As presented to the CIM Thunder Bay Branch, by Alan

Aubut, February 9, 2011 (with some minor additions)

Presentation is divided into 4 parts:

1. Introduction to Geostatistics
2. The Resource Estimation Process
3. Resource Classification
4. Conclusions
5. Bibliography
- Geostatistics is the analysis and prediction of
spatial or temporal phenomena, such as metal
grades.
- It is used to analyze and predict values of a
variable distributed in space that are
implicitly assumed to be correlated with each
other.
- Predictions at unsampled locations are made
using "kriging" or they can be simulated using
"conditional simulations.”
- Traditional estimation methods (ID, Polygonal,
Etc.) are biased:
- Weights assigned to each sample are strictly
arbitrary.
- Geostatistics determines weights based on the
correlation of the samples themselves
(determined using the Variogram).
- Kriging strives to reduce the error.
- BLUE : Best Linear Unbiased Estimator
Stationarity: the mean and variance of values do not
depend on location.
The Mean is assumed to be constant and
independent of location.
The variogram is assumed to exist and the
increments depend only on the difference
between any two points (sample locations).
You must choose a “Region of Stationarity”
- samples that are related to one another (a single
population)
 The foundation of Geostatistics.
 It is a two-point statistic that is the mean squared
difference for samples at a specific separation (h),
or “lag”.
1
𝛾 𝐡 = [𝑧 𝐮 − 𝑧(𝐮 + 𝐡)]2
2𝑁(𝐡)
𝑁(𝐡)
 Function of distance (h) and direction.
 Quantifies the variability between two random
variables separated by a distance vector h (the
lag).
 It is used to model the way pairs of samples in
space are correlated.
x x h=10’ x+h
x+h
h=5’

Plan View

n
Variogram=moment of inertia= 1 (xi - (xi +h))2
i=1
2 e
l in
+ 1 :1
Ni Grade at point (x +h)

+
+
+ + +
+ + +
+
++ +
+ + +
+

Ni Grade at point (x)

Each point is the average of all pairs for a particular “lag”.

 n
Variogram=moment of inertia= 1 (xi - (xi +h))2
i=1
2 e
1 lin
+ 1:
Ni Grade at point (x +h)

+
+
+ + +
+ + +
+
++ +
+ + +
+

Ni Grade at point (x)

H-scattergram for different lags”.

 Sill = Variance of the sample values

Variogram Model

Spatial Variance Range = Distance where maximum

variability between samples
is reached (ie the sill).
n
Variogram=moment of inertia= 1 (xi - (xi +h))2
i=1
2 e
in
1l
+ 1:
Ni Grade at point (x +h)

+
+
+ + +
+ + +
+
++ +
+ + +
+
Nugget =
Ni Grade at point (x)

analytical
error + natural
variability
Distance between Samples
A geostatistical estimation technique that uses a
linear combination of surrounding sampled
values to make a prediction.
Kriging allows you to derive weights from the
variogram that result in optimal and unbiased
estimates.
Weights are applied to each surrounding sampled
data.
Kriging attempts to: (a) minimize the error variance;
and (b) systematically set the mean of the
prediction errors to zero, so that there are no
over- or under-estimates.
Ordinary Kriging - the sum of the weights add up to one.
Simple Kriging – the mean and the variance of the system
are known
Indicator Kriging – the probability that a value is 0 or 1.
Multiple Indicator Kriging (MIK) – uses Indicator
Kriging to estimate the probability that a value is above a
cut-off, or “indicator”. The combined probabilities can
then be combined to get a point estimate. Used for
highly skewed distributions.
Universal Kriging – estimating data that has a regional
trend.
Resource Estimation- least biased method.
Data Interpolation– can be used with many data set that
are spatially of temporally distributed
- Must honour “Stationarity”
Conditional Simulation– best way to analyze “Risk”.
Used extensively in the Petroleum industry
Can be and should be used more in the Minerals
industry
Kriging (and all other weighted averaging methods)
over-smooths the values
Smoothing results in the blocks having a variance
lower than the true variance
Conditional Simulation builds many equally probable
densely informed point realizations.
 Each realization reproduces the input sample
histogram and the variogram
 Each realization is different because of the
uncertainty away from known points (thus
simulations and not estimates)
1. Turning Bands
- Simple method that does not require data to be on a regular
grid
- Requires two passes; non-conditional simulation and then
conditioning
- Requires Gaussian transformation of the data
2. Sequential Gaussian
- Requires a Gaussian transformation of the data
- Data needs to be on a regular grid
- Simulation is done an a random path through the grid utilizing
both hard and soft data at the same time
- Very fast computationally
3. Sequential Indicator
- Does not require a Gaussian transformation
- Best for categorical variables (i.e. Rock Type)
 Reproduces the Variability of the input data
HISTOGRAM

