imply dif-
Entire domain Thickness 348 4.45 5.95 5.28 0.27 0.051 −0.573 ferent N).
YLD8.5 79 66.40 92.50 83.31 4.59 0.055 −0.901
RTS 96 0.34 0.77 0.45 0.04 0.095 2.978 2.2. Local confidence intervals via the discrete Gaussian model
Test area Thickness 45 4.92 5.95 5.40 0.25 0.046 0.472
YLD8.5 15 66.40 89.00 83.04 5.63 0.068 −1.509
RTS 16 0.41 0.50 0.44 0.02 0.048 1.179 Kriging is an interpolation method that provides the estimated point
or block value (generally block values for resource estimation applica-
tions) and also gives an indication of the local precision of that estimate
for skewness or the proportional effect. The key assumptions of the via the kriging variance (David, 1977). However the kriging variance ac-
method are, firstly, that the domain is relatively statistically homoge- counts only for spatial variability (via the variogram model) and the
neous; i.e., that there are no major trends in the data within a domain geometry of the neighbouring sample data. It is independent of the
and that the variogram is a reasonable representation of behaviour of local data values employed in interpolation and its utility to build
the variable throughout the domain. Secondly, homoscedasticity is confidence intervals requires an assumption that the local error distri-
required. Thirdly, it assumes that the domain has been (or will be) bution is symmetric. In the presence of skewed distributions and pro-
systematically sampled on a regular grid and that the domain shape portional effect, the local error distribution will also be likely to be
matches the sample pattern without geometric effects on the mar- skewed and its variance will be related to the local mean via the propor-
gins. The fourth assumption is that errors are uncorrelated between tional effect; the local distribution of uncertainty therefore cannot be
blocks. fully specified by the kriged estimate and kriging standard deviation
The authors were comfortable with the first assumption within the (see Olea et al., 2011). Local confidence intervals derived from these
test area. The second assumption was met for Thickness and Yield but will therefore be erroneous to some degree. The DGM framework
clearly not for Sulphur (see Fig. 5). The third assumption is considered (Journel and Huijbregts, 1978; Vann and Sans, 1995) can be used to
to be approximately met for the three variables and the authors consid- overcome this difficulty.
er that block size (at the current drill spacing) is sufficient to assume A major advantage of the DGM framework is that, if the variable of
that errors are independent. interest conforms to the properties of the multiGaussian distribution
The GEV approach, as implemented and subsequently shown in (see Goovaerts, 1997), the local estimation error distribution is fully
Fig. 12, can be summarised as follows: specified by the mean and variance of the Simple Kriging (SK); see
Journel and Huijbregts (1978) and Goovaerts (1997). Of course, few
[A] Select or assume grid mesh dimensions X (m) × Y (m) — note earth science data sets in their original (i.e., untransformed) units Zα
that, because we have 1 block per sample in this exercise, the would adhere to multiGaussian properties. Therefore, in practice, a
block dimensions are the same as the sampling grid dimen- prior transform is required to transform data Z to Gaussian values
sions. Y; i.e. (Z → Y). A normal score transform (Z → Y): was performed to
[B] Using the appropriate variogram model, calculate the estima- calculate standard N(0,1) Gaussian values denoted as Yn corresponding
tion variance σe2 when a block of size X (m) × Y (m) is estimat- to each raw data value denoted as Zn. Then the relationship between Yn
ed using one central sample. and Zn was fitted using a summation of Hermite polynomials of Yn, as
[C] Calculate the number N of samples (and therefore blocks) at shown in the formula below:
the specified grid spacing required to cover the area of interest.
[D] Calculate the theoretical variance of estimation of the mean for
X

σ2 Z ¼ φðY Þ ¼ Ψi H i ðY Þ:
the variable of interest over the entire area as σ 2E ¼ Ne .
i¼0
[E] Calculate the equivalent standard deviation σE (i.e. square root
of the calculated variance).
[F] As a first approximation, calculate the relative precision, being The normal score transform does not guarantee that all of the
the ratio of the standard deviation in [E] to the global mean of multiGaussian properties are met by the Gaussian equivalents (Yα)
the variable of interest, calculated from the actual samples in but they can at least be tested to a bivariate level; see Goovaerts
the area): σ E =m . (1997). Assuming the properties of the multiGaussian distribution

Fig. 5. Mean versus standard deviation plots for Thickness, Yield and Sulphur; rho is the Pearson linear correlation coefficient.
120 A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124

