Spatial Interpolation

,
Kriging
Thomas K. Windholz
Outline
• Spatial Interpolation Basics
• Simple, Ordinary, and Universal Kriging
• Flavors of Kriging
Spatial Interpolation Basics
• Spatial interpolation allows us predicts
values at unsampled locations.
• In general, a model fitting samples can
be split into first and second order
components.
• First order can be captured by, for
example, a trend surface through
regression.
• Second order looks at residuals and their
Covariance (e.g., Kriging—named after
South African mining geologist D.G.
Krige)
Outline
Spatial Interpolation Basics
• Simple, Ordinary, and Universal Kriging
• Flavors of Kriging
Simple, Ordinary, and
Universal Kriging
• What is the difference?
• Derivation of simple kriging
• Equations for prediction & kriging
variance
• Analysis steps
• Augmentation to incorporate trend
Differences Among Kriging
Methods
• Remember equation for a spatial random
variable:


• Simple Kriging: No trend (only )
• Ordinary Kriging: Constant trend
• Universal Kriging: Polynomial trend
) ( ) ( ) ( s U s x s Y
T
+ = |
) (s U
Derivation of Simple Kriging
• Basic Idea:
¿
=
=
n
i
i i
s U s s U
1
) ( ) ( ) (
ˆ
ì
s
1

s
s
2

s
3

s
4

ì
1

ì
3

ì
4

ì
2

Derivation of Simple Kriging
• Basic Idea:

– How should differ from ?
) (
ˆ
s U
s
1

s
s
2

s
3

s
4

) (s U
ì
1

ì
3

ì
4

ì
2

Derivation of Simple Kriging
• One descriptor is expected mean square
error:
( ) ( ) ( ) ( ) ( ) ) (
ˆ
) ( 2 ) ( ) (
ˆ
) ( ) (
ˆ
2 2
2
s U s U E s U E s U E s U s U E ÷ + = ÷
¿ ¿¿
= = =
÷ + =
n
i
i i
n
i
n
j
j i j i
s s C s s s C s s
1
2
1 1
) , ( ) ( 2 ) , ( ) ( ) ( ì o ì ì
) ( ) ( 2 ) ( ) (
2
s c s s C s
T T
ì o ì ì ÷ + =
Derivation of Simple Kriging
• One descriptor is expected mean square
error:
( ) ( ) ( ) ( ) ( ) ) (
ˆ
) ( 2 ) ( ) (
ˆ
) ( ) (
ˆ
2 2
2
s U s U E s U E s U E s U s U E ÷ + = ÷
) ( ) ( 2 ) ( ) (
2
s c s s C s
T T
ì o ì ì ÷ + =
minimize
¿ ¿¿
= = =
÷ + =
n
i
i i
n
i
n
j
j i j i
s s C s s s C s s
1
2
1 1
) , ( ) ( 2 ) , ( ) ( ) ( ì o ì ì
Derivation of Simple Kriging
• Minimization (through differentiation) results
in:



• …which we can use back in our starting
equation of:
) ( ) (
1
s c C s
÷
= ì
¿
=
=
n
i
i i
s U s s U
1
) ( ) ( ) (
ˆ
ì
Derivation of Simple Kriging
• Note on the side:
) , (
j i
s s C
) , (
i
s s c
… Covariance matrix among all sample sites
… Covariance vector between
prediction location s and all sample sites s
i

Kriging variance
• The kriging variance results in
(substitute ì(s) in the expected mean
square error):
( ) ( ) ) ( ) ( ) ( ) (
ˆ
1 2 2
2
s c C s c s U s U E
T
e
÷
÷ = = ÷ o o
Analysis Steps
• Remove trend if it exists
• Calculate empirical variogram on
residuals
• Fit theoretical variogram
• Calculate C and c (actually I
-1
and ¸)
• Predict value & add trend
• Estimate error
Augmentation to incorporate
trend
• Ordinary kriging incorporates a constant
trend
• Universal kriging incorporates a trend of
order x
Ordinary Kriging
• General:

• Simple Kriging:

• Ordinary & Universal Kriging:
) ( ) ( ) ( s U s x s Y
T
+ = |
¿
=
=
n
i
i i
s U s s U
1
) ( ) ( ) (
ˆ
ì
¿
=
=
n
i
i i
s Y s s Y
1
) ( ) ( ) (
ˆ
e
Ordinary Kriging
• Can handle constant trend (mean)
through an augmented matrix C
+
and
augmented vectors e
+
(s) and c
+
(s).
+
C
) (s
+
e
=
=
) (s c
+
|
|
|
|
|
.
|