Reproduces the Spatial Continuity of the input data

VARIOGRAM

Measures the likelihood of the desired outcome (risk)

Recognizes that many equally-likely models of reality exist

Mean = 0
Variance = 1
Flowchart of the various stages of construction of a simulation
matching the histogram, the variogram and the data.
Summary of the conditioning process for the turning bands
method (after David, 1977, p. 327). From Olea, 1999, p. 162.
Sim10 Sim14

Variable grade dispersion

NN Kriged

Smoothed estimates
One Simulation Realization

Average of all the Simulations

Simulations preserves natural variability

 Conditional Simulations allow the spatial grade variability to be properly
characterized (assuming sample data histogram, variogram and domain size,
shape and location are representative).

The results from Conditional Simulations can be used to determine the

range of possible resources and reserves as defined by various levels of
mining selectivity.

Conditional Simulations allow practitioners to quantify the uncertainty

associated with any given reserve estimate and to perform better sensitivity
analyses to estimate the risk of not achieving the project’s financial targets.

Conditional Simulation provides an additional tool in mineral assessments

to enable key decision makers to better understand the risk associated with
projects.
1. Assemble drill hole sample data
2. Validate
3. Create Mineral Envelope
4. Capture and composite samples
5. Create unfold Strings
6. Unfold data
7. Variography
8. Create NN and OK Models
9. Apply Smoothing Correction
10. Report Resources
Validate Drill hole Data
Use Histograms
Use Scatter plots
Evaluate sections and plans
Check for Co-ordinate system errors
Check for Azimuth problems
Check for Elevation Problems
Check for sample Overlaps
 Dispersion - The degree of scatter of
mode data, usually about an average value

stdv

Skewness - the degree of

asymmetry of a distribution around its
mean

Median Mean Element of Interest (%)

(50%) (average)
Normal Distribution
– Symmetric
– Unimodal

Skewed Distribution
– Asymmetric
– can be positively skewed or
negatively skewed
Kurtosis - the degree of “peakedness” of
a distribution around its mode

arithmetic
mean 1m samples
5m samples

10m samples

Au (oz/t)
Bi-modal Distribution – There are two distinct
populations.
- Likely due to presence of multiple domains.
- Cannot assume “Stationarity”
- If possible need to separate the two populations.
 Ni = a * Cu + b

outlier

a
b
Ni (%)
Geostatistics is the study of spatial data.
Errors in the data result in errors in the analysis.
A single aberrant datum can have a significant
effect on the variogram.
Histograms and correlation plots can be used to
systematically evaluate the data prior to doing variography.
What was sample recovery and did it vary?
Are the samples consistently sampled and analyzed over
time and space?
Are the sample sizes (especially diameter) consist over
time and space?
Model the mineral system (define the Stationarity
System)
Do not model economic cut-offs
Snap interpretation lines to boreholes
Interpretation must be in 3 Dimensions
Need to be “snug” to mineralisation
Respect different mineralisation styles
Separate into different domains if need be
Must be Samples that are related to one another
Each domain is modelled separately
37000 Section 37250 Section
Use mineral domain wire frame to capture samples
Assess samples for missing assays and Density or Specific
Gravity (SG)
Use polynomial regression to calculate Density/SG
Composite to uniform sample length based on mean
sample length and intended block size.
All calculations from this point onward treat all data
as point data.
Each point represents the same volume, or “support”,
otherwise a bias is introduced.
You need to “Unfold” if:
1. There is a change in dip
2. There is a change in strike and dip
3. There are embayment structures
4. There are pinch and swell structures
5. There is folding of the mineralisation
• Needed because variogram is best
determined with a flat data set.
Use of Unfolding simplifies Domaining
 variable dip
“pinch and swell” structure

