Re: AI-GEOSTATS: Modelling anisotropy with GSTAT

[Digby Millikan]

Pei-Chun,  Question 1,  You may be interested in some of these points on lognormal kriging, as I have been involved with lognormal kriging of datasets, but not in use for unconditional simulation.  - Lognormal kriging can be very sensitive. If you plot experimental    variograms of your lognormal dataset, note that if your variograms    are well formed you are in luck, however if your variograms are very    erratic, note that your grades will be in error in proportion of the error    in your estimation of the sill.  - Yes prior to lognormal kriging you log transform the data so it has    zero skewness and is normally distributed. Note that if your data    belongs to a "three" parameter lognormal population you must also    make an estimte of the third parameter alpha, and add this to your    values before taking their logarithms.    If you plot a log probability plot of your data and it is a straight line    then you have a two parameter lognormal distribution and you can    take logarithms of the data to acheive zero skewness.    If the line drops of towards the origin you may have a three parameter    lognormal population with which you can estimate alpha from the    graph or alternatively iterativley trial different values to your population    until it has a skewness of zero i.e.                              z=ln(x+alpha)    where z is the transformed distribution which has skewness=0;  - Also note that the anti-logarithm of a number is not equal to the    logarithm of a number, so after your modelling, you cannot back    transform your data, by simply taking the antilogarithm of the values.    You will have to check a geostatistical text to see the procedure for    back transformation of data. Question 2,  I have just read Margaret Armstrongs "Basic Linear Geostatistics" and in the chapter on Structural Analysis provides three case studies, and in all case studies uses the same nugget effect for all directions, even though in two of the case studies there is variation of the nugget effect in different directions as in your case.  I would be inclined to use the omidirectional nugget effect in your case, the values are reasonably similar as in the case studies. It is possible the nugget effect varies in the different directions due to the different spacing of the data in the different directions, and the data is samples also, so may not perfectly follow the real underlying values of the actual continuous data. Regards Digby Millikan B.Eng   Geolite Mining Systems U4/16 First Ave., Payneham South SA 5070 Australia. Ph: +61 8 84312974 [EMAIL PROTECTED] http://www.users.on.net/digbym