Fig. 6. Modelled experimental variograms for Thickness (two directions shown: N0 is north, N90 is east), Yield (omnidirectional) and Sulphur (omnidirectional).

are met, a confidence interval on the Gaussian equivalents can be cal- Given this, the Gaussian confidence can be related back to original
culated at unknown locations using the SK mean (YVK) and variance units for blocks.
(σK2): n    o
K K
Pr φr Y V −1:96σ K < φðY V Þ < φr Y V þ 1:96σ K ¼ 95%
n o
K K
Pr Y V −1:96σ K < Y V < Y V þ 1:96σ K ¼ 95%:
In practice the data values are transformed into their Gaussian equiv-
alents and are centred into the blocks to be estimated. An anamorphosis
For points, these confidence intervals can be related back to the orig- model of the transform is fitted and block equivalent calculated. Next the
inal units using the block anamorphosis function introduced above block discretised Gaussian variogram is obtained from the variogram
φ(Y). In order to relate to block values it is necessary to calculate a model in original (Z) units and the block anamorphosis model. A SK
block support correction coefficient r (which is the Pearson correlation estimate and SK standard deviation (σK) is calculated for each block
coefficient between Gaussian points and Gaussian blocks) determined using the centred Gaussian data and the bounds of the 95% confidence
from the variance of blocks to give: intervals are calculated as:
 
K
Z min ¼ φr Y V −1:96σ K
X

i  
Z v ¼ φr ðY v Þ ¼ Ψi r H i ðY v Þ: K
Z max ¼ φr Y V þ 1:96σ K :
i¼0

Fig. 7. 95% confidence intervals for Thickness, Yield and Sulphur using the Discrete Gaussian Model approach; the longwall test area is also shown.
A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124 121

This approach was introduced by Roth and Armstrong (1996). In this study, CS realisations were generated for Thickness, Yield and
Fig. 7 shows the local block 95% confidence intervals for Thickness, Sulphur using the TBM algorithm. Declustering of the drillhole data
Yield and Sulphur derived using the DGM approach. was carried out using cumulative Ordinary Kriging (OK) weights. This
declustering approach is recommended above other alternatives be-
cause it is the only method which guarantees internal consistency with-
2.3. Conditional simulation in the multiGaussian algorithms. The ensemble of Gaussian realisations
will feature zero mean on expectation only if the weighting applied to
CS models (Chilès and Delfiner, 1999; Journel, 1974) consist of each sample during initial declustering is equal to the cumulative
sets of equiprobable, realistically variable, spatial realisations for rel- weighting that sample receives during the conditioning phase of the
evant coal attributes. These realisations are ‘conditional’ in the sense simulation process (which in TBM is by kriging). Given this the ensem-
that they agree with the known but limited drillhole sampling data ble of Gaussian realisations can be back transformed to match the
from the deposit. CS can also be extended to account for spatial corre- declustered sample mean on expectation, thus avoiding the introduction
lations between variables (Conditional Co-Simulation or CCS; Chilès of bias.
and Delfiner, 1999).
Importantly, a set of CS realisations collectively constitutes a versa-
2.3.1. Investigation into minimum number of realisations
tile model of uncertainty which can be interrogated at scales from pro-
In order to investigate the impact of choice of number of realisations
duction periods (e.g., annual, quarterly, monthly or weekly) down to
used to generate confidence intervals, 1000 realisations were generated
individual Selective Mining Units (SMUs). CS is thus a route to inferring
and subsets of 500 and 100 realisations were drawn from these; all
the uncertainty associated with specific volumes of coal and thus defin-
three ensembles of realisations were used to generate 95% confidence
ing estimation variance.
intervals on a block by block basis for Thickness; these are compared
CS is computationally intensive compared to kriging methods; how-
in Fig. 8. There are significant local differences between 95% confidence
ever, as previously mentioned, the data lend themselves to two dimen-
intervals derived from 100 and 500 realisations, but no global bias ex-
sional projections in this case (as is also the case for other coal seams
ists. There is little difference (evident as scatter around the 1:1 line on
where there is relatively limited post-depositional folding and faulting).
the charts) between the 95% confidence intervals derived from 500
This means that the generation of simulations was very rapid for this
and 1000 realisations, in this case therefore a set of 500 realisations
case study.
can be considered from a practical viewpoint to adequately characterise
There are numerous approaches to CS now available. The two
the full probability distribution, but 100 realisations arguably cannot.
main approaches used in the mining industry are the Turning Bands
For brevity we have shown Thickness results only; a sequence of poten-
Method (TBM; Journel, 1974) and Sequential Gaussian Simulation (SGS;
tial tests to validate a platform of conditional simulation realisations is
Deutsch and Journel, 1992). In our experience these methods, in their
provided by Emery (2008).
modern implementations, produce very similar outputs in most situa-
tions. The key commonality is that they both rely upon a multiGaussian
model and a spatial distribution of attributes which can be considered 3. Comparison between methods
to reasonably approximate the diffusion model (see Riviorard, 1994;
Vann et al., 2000). In this study we produced the simulations using TBM. 3.1. Local block-by-block comparison
It should also be noted that there is always uncertainty on the pa-
rameters and moments used to build the realisations of a CS model The GEV approach provides global results and therefore cannot be
(for example, uncertainty on the variogram models and histograms compared locally against the DGM approach or the CS approach. DGM
of various attributes). In a comprehensive approach to model uncer- and CS, however, can be compared both globally with GEV and locally
tainty of mineralisation, such ‘parameter uncertainty’ should be con- with each other. In order to make localised comparisons each of the
sidered. However, in this simple study we have used ‘best estimates’ 125 m × 250 m blocks in the test area has been processed into a one
of the various parameters required. dimensional time series string which can be considered to represent
As with the DGM approach, confidence intervals derived from CS a longwall mining sequence across part of the study area (which is
are local and account for spatial variability, skewness and any propor- shown in Figs. 3 and 7). The sequence assumes that the eight longwalls
tional effects which are associated with the variables of interest. (each representing approximately one year of production) are mined

