RECOVERABLE RESOURCE

ESTIMATION
- NON LINEAR METHODS
Recoverable resources
Concepts and applications

2
In situ vs. Recoverable resources
In situ: the whole metal existing within the deposit

Recoverable: takes into account, cutoff and selectivity (SMU)

Total Tonnage : 200 Mt Recoverable Tonnage : 90 Mt

Total Metal : 5 Mt Recoverable Metal : 3.5 Mt

(cutoff : 3%, SMU 5x5x3m)
Page 3
 From recoverable resources… to production

Pre-production
1. Exploration drilling
2. In situ estimation

Page 4
 From recoverable resources… to production

Pre-production
1. Exploration drilling
2. In situ estimation
3. Recoverable resources estimation
4. Cut-off application

Page 5
 From recoverable resources… to production

Pre-production Production
1. Exploration drilling 1. Grade control – Tonnage check
2. In situ estimation 2. Excavation
3. Recoverable resources estimation 3. Ore processing – Metal content check
4. Cut-off application 4. Reconciliation – Model update
5. Mine planning – feasibility study

Page 6
In real life
1. Grade control : delineating
the rich ore precisely
2. Blasting the bench
3. Ore/Waste is recovered by
engines (truck, shovel…),
their selectivity defines the
SMU (Selective Mining Unit).

Page 7
Recoverable resources
After the grade estimation, the richest SMUs are
considered to be recovered.

Ex: Zc = 500 ppm

Page 8
Recoverable resources
It leads to define new variables: image resolution smu size
T : tonnage estimated above the cut-off

Ex: T = 10 822 t

Page 9
Recoverable resources

Q: Metal quantity above the cut-off

Ex: Q = 7.58 t

Page 10
Recoverable resources
Additional variables are directly derived from T and Q:

M: Mean grade above the cut-off

=
Ex: M = 652 ppm

B: Conventional benefit

− = − . Ex: B = 1.71 t

Conv. benefit: Minimum metal content:

Proportional to Proportional to the
the profit production cost

Page 11
Grade Tonnage Curves

Zc= 0 ppm Zc= 200 ppm Zc= 400 ppm Zc= 600 ppm Zc= 800 ppm Zc= 1000 ppm

Page 12
Grade Tonnage Curves

Zc= 0 ppm Zc= 200 ppm Zc= 400 ppm Zc= 600 ppm Zc= 800 ppm Zc= 1000 ppm

Page 13
Q vs. T
l Note that the graphic Q vs. T also contains
information about the cutoff zc and M:

Page 14
Recoverable resources

= 0,10

= 0,85
,
From the variables
histograms, we deduce
, , , , ,
,& .

Page 15
Grade Tonnage Curves
l Two secondary variables are considered:
- Black: no correlation
- Blue: negative correlation

l The cut-off is applied on the main variable (orange).

Page 16
Grade Tonnage Curves

Zc= 0 ppm Zc= 200 ppm Zc= 400 ppm Zc= 600 ppm Zc= 800 ppm Zc= 1000 ppm

Page 17
Conclusion
l Recoverable resources are the available quantity of
metal given by operational parameters :
- Cutoff;
- Block size (SMU); and,
- Grade control density.

l The estimation of the recoverable resources is strongly

linked to an accurate estimation of the SMU distribution.

Page 18
Kriging Limitations

19
Kriging limitations: an illustration

This illustration uses the

“Walker Lake” data set, where
the true grades from a 2D
‘deposit’ are known.

20
Strategy
We want to test if it is possible to estimate recoverable resources from Kriging:
Reality

Different
Recoverable
Resources ??

‘Exploration’
Sampling

Kriging

21
Estimation
The estimation is performed on 5x5 blocks, using the
following variogram:

22
 Estimation
Estimated grades on 5x5 Real grades on 5x5
blocks from samples blocks

23
Recoverable resources at 500ppm

Wrong estimation of
Recoverable Resources!

T_est = 289 kt T_real = 521 kt

Q_est = 183.00 t Q_real = 350.73 t
M_est = 633.22ppm M_real = 673.20 ppm 24
Consequence on a real process

If the selected blocks were recovered,

we would lose some rich blocks and
process some poor blocks.

