You are on page 1of 54

KRIGING NEIGHBORHOOD

ANALYSIS
Chapter outline
• Kriging properties
• Neighborhood
• Kriging Neighborhood Analysis

2
KRIGING NEIGHBORHOOD
ANALYSIS
KRIGING PROPERTIES
Kriging properties
• The data values are not used to calculate the kriging
weights nor the kriging variance
₋ Hence the use of kriging for optimizing sampling patterns or
additional well locations.

• Weights remain unchanged when multiplying the sill of


the variogram by a constant. Lagrange parameter is
multiplied by this constant.

• The kriging variance is directly proportional to the sill of


the variogram.

4
Kriging properties

Kriging smoothes the real variability:


≈ − −2

5
Indications of kriging quality
Characteristics of estimated points or blocks to be checked:

• Conditional bias
slope of regression E[Z|Z*]

• Precision
estimation variance or standard deviation sk

• Smoothness
dispersion variance of the kriged estimates Var[Z*]

6
Indications of kriging quality
• In the stationary case:
- These parameters may help define the block size of estimation
and/or the neighbourhood. It is recommended to make numerous
tests to compare configurations -> KNA
- They can also be used for resource classification purposes according
to the search ellipsoïd
- All these parameters depend on the variogram model ( a different
variogram model would give different results). The uncertainty on
the variogram model could be of more importance.

7
• Three linked random variables are used to define Kriging
outputs : Z, Z* and Z-Z*

• Their relationship is characterized by the variogram: ℎ =


− ∗

--> We have access to probabilistic parameters, although we


only have one value of Z*.

− = = +

= =

= − −

, = − −
, ∗

, =

8
Conditional Unbiasedness
In addition to the global unbiasedness, a good interpolator should
be conditionally unbiased.

[
E Z0 Z = Z *
] *

9
Slope of regression
The linear regression between the true and estimated grades, with the
local condition of search ellipsoïd and variogram model

Slope = 0.95 and Slope = 0.67 and conditionally


conditionally unbiased unbiased ??

10
Kriging variance
• It depends only on the variogram model and the data configuration

• It is not suited for calculation of confidence intervals, or for calculating


probabilities to exceed a cut-off, unless :
1

0.9
- The grade histogram is Gaussian and 0.8

0.7 Kriging
- The block(s) are well estimated and 0.6 standard

fr e q u e n c y
0.5 deviation
- The stationary assumption is fulfilled 0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100
Possible Z values

11
Variance of Z*
Kriging provides an estimate of the dispersion variance of the estimator.
It is a global parameter locally estimated.

Low variance of Z* means the estimate is smoothed (range < sampling distance)
High variance of Z* means the local estimate is good and allows high selectivity

12
Smoothing effect of kriging
How is the variogram calculated on kriged values compared to the variogram of
the data?

The variogram computed on the kriging grid has a sill that is much lower than the
variogram computed on the real data. (2 times lower)
The range is highest on the variogram computed on the kriging estimation.

13
Kriging system



= ∗

≥ ∗




,

,


( , )
= ∗

∗ = 0
∗ − ∗=
( ) /
=
14
KRIGING NEIGHBORHOOD
ANALYSIS
NEIGHBORHOOD
Neighborhood
• Define the selection of samples used in the estimations
is a key problem;

• 2 types of neighborhood can be used:


- Unique:
o all data are used for each target estimate
o applicable only for small datasets

- Moving:
o Each time a new point/block needs to be kriged, a new set of limited number of
data receiving weights will be performed.

16
Unique Neighborhood
• All data used for each target estimate

• Kriging matrix is inverted once for all targets

• Provides maximum data for estimation

• Prevents artifacts from abrupt changes

• Impractical for many minerals projects due to memory


limitations and CPU time- kriging matrix is unworkable

Solution: to restrict the data selection to a s subset of points,


changing with the target -> it is called moving neighborhood

17
Moving Neighborhood
• Compulsory for large datasets or when stationarity
not fulfilled at any scale

• Data outside a 2D ellipse or 3D ellipsoid is


discarded

• Selection criteria (defined by user) define a subset


of the data inside the ellipse for estimation

• Each target will be kriged with a different set of


data

• How to optimize?
18
Moving Neighborhood
Search ellipse parameters depend
upon:
• Stationarity extension
• Data density
• Variogram parameters
⁻ spatial scale
⁻ anisotropy ratio
⁻ azimuth