Fig. 8. Comparison of 95% confidence intervals for Thickness derived from 100, 500 and 1000 conditional simulations.
122 A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124

Fig. 9. Comparison of DGM and conditional simulation 95% confidence intervals on Yield over an 8 year mining sequence on a block-by-block basis.

from west to east over an 8 year period with the direction of mining alter- stationarity conditions. DGM, however, requires SK and is consequently
nating between years. In Figs. 9 to 11 the block-by-block 95% confidence more sensitive to departures from stationarity.
intervals throughout the sequence derived from conditional simulation
are shown as a grey band, and the minimum and maximum of a 95% con- 3.2. 8 year production period comparison
fidence interval derived from DGM approach are shown as a dashed black
line series respectively. The comparison between the two methods is The GEV of the three variables at two standard deviations is shown
shown for Thickness in Fig. 9, Yield in Fig. 10 and Sulphur in Fig. 11. for drill spacings from 50 m × 50 m to 1000 m × 875 m over the full
In general, at the local level, the comparison is relatively close. For 8 year production period (representing 4.75 million m 2, as shown
Thickness the minimum of the DGM 95% confidence interval is gener- in Figs. 3 and 7) in Fig. 12. In addition each of the CS realisations
ally marginally lower than that derived from conditional simulation were averaged into the same area; the E-type estimate of the
but is otherwise similar. For Yield the DGM confidence interval mini- realisations (the block-by-block mean of the realisations) was also
mum and maximum are significantly lower than conditional simula- calculated over this area and the difference between each of the
tion in year 8 but otherwise match relatively closely. realisations and the E-type calculated. Intervals of 2 standard devia-
General similarity between these two approaches is expected, as tions of the resulting residuals can express confidence intervals, as
are some potential theoretical and practical differences. Both methods are shown in Fig. 12. These are plotted at the equivalent regular
account for sample data geometry, spatial variability, skewness and drill spacing (calculated as number of holes/total area) equating to
heteroscedasticity and both involve broadly similar multiGaussian the conditioning data. The implied confidence interval is labelled on
assumptions. However the centring of sample data which is required Fig. 12 as the “global standardised estimation precision”:
as part of the DGM approach can result some divergence from the
conditional simulation result. Additional discrepancy between the two
approaches is likely as a result of OK being used in the conditioning .
2σ E
 :
of the CS realisations; this allows for some local departure from m

Fig. 10. Comparison of DGM and conditional simulation 95% confidence intervals on Thickness over an 8 year mining sequence on a block-by-block basis.
A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124 123

Fig. 11. Comparison of DGM and CS 95% confidence intervals on Sulphur over an 8 year mining sequence on a block-by-block basis.