No proper short term mine

T_est = 289 kt T_recovered = 289 kt planning is possible based on
Q_est = 183.00 t Q_recovered = 195.60 t Kriged exploration data!
M_est = 633.22ppm M_recovered = 676.8ppm 25
Grade Tonnage Curves
SMU : 5x5

The Grade Tonnage Curves based on Kriging are far

from the reality.
The resulting recovered quantities are consistent
with Kriging but remain bad predictions.

Kriging is unable to estimate recoverable

resources on SMU based on exploration drillholes.

26
Grade Tonnage Curves
SMU : 20x20

If the SMU is larger, or the drilling tighter, Kriging

proves to be an adequate estimation technique for
the recoverable resources.

Kriging is an adequate method to estimate recoverable

resources on SMU based on grade control.

27
Conclusion

1. Kriging cannot estimate recoverable resources unless the block size

is consistent with the drilling pattern.

2. At exploration stage: Kriging is unable to estimate GTC properly,

except if the continuity is fairly high and the SMU size is close to the
sampling (ex: tabular deposits of Fe, Coal, Mn…).

3. At production stage: Kriging can produce accurate GTC based on

grade control.

4. To estimate properly the GTC from the exploration data, Non-

linear Geostatistics are required.

28
RECOVERABLE ESTIMATION
NON-LINEAR ESTIMATION
Non-linear models
Non-linear models are born from the inability of
Ordinary Kriging to reproduce the SMU histogram
accurately on SMUs and thus to estimate the
recoverable resources.
Idea -> to decompose the grade into independent
factors:
⇔ , ,…
Kriging of independent factors,
and deduction of recoverable
resources (QTM).

Page 30
Non-linear models
Three decompositions exist :
- Indicators (MIK)
-> Mosaic Model: No border effect for
increasing AND decreasing grade.

- Indicator residuals
-> Residual model: No border effect for
either increasing OR decreasing grade.

- Hermite Polynomials (DK / UC)

-> Diffusive model: Border effect for
increasing AND decreasing grade.

Page 31
 Transitions between grades
The cross-variogram of two indicators is a function measuring the
border effect of A on B:
∈ , ∈
= = +ℎ ∈ | ∈ , +ℎ ∉
∈ , ∉

32
 Which model?
The ratio is calculated on
5%
indicators.
It amounts the border effect
when going from one range of
values to another. 4%

Mosaic model :
3%
is constant for increasing
AND decreasing grades.

Residual model: 2%

is constant for either

increasing OR decreasing 1%

grades.
−

Page 33
 Diffusive model
Diffusive model : 650 ppm

increases with h for 550 ppm

increasing AND decreasing
grade.
450 ppm

Indicator variograms: Ranges

are decreasing when the cut- 350 ppm
off increases. High grades are
destructured.
250 ppm

Ratio : The border effect is

stronger when the difference 150 ppm

between cut-offs increases.

Poor and rich areas are
spatially separated.
− 50 ppm

Page 34
Methods for estimating recoverable resources

Non-Gaussian methods Gaussian based methods

• Median Indicator Kriging or • Global Change of

Multi-indicator kriging (MIK) Support (GCOS)
with Affine of Log-normal
• Disjunctive Kriging (DK)
support correction.
• Uniform Conditioning by
• Indicator residuals.
kriged panel grade (UC)

35
Assessment of the spatial variability
Profiles of Estimation on SMUs and Panels for a 50m-range
variogram, and for two different samplings:

100 m 25 m
Variability is smoothed Variability is kept

-> It is not possible to assess the local spatial variability

when data spacing is too large compared to the range
of the variogram.

Page 36
Page 36
 Estimating recoverable resources

Non-linear models assess the SMU

histogram with accuracy, but not the
spatial variability. SMUs can’t be
directly estimated.

At panel scale we can estimate

recoverable resources at the SMU
scale using a support change.

Kriging is applied with a good

quality consistent with drillholes
spacing.

Page 37
 SMU and Panel sizes

l SMU size is not defined by the geostatistician, but by the

mining engineer who requires a size consistent with the
mining equipment.
l Panel size however is linked to geostatistical considerations:
- Consistency with drilling pattern
- Requirement in terms of Kriging quality
- A minimum number of SMU inside one panel to insure the correct
interpolation of the distribution during the localization (eg: 16 is often
considered as a minimum)

Page 38
In Isatis
Workflows of MIK and UC are fully implemented in
Isatis. It is also possible to work with indicators
residuals, although there are no dedicated window.
Each method entails:
- an estimation of panels; and,
- a change of support to obtain the recoverable resources
at the SMU scale.