Used to deal with various geological

problems
Geological Distance
Cartesian Distance
 • This unfolding procedure resulted in the creation of a new 2
1 set of co-ordinates (UCSA [across dip direction], UCSB [down
dip direction], and UCSC [along strike direction]) for each
sample and each discretization point of the deposit

A – Across
B – parallel to unfold
strings
C – Orthogonal to B
3
The variogram
1. Independent of location and depends
only on distance between sample pairs
2. Measure of similarity of samples as a
function of distance
3. Foundation of geostatistics
4. Kriging relies on the variogram
• The variogram determines what
weight will be applied to each sample.
• All weights for each estimate must
sum to 1.
The variogram
“Search Cone”
and its
components.
Lag – distance
between
samples
 Variogram Model
C2
C1
“Experimental Variogram”

C0
Distance
a1 a2
Also known as “Structural Analysis”
Involves constructing a model of the experimental
semi-variogram.
Requires knowledge of the phenomenon being
modelled (i.e. grade)
Also requires good “craft” based on experience and/or
advice of those with experience.
The purpose of the variogram is to condense the main
structural features of the phenomenon being studied
into an operational form.
 Four primary models used and about a dozen available
Each is defined by a mathematical formula.
Spherical Model Exponential Model

Gaussian Model
Power Model

a – the “Range”
h – the “Lag”
c – the “sill” of the structure
Variations between samples are due to many causes
over a range of scales including:
Sample support due to variability in measurements.
Petrographic level due to variations between
minerals.
The mineralised horizon due to the variation
between mineralised material and waste.
All come into play simultaneously
Result in needing to “nest” models to represent all of
these structures – additively combine multiple models.
Isotropic – no difference with direction
Anisotropy – differs with direction
Geometric Anisotropy – all rise to the same sill
but each primary axis has a different range.
Typical in mineralised systems.
Zonal Anisotropy – there are at least two sills,
depending on direction.
Usually associated with multiple bedded
mineral horizons
The Dimensions are derived from the Variogram Ranges.
The data used to calculate the semi-variogram must
be of the same support (sample dimensions)
Core samples should not be mixed with samples of
another type (i.e. RC)
Hole sizes should be similar.
Sample lengths should be similar.
If not, composite to a uniform sample length
Use ranges determined from modelling to establish how
far we need to look for samples before we estimate block.
Try and eliminate bias through clustering by applying
sample selection restrictions.
Typically use an “Octant Search”
Search ellipsoid is divided into octants.
Minimum and maximum number of samples per
octant.
Minimum number of octants
To ensure all blocks are estimated use 3 nested searches
Parameters are less restrictive than for the previous
search
 Block Model Definition
- The size of the blocks in the model will best match
mining selectivity, sampling density and anticipated
grade control method (dynamic process if drilling
density is changed or a different mining method is
considered)
- SMU (Smallest Mining Unit)
- Selectivity will be a function of the
equipment/mining method that possibly will be
used
- Allows inclusion of internal waste and some
external waste
- Resultant model will be “diluted”
- Closer to mining reality
Model first using Nearest Neighbour estimation
Provides declusterised global statistics
Global Mean should be the same no matter what
method is used thus use NN Mean as check
(Stationarity).
Model using Kriging
Used because it reduces error to a minimum
Least biased
Was the grade distributed in the NN model in a
reasonable fashion?
Is the mean of the OK Model essentially the same as
the mean for the NN Model?
Was the grade distributed in the OK model in a
reasonable fashion?
Do both models honour the drill holes in a
reasonable fashion, allowing for “change of support”
smoothing?
NN Model OK Model
Kriging tendencies:
 Underestimates values above the mean
 Overestimates values below the mean
This distortion is called Conditional Bias
Alternative name is “Smoothing”
For skewed distributions it results in not
enough high grade samples (the upper tail is
fore-shortened)
Variance is lower than expected.
Evaluate using following relationship:
Sample Variance in deposit (Vd) = Variance between blocks (VB) + Sample variance in Block (Vb)
 Regression line
Zv*
(no bias)