Fig. 12 indicates that for Yield and Thickness, where the assump- over-estimates the estimation variance compared to that derived from
tions behind the GEV method are reasonably met, GEV provides a CS. This is undoubtedly due to the proportional effect associated with
very reasonable approximation of the estimation error provided by Sulphur (the test area is relatively low grade for Sulphur, see Fig. 3).
full CS. Crozel and David (1985) presented an alternative GEV approach The authors note that the observed overestimation of uncertainty by
using combined kriged error variances of block estimates for both 2D the GEV method is not always guaranteed and that in some circum-
and 3D situations. In a subsequent commentary on this work, Buxton stances an underestimation may occur; the key point being that if a
(1986) cites his doctoral research (Buxton, 1985) and makes conclu- strong proportional effect exists the utility of results generated by GEV
sions similar those made here that “…the classical method of composit- methods may be compromised.
ing extension variances is a perfectly adequate” in many applications
(Buxton, 1986, p694). The classical method referred to by Buxton is ex- 4. Conclusions
actly the GEV approach as used in this paper.
Probably no easier way exists to calculate global estimation variance. This study indicates that prior testing of the assumptions associat-
For Sulphur, where the required GEV assumptions are not met (in par- ed with each method is critical. Where the required assumptions of
ticular the assumption of homoscedasticity), then GEV significantly the GEV method are met it provides a reasonable approximation of

Fig. 12. Comparison of global standardised estimation precision derived from global estimation variance (GEV) approach and conditional simulation (CS) for the three variables of
interest over an 8 year production area.
124 A. Cornah et al. / International Journal of Coal Geology 112 (2013) 114–124

