Kriging

Applied Geostatistics
For Mining Professionals
Presented by
John Vann
BAppSc BSc(Hons) MSc(Mining Geostatistics)
Managing Director
Quantitative Geoscience, Perth
Aims of this Session

Aims of this Session:

 How to express the estimation variance
for interpolation?

 How to minimise this?

 OrdinaryKriging!
 Setting the search quantitatively...
Definition

Ordinary kriging is a linear interpolator which ensures

minimum estimation variance for a given block support, data
geometry and specified variogram model.
What Do We Want?

Best possible precision (min. estimation variance)

and unbiased.
 Take into account data quality

 Take into account sampling geometry (clustering)

 Take into account spatial continuity
 Index of reliability
 Support!
 Why Krige?

Kriging takes into account all the factors we just stated.

 B.L.U.E.

 Interpolate by weighted average Q: how to get weights?

 Variogram: model of uncertainty (at / between samples)

 Probabilistic approach allows us to calculate and thus

minimise estimation variance
Objectives of Kriging (1)

1. Minimise estimation variance

2. Unbiased (sum of weights = 1)

Objectives of Kriging (2)

 The OK estimator provides min estimation variance

for any linear estimator.
 By taking a probabilistic approach to the problem we
can calculate the estimation variance.

 If we can express the estimation variance as an

equation, we can minimise it.
 The Unbiased Condition

The unbiased condition must be met.

 This ensures global lack of bias (not local)

 Polygons, nearest neighbour and IDW all meet this

condition
N


 How?
 The sum of the weights i  1.0
must add up to 1.0 i 1
More Than 1 Sample

 We previously saw estimation variance for Polygonal

(only 1 sample)

 What about an interpolation?

N
*
ZV  Z x  
i 1
i i
The Polygonal Case

 First consider the Polygonal case:

 e2  xi ,V   2  xi ,V    V ,V 

 Note that we are assuming here that the support of

our samples are very small compared to our block.
What Additional Factors?

For interpolation by n samples, what extra terms?

e  xi , V   2  xi , V    V , V 
2
 1 sample, 1 block: 
 What else?
 The weights themselves (...λ’s)
 Support of v and V, plus location of samples wrt
block? Aready accounted for….

 The location of samples within the search….

Clustered? Location of samples w.r.t. other samples
The Equation for n>1 (1)

 1 block 1 sample:

 e2  xi , V   2  xi ,V    V , V 

 The following slide shows the generalised case...

That is: for n>1 (more than one sample)
 The Equation for n>1 (2)

 e2  xn , V   2    ( x ,V )      ( x , x )   (V ,V )
i i i j i j

 This is the variance of the error we incur when we

estimate a block V using n samples located within a
search.
 The new terms are (I) the weights.
 …and (II) the location of samples w.r.t. other
samples
 Looking at the Equation

The estimation variance for interpolation :

 e2  xn ,V   2    ( x ,V )      ( x , x )   (V ,V )
i i i j i j

 This is a set of equations (1 per sample)

 How do we find the minima of a function?

 Calculus (set derivatives to zero)

Minimising (1)

 Differential calculus can yield the minima

 Some functions can have more than 1 minima
 We require the minimum under a constraint: what?
N


i 1
i  1.0

 When we differentiate we use a “dummy” variable μ

 This is the ‘Lagrange multiplier’ (constrained minima)

Kriging Equations (1)

The results are the kriging equations:

N N

    ( x , x )     ( x ,V )
i 1 j 1
i i j i i, j  1,2,..., N


i 1
i 1

 For a multivariate (the weights!) calculus problem the

minima is a set of equations
Kriging Equations (2)
This set of equations is not intuitive to look at!

 nb: when we solve the equations, we get the weights.

N N

    ( x , x )     ( x ,V )
i 1 j 1
i i j i i, j  1,2,..., N


i 1
i 1

 This is set of equations. How do we solve these?

Kriging Equations (3)

 N equations: 1 each for each sample found

 +1 (sum of the weights = 1.0)
 Remember simultaneous equations?

 How do we solve?
3x  2 y  5
4x  3y  3
Kriging Equations (4)

What if we have 2 equations and only 3 variables?