Page 39
Workflows
MIK UC
Indicator pre-processing:
• Creation of indicators

Variography of indicators Variography of the raw variable

Kriging of indicators on panels Kriging of the raw variable on panels

Support change
- Point anamorphosis
- Block anamorphosis

Indicator Post-processing: Uniform conditioning

• Support change • Recoverable ressources on panels
• Recoverable ressources on panels

Page 40
Support effect
In one panel, an ideal « punctual » selectivity cannot be
reached by the mining operators (shovel, trucks..), only a
coarse « block » selectivity is possible :
Panel

SMU

Page 41
Support effect – Krige’s relationship

Var Z1 > Var Z2 > Var Z3

Page 42
Support effect – Krige’s relationship

, is the
average point
variogram inside
the block.

It measures the
variability absorbed
inside the block.
= 10 = 10

̅, = 6.09 ̅, = 0.67

Page 43
Support effect – Krige’s relationship

, is the
average point
variogram inside
the block.

It measures the
variability absorbed
inside the block.
= 10 = 10

̅, = 6.09 ̅, = 0.67

= 3.91 = 9.33

The support effect is always taken into account through the block variance
calculated by the Krige’s relationship:

= − ̅
( , )

Page 44
Krige’s relationship
ℎ : = − ̅
( , )

The term , is the average point variogram inside the

block. It measures the variability absorbed inside the block. The
variance is reduced of its amount.

1
, = − .
²

Page 45
 Block Variance
ℎ : = − ̅
( , )

For one example block:

Long range: Short range:

̅
( , )→0 ̅, →

→ →

Page 46
 How to calculate ?
̅ , is the average variogram value within one block.
To calculate it, the block is discretized:
To sample every distance, a random secondary
set of 3x3 points is used :

From ℎ we get the

variogram value ℎ .

Ex: discretization 3x3 Each variogram value between a regular point and a
random point is assessed. The average of these values
is ̅ , . For a 3x3 discretization :

, = ∑ ∑ ,

Page 47
Experimental ( , )
How many points of discretization ?

Guidelines:
• For a continuous variogram, a small number of discretization
points is sufficient to converge.
• For an erratic variogram, a higher number is advised.

Page 48
Global Support Change

49
Uniform Conditioning Workflow
l Support change is performed beforehand, the SMU
distribution is defined based on the data distribution.
l This global SMU distribution is then localized on panel
using kriging values for conditioning:

Page 50
Recoverable resources

l The global variables Q,T,M and B can be calculated for

any support size, once we know the histogram of
grades over that support
• The exact location of grades
is not necessary

• The point histogram is only

available, and not the SMU
histogram…

• … we require an estimate of
the SMU distribution

Page 51
 Support change
l Changing the support means to average small support
objects into greater support.
- The mean remains the same
- The variance decreases

Support change from small blocks to bigger blocks.

Page 52
 Support change

Change of support
model

« Point » support: Data Block support: SMU

From Krige’s formula : = − ̅

( , )

Page 53
Discrete gaussian model
Z(x) points Y(x) points r characterizes the relationship
F between gaussian point
variables and gaussian SMU
variables, that is assumed to
be bigaussian.

r
It is called the coefficient of
Z(v) SMUs Y(v) SMUs support change.

Fv The SMU anamorphosis Fv

can be calculated using r and
the point anamorphosis F.

Page 54
Gaussian Anamorphosis
A continuous variable Z could be considered as the
transform of a standard Gaussian variable Y (mean 0,
variance 1), through an anamorphosis function Φ

55
The Gaussian Anamorphosis
Transforming a non Gaussian distribution into a
Gaussian distribution (“anamorphosis” means
“transformation”)
F G
1 1

F( z) = G ( y)

0.5 0.5

0 0
0 10 20 z 30 40 Z -3 -2 -1 0 y 1 2 3 Y

Anamorphosis between an empirical distribution and a normal

distribution (cumulated distribution functions)

56
 Point Anamorphosis

Experimental
Experimental
anamorphosis
anamorphosis
The anamorphosis links the
raw distribution to a
Gaussian one. Raw values: Z

The experimental
anamorphosis is modeled Model
Model
by a continuous function. anamorphosis
anamorphosis

=Φ

= Gaussian values: Y

= =

=
Page 57
Coefficient of support change
The bigaussian relationship between
this pair of gaussian variables can be
described using the following
formula, involving the coefficient of
support change:

