Scribd
Upload
752 views2 pages

Spatial Statistics For Dummies

The document discusses spatial statistics and variography. It defines key concepts like the variogram and semivariogram which measure the similarity between values of a regionalized variable at different locations. The semivariogram is used to model spatial dependence and estimate unknown values using kriging, a method that produces minimum mean squared error predictions based on the semivariogram. Kriging allows for predictions at any location based only on the variogram model and relative point positions. The document also discusses changing support from point data to block averages and principles of spatial sampling based on variogram range.

Uploaded by

Muttaqien Abdurrahaman AR
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
752 views2 pages

Spatial Statistics For Dummies

The document discusses spatial statistics and variography. It defines key concepts like the variogram and semivariogram which measure the similarity between values of a regionalized variable at different locations. The semivariogram is used to model spatial dependence and estimate unknown values using kriging, a method that produces minimum mean squared error predictions based on the semivariogram. Kriging allows for predictions at any location based only on the variogram model and relative point positions. The document also discusses changing support from point data to block averages and principles of spatial sampling based on variogram range.

Uploaded by

Muttaqien Abdurrahaman AR
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
You are on page 1/ 2

Spatial Statistics for Dummies The variogram.

The first law of geography is that nearby things are more similar than far-apart things. A regionalized variable Zx is some random variable Z that is defined at geographical (twoor three-dimensional) location x. For example, Z might be the amount of wood at location x in a forest (here x would be two-dimensional), or it might be the concentration of iron at a particular three-dimensional location in an iron ore deposit. In line with the first law of geography, we can measure the degree of spatial similarity of Z by the semi-variogram. This is defined for a vector offset h by:1 E[( Z x + h Z x ) 2 ] 2 The expectation is over all location x at which the regionalized variable is defined. Implicit in this is the assumption of second order stationarity: that all portions of this area are exchangeable.

( h) =

Sample semivariogram
1.50

1.25

1.00

gamma

0.75

0.50

Here is a generic semi-variogram . The maximum value that it attains is called the sill; this semi-variogram has a sill of 1.5. Lag We can loosely interpret this as the variance of the regionalized variable. The lag h at which the semi-variogram reaches the sill is called the range this variogram has a range of about 12. Note that while clearly (0) = 0, as h approaches zero, (h) may approach a strictly positive limit; this limit is called the nugget effect.
0.25 0.00 0 2 4 6 8 10 12 14 16 18

20

Things are easiest when the regionalized variable is isotropic. This means that (h) depends only on the length of h and not on its orientation. The opposite of isotropic is anisotropic. A standard approach in variographics is to estimate the semi-variogram from your collection of spatial data, and then fit a mathematical model. For example, you might find that a simple model like (h) = c1 c0 e-ah (the exponential semi-variogram) fits the sample variogram well. If this is the case, then you can estimate the three parameters c0, c1 and a and use these estimates to reconstruct an estimated variogram for any lag h of interest. Kriging. Named for Daniel G. Krige, this is a least-squares prediction methodology. For example, if we have samples at locations x1, x2, , xn, at which points we know the value of Z, we might want to predict the value of Zw at some other location w. The kriging estimator is n Z w = i =1 bwi Z xi where the kriging weights b are chosen to minimize the expected mean squared error subject to the constraint i bwi = 1, a constraint imposed to ensure unbiasedness. Set up the square matrix A of order n+1, with elements aij = (xi xj) for i, j = 1, 2, , n ai,n+1 = an+1,i) = 1, i=1, 2, , n an+1, n+1 = 0

and the n+1 component vector B with elements bi = (xi-w) i = 1, 2, , n and bn+1 = 0. Then the kriging weights are the first n elements of C, the solution to the equations AC = B (the last element of C is a Lagrange multiplier, whose value turns out also to be useful.) As a byproduct of this equation solving, you also get an estimate of the kriging variance, the expected mean squared difference between the estimate and the actual mean value of the regionalized variable at that location. The crucial feature of the kriging equations is that they involve only the semivariogram. This means that if you want to find the weights for an arbitrary pattern of known data and for prediction at an arbitrary location, all you need is the relative positions of the samples and the unknown location, and a mathematical formula for the variogram and you can set up the kriging equations. Thus once you have a model for the semi-variogram, kriging allows you to do least squares regression without actually having any observed values of your dependent. Change of support. The semi-variogram is defined for point samples. We may be interested in the content of non-point supports. For example, we may be interested in the total timber resource in W a 100 meter square of a forest, or the amount of copper in W - a particular cube of side 3 meters. Going from the statistics of points to the statistics of finite areas or volumes is called regularization. To predict the mean value of the regionalized variable within a block using observed values at sampling points is another quite simple application of the kriging equations. You just need one new concept the semi-variogram between the block W and the point samples. This is defined by the average of the variogram between each point within W and the sample xi in equations (w xi)dw (W , xi ) = wW dw
wW

Use this on the right side of the kriging equations and you get the weights for predicting the content of the block W, and also get a standard error for the estimate. Sampling precepts. Observations Z that are separated by a distance larger than the range are uncorrelated. Each sample point is the center of a zone of influence those locations separated from it by less than the range of the variogram. In principle, you can use a sample value to make predictions at any other location within its zone of influence. Beyond the zone of influence, the sample has no particular useful information except information about the global mean of the regionalized variable. This concept has implications for sampling. You will get the best predictions when every point in the field is as close as possible to some sample value, and so you should use sample locations that minimize some measure of typical closeness. A good choice is a tessellation in equilateral triangles, aka honeycomb pattern. The more intuitive choice is a square grid; this is actually not as good as the honeycomb because it leaves bigger gaps on the diagonal of your squares. Literature Noel Cressie Statistics for Spatial Data (1993) is overkill but in a way easier reading than the earth science books. A more specific, useful source is Grondona, Martin O. and Cressie, Noel (1991) Using spatial considerations in the analysis of experiments Technometrics, 33, 381-392

You might also like