Kriging: a Linear Estimator


Z 0 = l1Z1 + l2 Z 2 + ... + ln Z n
Z1 , Z 2 ,..., Z n Data Points
l1 , l2 ,..., ln Weights

19
Moving neighborhood parameters
• Range: ellipsoid radius should be slightly greater than
variogram range
₋ High variance model (high nugget, short range) and/or scarce data: large
ellipsoid is advised
₋ High continuous model (low nugget, large range) and/or numerous data:
large ellipsoid is not required.
₋ Definition of ellipsoid size and shape should be consistent with the
stationary hypothesis

• Orientation of ellipsoid is often parallel to anisotropy


but it is not compulsory, it depends on model
orientation and data configuration

20
Anisotropic distances

When neighborhood
ellipse is parallel to
variogram anisotropy:
Using anisotropic
distances during data
selection phase
increases selection of
data points in
direction of maximum
continuity

Using anisotropic distances Using euclidean distances


21
Moving neighborhood parameters
• Sectors force the search to be spread around the target, to avoid
selecting samples from just one direction or drill hole

• Minimum number of samples: common to need estimate for all


points/blocks, to be balanced against quality

• Maximum number of samples: more data is generally better but


can induce negative estimates

• The combination of range and maximum samples is important-


need to balance need to estimate (all) blocks against quality

22
How to choose the model?
• The variogram model should not contradict the
naturalist interpretation

• The cross validation can help to choose the model:

- For a given data set the cross validation checks the


consistency of the model of spatial correlation with the data.
- The procedure is the following :
o Each data point is removed in turn
o It is then kriged using the other data.
o Statistics on the errors are then achieved.

23
KRIGING NEIGHBORHOOD
ANALYSIS
KNA
KNA or QKNA
Quantitative Kriging Neighbourhood Analysis for the Mining Geologist — A Description of the Method With Worked Case Examples
J Vann, S Jackson and O Bertoli

“Arbitrary specification of the kriging neighbourhood is very risky


because the kriging weights are directly related to the variogram
model, data geometry and block/sample support involved in the
kriging. Whilst kriging is commonly and correctly described as an
‘minimum variance estimator’ this is only true when the
neighbourhood is properly defined”

The neighbourhood may be appropriately defined using Kriging Neighborhood Analysis


(KNA) or Quantified Kriging Neighbourhood Analysis (QKNA).

25
Kriging Neighborhood Analysis (KNA)
• Defining an appropriate search neighborhood is important:
- Improved estimation quality;
- Unbiased estimates; and
- No negative estimates.
• KNA is a method allowing you to optimize and justify
neighborhood parameters.

26
Kriging Neighborhood Analysis (KNA)
• The results of KNA can assist with block size selection, choice of discretization
and mineral resource classification decision…

• The misconception “Search to the range” of variogram is not good strategy


for defining kriging neighborhood -> the choice of radius neighborhood should
be influenced more by the slope of the variogram model at short lags and the
relative nugget effect.
- Pure nugget: the most reliable estimate will be made the largest number
of samples ->whole domain;
- Short range of variogram: the neighborhood required for a good
estimation will progressively get larger;
- Large range of variogram: only nearby samples will be required to ensure
that a good estimate;

27
KNA – General Process

Define sets of parameters and


a section

Run geostatistical criteria


statistics

Selection optimum parameter


set that optimizes the
geostatistical criteria

28
KNA – General Process
• A representative block or selection of blocks is estimated for each
test.

• The output parameters are calculated on this support.

29
How to analyse estimate quality?
• Evaluation Criteria
- Slope of regression
- Kriging Efficiency
- Sum of Positive Weights
- Weight of the Mean
- Correlation Z|Z*
• Performance criteria:
- Number of estimated blocks
- Running Time

30
Slope of Regression
• Kriging smoothes the real variability and the slope of regression sum up
the degree of oversmoothing of low and high values of variable
estimated.

• It is the regression slope between


the corresponding true variable
and the estimated block.

• It is calculated as:

∗ =ρ
[ ∗]
( , ∗)
∗ = ∗

C00 means covariance between the target point and itself. In the case of block kriging,
Cvv (covariance between the target block and itself) is used instead. 31
Slope of Regression
• The slope approaches one when the Lagrange multiplier will be small->
good data coverage, no extrapolation & limited clustering -> limited
oversmoothing;

• When the Lagrande multiplier is large it dominates the equation and the
slope tends towards a half.
− +μ
=
− +2∗μ
Where BV = Block Variance (i.e. the variance of actual block values) and KV = Kriging Variance
(i.e. the error variance of the block estimate);

32
Correlation true and estimated blocks

The covariance between true and estimated blocks Cov ( , ) can be calculated:


Cov , = ( , )

Where ( , ) is the mean value of the covariance function between each sample location
and the block to be estimated V.