= × + 1− × 0,1

=1× + 1−1 × 0,1

=

r = 0.9
1.0

Page 58
Coefficient of support change
The bigaussian relationship between
this pair of gaussian variables can be
described using the following
formula, involving the coefficient of
support change:

= × + 1− × 0,1

=1× + 1−1 × 0,1

=

r = 0.5
0.9
0.7
1.0

Page 59
Coefficient of support change
The bigaussian relationship between
this pair of gaussian variables can be
described using the following
formula, involving the coefficient of
support change:

= × + 1− × 0,1

=1× + 1−1 × 0,1

=

=0× + 1−0 × 0,1

= 0,1

r = 0.0
0.9
0.7
0.5
0.2
1.0

Page 60
Cartier’s relationship
( )= ( )| ( )

Using gaussian anamorphosis:

Φ ( ) = Φ |

From Hermite polynomials’ properties, we deduce :

=Φ = ( )

Where the distribution of SMUs is modelled using the point

Hermite coefficients and the coefficient of support change.

Page 61
 Cartier’s relationship
1
( )= ( )| ( ) =

Using gaussian anamorphosis:

Φ ( ) = Φ |
= ∑ |
= ∑ |

From Hermite polynomials’ property : ( ) | ( ) = ( ( ))

We deduce :
=Φ = ( )

Where the distribution of SMUs is modelled using the point Hermite

coefficients and the coefficient of support change.

Page 62
Range of Distributions
For each r, we can compute the variance of the block distribution
based on the properties of Hermite polynomial:

1 0.75 0.5 0.25 0

124.7 66.4 28.4 7.0 0

Page 63
How to calculate r ?
ℎ : = − ̅
( , )

0 r 1

Page 64
 SMU support size:
1m
Variogram – Range 30 m

Variogram – Range 300 m

r = 1.00

r = 1.00

Page 65
 SMU support size:
101mm
Variogram – Range 30 m

Variogram – Range 300 m

r = 0.69
1.00

r = 0.99

Page 66
 SMU support size:
25
101mm
Variogram – Range 30 m

Variogram – Range 300 m

r = 0.56
0.69
1.00

r = 0.99

Page 67
 SMU support size:
100
25
50
75
101mm
m
Variogram – Range 30 m

Variogram – Range 300 m

r = 0.56
0.37
0.27
0.69
1.00
0.21

r = 0.90
0.99
0.94
0.97
1.00

Page 68
 SMU support size:
1000
500
200
100
150
25
50
75
101mm
m
mm
Variogram – Range 30 m

Variogram – Range 300 m

rr =
= 0.00
0.05
0.56
0.37
0.27
0.69
1.00
0.14
0.21
0.11

rr =
= 0.00
0.90
0.99
0.72
0.39
0.94
0.97
1.00
0.81

Page 69
Grade Tonnage Curves
l Same variable on three different supports:

Zc = 6 Zc = 6 Zc = 6
1m x 1m 5m x 5m 20m x 20m

Page 70
Grade Tonnage Curves

• 1m x 1m (orange)
• 5m x 5m (black)
• 20m x 20m (blue)

Page 71
Global support change

Data distribution

Variogram

« Point » support: Data Application of the

DGM

SMU distribution

Block support: SMU Q,T,M & B

Page 72
Information Effect

Exploration Data Grade Control + Reality

Exploration Data

l An increasing level of information brings the model

closer to the reality.
l Unknown information leads to errors in the model.

Page 73
 Information Effect and QTM

Ideal selection on Zv Poor SMUs

recovered

Zv *

Rich SMUs
not recovered

Zv
Real selection on Zv*
˜ ˜ ˜ ˜

˜ ˜ ˜ ˜

˜ ˜ ˜ ˜

˜ ˜ ˜ ˜
Information effect
decreases the metal
Zv
recovery

Page 74
Information Effect and QTM

Page 75
Information effect
SMU Kriging from
exploration data
SMU Kriging from
grade control

Theoretical SMU
distribution from
DGM

l During the grade control process, the density of information allows

to get closer to the SMU distribution obtained from the DGM
which is considered as « true ».

Page 76
Info. Effect and Tonnage (T)
l A difference remains between the estimated SMU distribution and
the « true » distribution, because the information cannot be
completely exhaustive.

l SMUs to be sent to the processing plant (Tonnage) are eventually

chosen based on the grade control information.

l We can measure the information effect on T with the variance of

Zv*, if it is close to the real variance then the information effect is
low.