full CS, even in a semi-local setting (e.g., for subsets of the coal seam Coward, S., Vann, J., Dunham, S., Stewart, M., 2009. The primary-response framework
for geometallurgical variables. Proceedings 7th International Mining Geology Confer-
approximating annual production intervals). Since GEV is conceptually ence. The Australian Institute of Mining and Metallurgy, Melbourne, pp. 109–113.
simple and expeditious, its use should not be excluded in situations Crozel, D., David, M., 1985. Global estimation variance: formulas and calculation.
where simple geological seam geometry that can be approximated rea- Mathematical Geology 17, 785–796.
David, M., 1977. Geostatistical Ore Reserve Estimation. Elsevier Scientific Publishing
sonably in two dimensions and the assumptions and limitations are Company, Amsterdam.
likely to be acceptable. The key limitations are that it does not perform Deutsch, C.V., Journel, A.G., 1992. GSLIB: Geostatistical Software Library and User's
well in the presence of high skewness and proportional effect and that it Guide. Oxford University Press, New York.
Diessel, C.F.K., 1999. An Assessment of the Zone of Coal Deterioration at Moranbah North
only quantifies uncertainty associated with (at best) the semi-locally Mine Based on Petrographic Analyses. Consultant Report Prepared for Anglo Coal
estimated mean value. Australia Pty Ltd.
DGM and CS quantify uncertainty associated with local estimations of Emery, X., 2008. Statistical tests for validating geostatistical simulation algorithms.
Computers and Geosciences 34, 1610–1620.
the mean. The DGM approach represents a computational saving over CS
Faliven, O., Cabrera, L., Sáez, A., 2007. Optimum and robust 3D facies interpolation strat-
and is applicable to large scale deposits which are modelled in three egies in a heterogeneous coal zone (Tertiary As Pontes basin, NW Spain). Internation-
dimensions; similar key assumptions underlie both methods. The CS ap- al Journal of Coal Geology 71, 185–208.
proach has the most flexibility and also provides a platform for address- Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation (Applied Geostatistics
Series). Oxford University Press, New York.
ing many other issues related to coal quality variability, drill spacing, and Heriawan, M.N., Koike, K., 2008a. Identifying spatial heterogeneity of coal resource
blending for example. The price to pay is increased time requirements; quality in a multilayer coal deposit by multivariate geostatistics. International
this relates not only to computational time (which in this case was not Journal of Coal Geology 73, 307–330.
Heriawan, M.N., Koike, K., 2008b. Uncertainty assessment of coal tonnage by spatial
significant), but also the manipulation and checking of CS models can modeling of seam distribution and coal quality. International Journal of Coal
be time consuming for the professionals involved. The DGM approach Geology 76, 217–226.
should be considered as a viable alternative to CS where the full granular- Journel, A., 1974. Geostatistics for the conditional simulation of ore bodies. Economic
Geology 69, 673–688.
ity and flexibility of the latter is not required. Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. Academic Press, London.
Karacan, C.Ö., Olea, R.A., Goodman, G., 2012. Geostatistical modeling of the gas emission
ozone and its in-place gas content for Pittsburgh-seam mines using sequential Gauss-
Acknowledgements ian simulation. International Journal of Coal Geology 90–91, 50–71.
Laws, A.J., Bos, F., 2000. Moranbah North and Grosvenor geological review (6 Volumes).
Shell Coal Unpublished Company Report Dated June 2000. Anglo Coal ref: CR3448
The authors wish to acknowledge the staff of Anglo American Met-
MN/GV (01).
allurgical Coal for their permission to conduct this work and publish Michaelsen, P., Henderson, R.A., 2000. Facies relationships and cyclicity of high-
the results. The original study referred to in the text was conducted latitude, late Permian coal measures, Bowen Basin, Australia. International Journal
by a team consisting of two of the current authors (John Vann and of Coal Geology 44, 19–48.
Morgan, M.G., Henrion, M., 1990. Uncertainty: A Guide to Dealing with Uncertainty in
Ian Driver) along with Olivier Bertoli (then of Quantitative Group, Quantitative Risk and Policy Analysis. Cambridge University Press, Cambridge.
now at Geovariances), see Vann et al. (2004). The authors would Olea, R.A., Luppens, J.A., Tewalt, S.J., 2011. Methodology for quantifying uncertainty in
also like to thank Michael Harley of Anglo American and Mike Stewart coal assessments with an application to a Texas lignite deposit. International
Journal of Coal Geology 85, 78–90.
of Quantitative Group and the two anonymous reviewers; their con- Riviorard, J., 1994. Introduction to Disjunctive Kriging and Non-Linear Geostatistics.
tributions improved the paper. Any remaining deficiencies are of Clarendon Press, Oxford.
course entirely the responsibility of the authors. Roth, C., Armstrong, M., 1996. Confidence intervals for local estimation: application to
the Witwatersrand basin. In: Ramani, R.V. (Ed.), 26th Proceedings of the Applica-
tion of Computers and Operations Research in the Mineral Industry, APCOM'96.
Society for Mining, Metallurgy, and Exploration, USA.
References Vann, J., Sans, H., 1995. Global resource estimation and change of support at the Enter-
prise gold mine, Pine Creek, Northern Territory: application of the geostatistical
Bertoli, O., Job, M., Vann, J., Dunham, S., 2003. Two-dimensional geostatistical methods —
Discrete Gaussian model. Proceedings of APCOM XXV Conference, Brisbane,
theory, practice and a case study from the 1A shoot nickel deposit, Leinster, Western
Queensland, 9-14 July, 1995. The Australasian Institute of Mining and Metallurgy,
Australia. Proceedings 5th International Mining Geology Conference. The Australian
Melbourne, pp. 171–180.
Institute of Mining and Metallurgy, Melbourne, pp. 189–195.
Vann, J., Guibal, D., Harley, M., 2000. Multiple indicator kriging — is it suited to my deposit?
Buxton, B., 1985. Geostatistical determination of the precision of global recoverable
Proceedings of the 4th International Mining Geology Conference, Coolum, Queensland.
reserve estimates: Ph.D. thesis, Department of Applied Earth Sciences, Stanford
The Australasian Institute of Mining and Metallurgy, Parkville, pp. 171–179.
University.
Vann, J., Bertoli, O., Driver, I., 2004. Geostatistical resource management: Assessment of
Buxton, B., 1986. Letter to the Editor: global estimation variance. Mathematical Geology
clean coal quality and classification guidelines based on a Moranbah North case
18, 693–696.
study. Anglo Coal Australia Pty.Ltd., Unpublished Report Dated April 2004.
Carrasco, P., Chilès, J.P., Segurét, S., 2008. Additivity, metallurgical recovery and grade.
Wackernagel, H., 2003. Multivariate Geostatistics: An Introduction with Applications,
Proceedings Eight International Geostatistical Congress, pp. 237–246.
3rd edition. Springer, Berlin.
Chilès, J., Delfiner, P., 1999. Geostatistics — Modelling Spatial Uncertainty. John Wiley
and Sons, New Jersey.

Das könnte Ihnen auch gefallen