Kriging

Using Geostatistical Analyst, ESRI Chang, Kang-tsung, 2006, Introduction to Geographic Information Systems,

So Far
We dealt with deterministic models that can not provide estimates for accuracy/certainty in predictions. Kriging is a stochastic model that provides estimates for accuracy/certainty in predictions.

Kriging

Spatial variation consists of 3 components Random spatially correlated component A drift or structure/trend Random error term

How to measure the spatially correlated component?

Uses Semivariogram to measure spatially correlated component, also called spatial autocorrelation

Creating a Semivariogram?
First what is semi-variance between two point separated by distance h?

Where:
= semivariance between point Xi and Xj h Z = distance separating Xi and Xj = attribute value (height, ore quality, etc)

Semivariance between one point and all other points

If I calculate all possible semivariance for point A (red dot), I get 11 points
A

Repeat for Semivariance calculations for every possible pair

Much more semivariance computations if we repeat for all possible pairs

If all pairs are considered, we get a Semivariogram cloud

Distance

Semivariogram allows us to investigate spatial dependence If there is spatial dependence points that are closer together will have small semi-variance and vice versa Because all pairs are plotted on the semivariogram, the semivariogram becomes difficult to manage/interpret

Binning is the solution

A process that is used to average semivariance data by distance and direction
First group pairs of sample points into lag classes. If lag size is 2000 meters, lag classes are: (<2000, 2000-4000, 4000-6000, etc) Second group pairs by direction (e.g., using grid cells) Each grid cell will have pairs with similar directions & distance

Continue

Where:
N

= average semi-variance between sample points separated by lag h = is the number of pairs of sample points sorted by direction in the bin

Third: For each group of pairs with similar lag and direction (in the same grid cell), we compute the average semivariance between sample points separated by lag h

Semivariogram after binning

Same lag, different directions

On a semivariogram, if spatial dependence exist, semivariance is expected to increase with distance

Wind direction Plume

Anisotropy

Pollutant concentration

high

low

Anisotropy is a term describing the existence of directional differences in spatial dependence From semivariogram data, directional dependence can be extracted if it exists

Continue

Fourth: To use the semivariogram as an interpolator, the semivariogram data must be fitted with a mathematical function/model

Semivariogram
Measures the variability of data with respect to spatial distribution Looks at variance between pairs of data points over a range of separation scales Forms the basis of every geostatistical study

Semiveriogram conditions:
Stationarity
The entire dataset can be described with one statistical model Under the condition of stationarity, you can see the distance of correlation in your data
Correlated at any distance

(h)

Correlated at a max distance

Semiveriogram types

Spatial dependence levels after a certain distance

Spatial dependence Decreases exponentially

Many functions to choose from; Geostatistical analyst provides 11 models Examples include: Gaussian, Linear, Spherical, Circular, and Exponential Popular models are the spherical and exponential

Semivariogram: Model Parameters

C0 &C1

C1

C0 a
(h)

Nugget (C0): Semivariance at distance 0: Represents microscale variations or measurement errors Range (a) : Distance at which semivariance levels off: Represents the spatially correlated portion Sill (C0 + C1): Semivariance at which leveling takes place N.B.: Every data set will have unique model features (model parameters)

Extract model parameters

C0 &C1

C1

C0 (h) a

Fifth: Extract model parameters (C0, C1, and a) for the dataset under consideration from the semivariogram. This will allow us to calculate a modeled semivariance between any 2 points knowing the distance h separating the two points.

Continue
Sixth: Use the model parameters to extract semivariance values for any point by solving a set of simultaneous equations

Let us consider the following Problem: Knowing three points: 1, 2, 3, we want to use krigging to estimate point 0

How?
Problem: Knowing three points: 1, 2, 3, we want to estimate semivariance values and solve equations to determine value at point 0

Where: (hij) is the semivariance between known point I & J (hi0) is the semivariance between known point I & point to be estimated Lagrange multiplier to ensure minimum estimation of error
More like a fudge factor

Thus
Derive semivariance between known pairs of known points (1,2,3) and each of the known point and unknown point knowing model parameters Derive weights solving by solving the simultaneous equations Plug in weights in the equation below to calculate the value of the unknown point (Z0)
Z0 = Z1W1 + Z2W2 + Z3W3

What if we had > 3 points

(1) Apply Matrix algebra to solve simultaneous Equations

What if we had > 3 points

(2) Estimator equation becomes

Where:
= Estimated value = Known value at point x = Weight associated with point x
S

= number of sample points used in estimation

How is this different from Inverse Distance Method

Krigging Calculation of weights involve (1) (2) Measuring reliability Variance between point to be estimated and Known points Variance between Known points (1) IDW Variance between point to be estimated and Known points

A measure for the reliability of the estimate

None

Kriging Types in ArcGIS

Ordinary kriging Simple kriging Universal kriging Indicator kriging Probability kriging Disjunctive kriging Cokriging

From before

Spatial variation consists of 3 components Random spatially correlated component A trend component Random non correlated component (error term)

Components of Kriging

The value of z depends on: (1) trend component, (2) random autocorrelated component, and (3) random non-correlated component (for simplicity not represented in figures)

Z(s) = (s) + (s)

Where:
Z = Value at point s = Trend component value at point s (first order or second order polynomial) = Random, autocorrolated component

Ordinary Kriging

Assumes there is no trend

Assumes m(s) is unknown and constant

Focuses on the spatially correlated component

Simple Kriging

Assumes (s) , the mean of data set is known and is constant Assumes there is no trend component In the majority of cases this is unrealistic assumption

Indicator Kriging

(s) is constant, and unknown Values are binary (1 or 0) Example, a point is forest or non forest

Universal Kriging

Assumes z values change because of a drift (trend) in addition to autocorrelation. (s) is not constant Trend component expressed as a 1st order (plane) or 2nd order polynomials (quadratic surface) Kriging is performed on residual after the trend is removed

Cokriging
Adds second variable which is correlated with the primary function It assumes the correlation between the variables can be used to improve the prediction of the primary variable Example better kriging of precipitation can be done if elevation is included as a secondary variable

Kriging outputs & assumptions

Prediction map Prediction map reports the mean Assumes a normal population at each point & thus,

Kriging outputs & assumptions

5%

35%

50%

Probability maps show the probability of finding a value at a location(e.g., > threshold value of 1) Assumes a normal population at each point

Kriging Outputs

Quantile map: Quantile map showing the probability for each location for having an extreme value (highest 5%) Assumes a normal population at each point

Kriging outputs
Error maps refer to how good we can map these predictions

Error Assesment

How can we choose the most appropriate method for interpolation?

Cross validation:
(1) (2) (3) (4) (5) Remove a known point from data set Use remaining points to estimate the value at the point removed Compare the estimated to known value Repeat for all points & calculate root mean squares The method that produces the least difference is selected (closer to 1)