« True » Tonnage from DGM

« Estimated » Operational
Tonnage from grade
control
Page 77
Info. Effect and Quantity of Metal (Q)
l To obtain the new value of Q
we need the covariance Zv *

between Zv and Zv*.

l If the covariance is high, we are

closer to an ideal selectivity. Zv

Zv

Page 78
How to calculate s ?
l To obtain Var Zv* and Cov (Zv, Zv*), we will make the assumption
that the estimation from grade control was done by Kriging.

l It is then possible to know Var Zv* and Cov (Zv, Zv*) from a
sampling pattern, an SMU size and a variogram:

Ex: Grade control sampling pattern: Kriging equations give:

∗
= − −
∗
, = − −

Page 79
Transformation into Gaussian
l To work in the Gaussian space Var Zv* and Cov (Zv, Zv*) have to
be transformed (like Var Zv was transformed into r).

∗
( , )
∗

0 sr 1 0 ∗ 1

∗ ∗
Curve: =∑ Curve: , =∑ ∗

Page 80
 Information effect and DGM
l Finally, the information effect is modelled through two additional
parameters:
Y (x) point

∗
→s
→

Y ∗ (v) SMUs Y (v) SMUs

, ∗
→ ∗

Page 81
Uniform conditioning

82
 Estimating recoverable resources

Non-linear models assess the SMU

histogram with accuracy, but not the
spatial variability. SMUs can’t be
directly estimated.

At panel scale we can estimate

recoverable resources, which are
distributions of SMU.

Kriging of factors is applied with a

good quality consistent with
drillholes spacing.

Page 83
Uniform Conditioning overview
Panels SMUs

• Quality estimation • Adequate size for recoverable resources

• Consistent with drillholes spacing • Non-consistent with drillholes spacing
• Inappropriate size to calculate
recoverable resources

Use the panels

estimation to infer
the distributions

Page 84
SMU distribution - Panel relationship
Z(V) panel Y(V) panel

z(V)
FV y(V)

Direct link between

Gaussian distributions

Z(v) SMUs Y(v) SMUs

Fv

Page 85
SMU distribution - Panel relationship
SMUs Direct link between Gaussian
distributions using the
correlation coefficient:

= × + 1− × 0,1
Panels = 0.7 × 1 + 1 − 0.7 × 0,1

= 0.7 + 0.714 × 0,1

Page 86
How to calculate ?
l Panels to SMUs relationship

0 R r 1

l The coefficient amounts the variability of SMUs within

one panel.

Page 87
Information effect
l It is taken into account for the panel distribution too. As for the SMU
distribution its definition derives from the kriging, but here the
dispersion variance (Var Z*) is obtained locally for each panel.

l In UC it is always used for panels (S), not systematically for SMUs (s).

0 S R r 1

Page 88
Information effect – Dispersion variance
l The dispersion variance (Var Z*) entails the smoothing made
during kriging:
- High variance : no smoothing effect, the estimate can potentially take
values between high and low grades.
- Low variance : smoothing effect, the estimate will be close to the average
grade.

Estimate Z* Var Z*

Page 89
 Dispersion Variance
l When taking the information effect into account on panels: =
∗
- If is high, the real variability of SMUs within one panel is
used: S = R
∗
- If is low, the variability of SMUs within one panel is higher
due to the uncertainty of the estimate. S -> 0

Page 90
 Map of distributions

Panels estimates: Z*

UC Results

The average value of the distributions are Z*.

Their variances are function of S :
- Small S gives a high variance of SMUs within the panel
Panels dispersion variance: Var Z* - High S gives a low variance of SMUs within the panel

Page 91
Kriging quality and UC
The Lagrange multiplier amounts the error made
by estimating the local mean in OK.
UC is assuming that the local mean is properly
estimated, thus that µ~0 and that the Var Z* is
inversely proportional to the Kriging variance.