3x  2 y  7 z  5
4x  3y  2z  3
 Can’t solve!!! Need one equation per variable
Solving

 To solve we needed the “dummy” variable

 This variable is μ (the “Lagrange multiplier”)

 When we solve, μ is approximately equal to zero
Kriging Variance

We can now express the minimum variance, which is

the “kriging variance”
N
2
 OK  Var( zV* , zV )     ( x ,V )   (V ,V )  
i 1
i i
 A Better Way to Look at
the Kriging Equations
There is a better way to view the kriging equations
 Consider our simple simultaneous equations:

3x  2 y  5 3 2   x  5
4  3   y   3
4x  3y  3      
 Now look at them in matrix form
 To solve large systems of simultaneous equations, it
is far more efficient to use a matrix approach.
Matrix Form

  11  12   1N 1  1    ( x1 ,V ) 
     
 21  22  2N 1   2    ( x 2 , V ) 
          
     
 N 1  N 2 NN 1 N   ( x N ,V )
 1 1  1 0     1 
point-point gamma
map of sample locations weights Point-block
using variogram distance [unknowns]
Properties of Kriging

 Exact Interpolator (points not blocks)

 OK doesn’t depend on data values (stationarity!)

 Unique solution (providing we use an authorised

model)

 Combining kriged estimates ….

‘Screen’ Effect

 Small nugget - large screen effect

 Large nugget - NO screen effect

 Increasing declustering as nugget decreases

 Choosing how far to search depends on screen effect

 Quantified ‘Kriging Neighbourhood Analysis’

Neighbourhood Analysis (1)

 Objective: minimise conditional bias

 In other words: minimise estimation variance

(linear case)

Quantify! Measure! Science!

Neighbourhood Analysis (2)

How? What criteria should we consider?

(1) Look at the weights themselves: close to 0 at edges
of the searched volume/neighbourhood. (negative?)
(2) Slope of the regression (Z*v|Zv) approx. = 1.0

(3) Weight of the mean for simple kriging (SK) is low.

 What is SK????
Neighbourhood Analysis (3)
SK: Simple Kriging
 SK presumes strict stationarity

 Thus, the weights in the local neighbourhood < 1.0

 Remaining weight is allocated to the mean of

samples in the domain.
N This is the
*
Z SK V
  Z  x i    i   m  ‘weight of the
m mean’
i 1

 Q: Pure nugget case?

Using SK

 We use SK only for neighbourhood analysis!

…. i.e. the screen effect
 If the weight of the mean is significant: search more
 If the weight of the mean is low, then the samples
found are adequate to minimise conditional bias
 Again, recall pure nugget case.

 Very useful statistic for neighbourhood analysis

 Along with slope and values of weights
How To Do It! (1)

This is a critical step and it distinguishes “scientific”

linear interpolation from a crude job.
 Select various estimation geometries

 Generate key statistics: KV, slope, weight of the

mean, bar gamma v,v etc.
 Iterative: choose the search with the best results
How To Do It! (2)

Block size with impact on statistics!

 We can also look at discretisation sensitivity

 Search far enough to ensure a good result.

 Some smoothing is required!
 Smoothing Relationship (1)

 Some smoothing is required to minimise estimation

variance (we are estimating BLOCKS!)
 The so-called ‘smoothing relationship’ of kriging:

2 * 2 2
D [ Z (v ) | V ]  D [ Z (v) | V ]   OK  2
Smoothing Relationship (2)

Estimated blocks will be smoother than true blocks

 Information effect
 Nb. Langrange is usually a
negligible term
Point Kriging

We can also krige points (as opposed to blocks).

For example, in contouring applications.

A very simple   11  12   1N 1  1    ( x1 , x0 ) 
 
 1  2    ( x2 , x0 ) 
change to the  21  22  2N
         
equations:     


 N 1  N 2 NN 1  N   ( x N , x0 )
 1
 1  1 0     1 


Point-point
point-point gamma
weights
map of sample locations
using variogram distance [unknowns]
Steps in Kriging (1)

1. Check the data. An advantage of the geostatistical

approach is that it is intolerant of poor data

2. If the variogram is horrible, what should we do?

3. Set up a block model. Avoid very small blocks!

Steps in Kriging (2)

4. Kriging neighbourhood analysis

5. Krige!

6. Check results. Should they honour the data,

generally?