\
|
0 1 1
1 ) , ( ) , (
1 ) , ( ) , (
1
1 1 1


   

n n n
n
s s C s s C
s s C s s C
|
|
|
|
|
.
|

\
|
) (
) (
) (
1
s
s
s
n
v
e
e

|
|
|
|
|
.
|

\
|
1
) , (
) , (
1
n
s s C
s s C

Ordinary Kriging
• Can handle constant trend (mean)
through an augmented matrix C
+
and
augmented vectors e
+
(s) and c
+
(s).
+
C
) (s
+
e
=
=
) (s c
+
|
|
|
|
|
.
|

\
|
0 1 1
1 ) , ( ) , (
1 ) , ( ) , (
1
1 1 1


   

n n n
n
s s C s s C
s s C s s C
|
|
|
|
|
.
|

\
|
) (
) (
) (
1
s
s
s
n
v
e
e

|
|
|
|
|
.
|

\
|
1
) , (
) , (
1
n
s s C
s s C

Simple kriging
Ordinary Kriging
• Constant trend will be simultaneously
predicted.
• Can estimate variogram from y values
without removing trend (since it is
constant).
• Usually works within a neighborhood and
not with entire dataset.
• Since it works in a neighborhood trend only
has to be constant in the neighborhood (be
cautious with this statement)
Universal Kriging
• Can handle polynomial trend:
+
C
) (s
+
e
= ) (s c
+
|
|
|
|
|
|
|
|
.
|

\
|
0 0 ) ( ) (
0 0 ) ( ) (
) ( ) ( ) , ( ) , (
) ( ) ( ) , ( ) , (
1
1 1 1
1 1
1 1 1 1 1 1
 
     
 
 
    
 
n p p
n
n p n n n n
p n
s x s x
s x s x
s x s x s s C s s C
s x s x s s C s s C
|
|
|
|
|
|
|
|
.
|

\
|
) (
) (
) (
) (
1
1
s
s
s
s
p
n
v
v
e
e


=
|
|
|
|
|
|
|
|
.
|

\
|
) (
) (
) , (
) , (
1
1
s x
s x
s s C
s s C
p
n


Universal Kriging
• Polynomial trend will be simultaneously
predicted.
• Cannot estimate variogram from y
values without removing trend first!!!
• Thus, trend has to be removed first
anyway to estimate variogram!
• Neighborhood not such an issue as with
ordinary kriging.
Outline
Spatial Interpolation Basics
Simple, Ordinary, and Universal Kriging
• Flavors of Kriging
Flavors of Kriging
• Block Kriging
• Co-Kriging
• Others: Robust, Disjunctive, and
Indicator Kriging
Block Kriging
• Uses a block A rather than locations s.
• For example:
) , (
i
s s C
A
ds s s C
s A C
A
i
i
}
=
) , (
) , (
) , ( s s C
2
) , (
) , (
A
s dsd s s C
A A C
A
}}
' '
=
Co-Kriging
• Basic idea is to use a second, highly
correlated, variable in locations where
primary variable is (or cannot) be
measured.
… Primary and secondary variable
Samples with:
Co-Kriging
• Basic idea is to use a second, highly
correlated, variable in locations where
primary variable is (or cannot) be
measured.
… Primary and secondary variable
… Secondary variable only
Samples with:
Co-Kriging
• Essential component of co-kriging is the
cross-covariogram or cross-variogram:
( )( ) ( )
X Y YX
s X h s Y E h C µ µ ÷ ÷ + = ) ( ) ( ) (
( )( ) ( ) ) ( ) ( ) ( ) ( ) ( 2 s X s X s Y h s Y E h
YX
÷ ÷ + = ¸
Co-Kriging
• … and the empirical cross-variogram
can be estimated through:
¿
= ÷
÷ ÷ =
h s s
j i j i YX
j i
x x y y
h n
h ) )( (
) (
1
) (
ˆ

Co-Kriging
• Thus, we can model the variable Y(s)
through:



• With the solution for e again as an
augmented system e
+
.
¿ ¿
+
= =
+ =
m n
j
j xi
n
i
i yi
s Y s s Y s s Y
1 1
) ( ) ( ) ( ) ( ) (
ˆ
e e
Spatial Interpolation,
Kriging
Thomas K. Windholz