∗
= − −

Page 92
 Classes of VarZ*
∗
∗

Page 93
A Variety of distribution

Page 94
SMU Information effect

∗ = −1.5
=0

Page 95
SMU Information effect

∗ = −0.5
−1.5
=0

Page 96
SMU Information effect

∗ = 0.5
−1.5
−0.5
=0

Page 97
SMU Information effect

∗ = 1.5
−1.5
−0.5
0.5
=0

Page 98
SMU Information effect

∗ = 2.5
−1.5
−0.5
0.5
1.5
=0

Page 99
Uniform conditioning
Panel Z(V) panel Y(V) panel

z*(V)
FV y*(V)
z*(V)
Var z* ∗
→

Panel Z(v) SMUs Y(v) SMUs

Fv

Q,T,M,B

Page 100
Q & T calculation
l The results of UC are the QTM values over panels. They can be modeled
with Hermite polynomials too, knowing and the panel estimate ∗ for
each cutoff we have:
∗
- =1−

=∑ ∗
- ., ,

Example of Q and T for a given cutoff :

For each panel, Q and T are the recoverable Metal and Ore at a given cutoff. To get
real tonnages, one needs to multiply them with the panel mass.
Page 101
Results
l Results of UC can be compared to the theoretical SMU distribution from
the Global Change of Support. Their Grade Tonnage Curves should
match, although some differences due to non-stationarity and clustering
are acceptable.

Page 102
 UC robustness

l Given a bad SMU distribution as input (no declustering weights), the UC

manages to correct this bias and to come very close to the reality (Walker Lake
case study) thanks to the Kriging declustering effect inherited from the Panel OK.

Page 103
Multivariate UC
l In Multivariate UC, we only model bivariate relationships
with the main variable.
l The main variable UC results are the same in univariate and
in multivariate.
Sec. Var 1
Ex: Al2O3

Main Var
Ex: Fe
Sec. Var 2
Ex: P

Sec. Var 3
Ex: SiO2
Page 104
SMU distribution - Panel relationship
Z (V) panel Y (V) panel Y (V) panel
z (V) FV2
y (V)
y (V)

Simple
Kriging
(v) SMUs Y (v) SMUs Y (v) SMUs
Fv2

Page 105
Building the regression
Y (V) panel Y (V) panel
∗ ∗
( , )→ ∗ ∗
y (V)
y (V)
Hypothesis of
conditional
∗ ∗
→
independancy
→
→ →

Y (v) SMUs Y (v) SMUs

« Data - Data » 1 ∗ ∗ ∗ ∗
« Data - Target »
∗ ∗ 1 × =
∗ ∗

∗ ∗ 1
Page 106
Hypothesis of conditional independency
l If we want to know the covariance between two variables,
knowing a third one the theoretical formula is:

l , = , | + | , |

|
, |

|
Page 107
 Hypothesis of conditional independency
If the density probability of the three variables is not known,
then one can make assumptions in order to assess
, :
1. If the variables follow N(0,1) distributions, we have:
, = , = , | + , ,

2. The hypothesis that Y1 and Y2 are independent

conditional on Y3 let us write:
, = , = , ,

Page 108
Practical key points of UC
l Quality Kriging ensures quality UC
-> make sure that the Var Z* is coherent.
l The change of support is sensitive to the first part of the
variogram : take good care of the nugget effect.
l Use a lot of cutoffs: it doesn’t increase significantly the time
of the process, but improves quality, consistency and later on
the localization process.
l In multivariate the model may be slightly inconsistent, check
the input parameters if s2vk is above 1.1.

Page 109
UC vs. Conditional simulations
l When proper hypothesis are respected for UC and for
conditional simulations, the QTM obtained with UC
are the same as the average QTM of simulations:

Page 110
QTM post-processing

QTM VALIDATION

111
Valid QTM ?
l We expect that the QTM variables follow certain
rules:
- T & Q are always decreasing, and positive;
- T0 = 1;
- ‘T constant’ is equivalent to ‘Q constant’;

Page 112
 Valid QTM ?
l Between two increasing <
−
<
, ,
−
cutoffs, the loss of metal and
tonnage have to be consistent. , ,

, = 200

, = 300

Page 113
Correction
l Non-linear methods may produce slightly inconsistent
QTM.
l In Isatis it is possible to correct QTM which are not
coherent. They are made consistent with basic
corrections.

Page 114
QTM post-processing

LOCALIZATION

115
 Localizing UC results

Q (t)
Cut-off zero
Δ
Δ
=
.
Δ .
.
Δ
=

16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 T (t)
= / = /
Mean grades are deduced for every SMU within the panel
according to their rank and the GT curves.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

5.1 5.8 6.0 6.1 6.3 6.6 6.7 6.8 6.9 6.9 7.2 7.5 7.7 7.9 8.3 8.7

Page 116
Localization
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

5.1 5.8 6.0 6.1 6.3 6.6 6.7 6.8 6.9 6.9 7.2 7.5 7.7 7.9 8.3 8.7

What location?