waste
Estimated Grade

processed

m*

Ore
dumped

Zc m Zv

Actual Grade
Two methods commonly used to correct for Smoothing
1. Affine
A direct variance correction
All secondary modes are reproduced
Can result in negative values
2. Log Normal Shortcut
For highly skewed (~ Log Normal) distributions
Both restore variance to expected levels
Classification of Resources is wholly dependant of
Confidence in the estimate
Two primary factors to consider
1. During which search was the block
estimated?
 Most restrictive criteria result in highest
confidence in resulting estimate.
2. What confidence do we have in the Variogram
which was used to determine our primary
search criteria (the ranges) and from which
the Kriging Weights were derived?
A Mineral Resource is a concentration or occurrence of
diamonds, natural solid inorganic material, or
natural solid fossilized organic material including
base and precious metals, coal, and industrial
minerals in or on the Earth’s crust in such form and
quantity and of such a grade or quality that it has
reasonable prospects for economic extraction. The
location, quantity, grade, geological characteristics
and continuity of a Mineral Resource are known,
estimated or interpreted from specific geological
evidence and knowledge.
Some practitioners interpret “reasonable prospects for
economic extraction” to mean that mineability
must be demonstrated.
 a mining method must be selected
 a “transfer function” needs to be applied.
• Use “Whittle” to demonstrate mining using open pit
method
• Use “Floating Stope” to demonstrate mining by
underground methods.
Reasonable – sensible
Prospects - Something expected; a possibility.
 Application of a Transfer Function to select what will be
called Resources is too restrictive.
1. No feasibility study has been made to determine the
economic viability of a mining method.
2. The total resources initially in-place is constant.
 Not the case if filtered using a transfer
function.
A

D C A – mineral domain
B – Inferred Resources
C – Indicated Resources
D – Measured Resources
B+C+D = “the Resource”
 Exploration Results

Mineral Resources Mineral Reserves

Increasing level of Inferred Possible

geological
knowledge and Indicated Plausible Probable
confidence
Measured Presumed Proven

Increasing Probability of economic viability

Consideration of
mining, metallurgical,
Consideration of only economic, marketing,
mining and economic legal, environmental,
factors (the "Transfer social and
Function") governmental factors
(the "Modifying
Factors)
[Link]
Geostatistics uses the data from the mineral domain to
determine the weights to use when estimating.
 The Variogram
Estimation is done using Kriging
 Reduces error to a minimum
 If done properly it is the least biased estimation
method.
Resource Estimation requires an integrated approach
with emphasis on reducing bias at all stages.
A smoothing correction needs to be applied to
compensate for too much averaging (high grade will
be under reported).
Classification of resources is dependant on confidence
in the estimate which in turn is dependant on
confidence in the Variogram.
Currently there is a tendency to under-report resources
by application of a transfer function.
In part due to current classification scheme
UN sponsored classification scheme that is more
robust has been developed and should be adopted.
Some good references:
Sinclair, A.J. and Blackwell, G.H. (2006) Applied Mineral Inventory
Estimation; Cambridge University Press, New York, 381p.
Leuangthong, O., Khan, K.D., and Deutsch, C.V. (2008) Solved Problems
in Geostatistics; John Wiley and Sons, New Jersey, 207.
Isaaks, E.H. and Srivastava, R.M. (1989) Applied Geostatistics; Oxford
University Press, New York, 561p.
Deutsch, C.V. and Journel, A.G. (1998) GSLIB Geostatistical Software
Library and User's Guide; Oxford University Press, New York, 369p.
Journel, A.G. and Huijbreghts, C.J. (1991) Mining Geostatistics; The
Blackburn Press, New Jersey, 600p.
Goovaerts, P. (1997) Geostatistics for Natural Resources Evaluation;
Oxford University Press, New York, 483p.