The second element of importance that can be taken
into account is the distribution and the maximal absolute value of the
estimated kriging variance (which can be displayed by FUZZEKS). One must
be very careful with the estimated kriging variance, because this is
a value that states about the quality of the kriged values that
are calculated based on the theoretical variogram as well as the
estimated kriging variance is also calculated based on the
However, there are situations where this is acceptable, if e.g. the absolute values of two kriging variances may be compared in the case of the same theoretical variogram but different data sets.
Example: You have areas with an insufficiant number of data points and therefore a high kriging variance. To improve the situation, fuzzy data points - expert knowledge in the form of fuzzy numbers - are added to the input data set. The same theoretical variogram is used for both old and new data set. Then you can compare the absolute values of the two kriging variances.
A third idea is in fitting a theoretical variogram is to check the actual kriging result: If the isolines for the most possible values are not typical for this sort of parameter (judged by an expert), the theoretical variogram should be changed.
The last method to check the theoretical variogram uses cross-validation. This procedure leaves out sequentially any single point of the input and kriges a value at this location (with the investigated variogram) in order to compare real values with estimated ones. The cross-validation result is typically a number calculated by adding the squares of the differences of each of of the two values (which can be minimized by trying out different variogram types and parameters then). In the current version of FUZZEKS support for cross-validation is not implemented.
It has been found out that a decrease of the estimated kriging variance often
results in an increase of the cross-validation and vice versa.
So typically one has to find a theoretical variogram that is a compromise between the different criteria.
Remark: It can be possible that it makes no sense to use kriging (because
of the mathematical assumptions).
A sign for this can be a large nugget-effect (compared to the sill).
A large nugget-effect could mean that the range for the investigated
parameter is smaller than the
smallest distance between two points and can not be found out by using
this data set.
In these cases other interpolation methods (especially the simpler and less statistically based methods) tend to have no advantages either. In these situations, one can only get some interpolation, which does not need to have much in common with the spatial character of this parameter. You can also use kriging (FUZZEKS) then. Typically one would use a linear theoretical variogram, because for the kriging result with this, the range parameter makes no difference (for sill and nugget-effect holds the same). If one thinks of drawing the isolines by hand because no suitable method is found, kriging with linear theoretical variogram can be a pretty good method.
The next pages deal with