5.9
1 6.3
5 7.0
10 7.1
12 5.1 6.3 6.9 7.5

6.0
2 6.3
6 7.4
14 7.7
16 5.8 6.6 7.9 8.7

3
6.1 7
6.5 13
7.2 15
7.5 6.0 6.7 7.7 8.3

4
6.2 8
6.8 9
6.9 11
7.0 6.1 6.8 6.9 7.2

1. OK of the grade on SMUs 3. Localization

2. Estimates ranking

Page 117
 Localization
Var 1: Main Variable Var 2: Secondary Variable
Q (t) Q (t)
Δ Cut-off zero Cut-off zero

Δ Δ

16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 T (t) 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 T (t)
= / = / = / = /

Mean grades are deduced for every SMU within the panel
according to their rank and the GT curves.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

Var 1 5.1 5.8 6.0 6.1 6.3 6.6 6.7 6.8 6.9 6.9 7.2 7.5 7.7 7.9 8.3 8.7

Var 2 2.48 1.85 1.61 1.15 0.95 0.71 0.59 0.52 0.46 0.40 0.35 0.32 0.27 0.23 0.15 0.12
Page 118
 Multivariate Localization
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

Var 1 5.1 5.8 6.0 6.1 6.3 6.6 6.7 6.8 6.9 6.9 7.2 7.5 7.7 7.9 8.3 8.7

Var 2 2.48 1.85 1.61 1.15 0.95 0.71 0.59 0.52 0.46 0.40 0.35 0.32 0.27 0.23 0.15 0.12

What location?
5.9
1 6.3
5 7.0
10 7.1
12 5.1 6.3 6.9 7.5 2.48 0.95 0.40 0.32

6.0
2 6.3
6 7.4
14 7.7
16 5.8 6.6 7.9 8.7 1.85 0.71 0.23 0.12

6.1
3 6.5
7 7.2
13 7.5
15 6.0 6.7 7.7 8.3 1.61 0.59 0.27 0.15

6.2
4 6.8
8 6.9
9 7.0
11 6.1 6.8 6.9 7.2 1.15 0.52 0.46 0.35

1. OK of the Var1 grade on SMUs 3. Localization

2. Estimates ranking

Page 119
QT discretization
l In order to model the curve Q vs. T we need a lot of
cutoff during UC. Otherwise some SMUs will have the
same values:

734 487 473 448 302 302 302 302 302 204 99 99 99 99 99 99

Page 120
 QT discretization
If the number of SMU per panel is too small, their values may be slightly
smoothed, and some extreme SMU values are not reproduced:

739 546 446 423 373 339 322 277 226 198 170 143 123 74 30 19 538 328 184 61

16 SMUs per panel 4 SMUs per panel

Page 121
Artefacts
l If the variogram is very erratic, with respect to the sampling (1 sample per panel):

Reality: Kriging: Localization:

Real global variance Smoothed global variance Real global variance
Real spatial variability Smoothed spatial variability Smoothed spatial variability

Ø The SMU values are accurate but not at the right location within the panel.

Page 122
 Summarized method
Z(V) panel Y(V) panel
3. Panel anamorphosis:
Experimentally defined z*(V)
FV y*(V)
z*(V)
through dispersion variance Var z*
Var z*
∗
→
3. Panels Ordinary Kriging
=

Z(v) SMUs Y(v) SMUs

Fv
2. SMU anamorphosis:
Global Suppor t Change

4. Uniform Conditioning

Z(x) points Y(x) points

F
1. Point anamorphosis

5. Localization

Page 123
UC - Results Analyse

Kriging V

LUC V

Page 124
UC - Results Analyse
Kr V - Panel LUC

Kr V - SMU True V

Page 125
QTM post-processing

MULTI-DOMAIN LOCALIZATION

126
LMUC on multidomains
Domains are defined on SMU.
Thus panel may contain different domains.
UC is applied on panels that contain at least one SMU

Page 127
LMUC on multidomains
UC is performed separately on both domains.

UC on Domain 1 UC on Domain 2

Page 128
LMUC on multidomains
UC is performed separately on both domains.

UC on Domain 1 UC on Domain 2

Page 129
LMUC on multidomains
Method 1: Method 2:
- LUC on each domain - UC average on panels
- LUC global using OK of every domain

Page 130