Note that given the kriging equations which state for any l: ∑ , − = ( , )

∗ ∗
We have: , = ( ) −

So that: =1− ( ∗)

Where µ is the Lagrange multiplier used in solving the kriging system written the covariance
function

33
Kriging Efficiency
( − )
=

Where BV = Block Variance (i.e. the variance of actual block values) and KV = Kriging
Variance (i.e. the error variance of the block estimate);

• For perfect estimations:

K = , the dispersion variance = , then:


( − )
= = %

• Where only a global estimate of all blocks is practical, all blocks


will be valued at the global mean, i.e.:
( − )
= = = = %

Krige (1996)
34
Kriging Efficiency
( − )
=

• Usually blocks are estimated imperfectly, in other words with no


conditional biases:
( − )
= − and = =

• However, with conditional biases present this relationship does


not hold and then:

> − because of insufficient smoothing, and


( − )
< =

Krige (1996)
35
Kriging Efficiency
• The efficiency can even be negative if KV>BV. Such a situation is
ridiculous and the block valuations will be worthless:
( − )
=

Krige (1996)
36
Sum of positive weights
• Condition pour le KO:

=1

• Sum of the Positive kriging weights assigned to the samples used


to estimate each block ->the closer to one, the better the
estimate.

Krige (1996)
37
Weight of the mean

• The weight assigned to the mean when estimating a block with


Simple Kriging.
- Simple Kriging: Kriging with known mean
- Ordinary kriging : the mean is not known
• The weight of the mean represents the lack of information;

λm → 1 is equivalent to an uncertain local estimation


λm → 0 is equivalent to a well-defined local estimation

38
Weight of the mean

It is derived from a simple kriging formula:


1. To perform a Simple kriging from data;
2. To calculate the weight assigned to the mean;


= + × =1−

Is the weight assigned to the global mean


grade or “ weight of the mean”

The weight of the mean gives a clear indication of the kriging


quality

39
Correlation Z|Z*
The covariance between actual value and estimated value
Cov ( , ∗ ) can be calculated:

Cov , ∗ = − −2

40
Kriging variance
Kriging variance is the statistical variance of the error between the block
estimate and its real grade


= 0

− ∗=
/

41
Variance of Z*
Kriging provides an estimate of the dispersion variance of the estimator.
It is a global parameter locally estimated.

= ∗

≥ ∗

42
Block size and sampling
Main grade Secondary grade

« Reality » - points

« Reality » - blocks
-> Size consistent
with sampling

Estimates - blocks
-> Close to reality

43
Block size and sampling
Main grade Secondary grade

« Reality » - small blocks


-> Size not
consistent with
sampling

Estimates – small blocks

-> Acceptable.. -> Smoothed

44
Block size
Scatter Diagram V true vs. V estimated

Histogram of Kriging Standard Deviation

1x1 5x5 20x20


45
Block size

Increasing the block size improves the quality of estimation.

46
Block size
• Block size is chosen so as to
ensure quality Kriging.

• One sample per block is


considered as sufficient. For a
continuous grade, a smaller
block may be used. Conversely a
larger block is advised for erratic
grades.

47
Number of Data

KE and the slope are strictly increasing with the number of data involved
in the estimation.
-> Need for other output parameters!

48
Number of data

The running time or the number of positive weights allow to choose an


optimum number of data.

49
Size of the ellipsoid

KE and the slope strictly increase with


the ellipsoid size.
Be careful to choose a neighborhood
consistent with the stationarity
hypothesis.

50
Variogram and ellipsoid size

Short range variogram Long range variogram

The ellipsoid size required to optimize the quality of estimation is smaller


for the long range variogram than for the short range variogram.

51
Size of the ellipsoid

52
Minimum number of data

Increasing the minimum number of data will improve the quality of


estimation. It is due to the elimination of poorly estimated blocks.

53
Use Anisotropic Distance

• Unenabled: the distance will be isotropic standard distances:


Direction W

a
Direction V
a
2a

• Enabled: the distance will be anisotropic;


Direction W

a
Direction V
a
2a

54

You might also like