Page 1 of 1

[SOLVED] units problem for variogram automatic fit

PostPosted: Wed Mar 07, 2018 11:15 am
by bez
When making an automatic fit to a variogram with units say in e-07, the automatic fit provides non relevant outputs.
when I change the units of Z and thus those of the variogram to get "reasonnable" units, says between 0 and 1, then the automatic fit is correct.

Re: units problem for variogram automatic fit

PostPosted: Fri Mar 09, 2018 4:34 pm
by Nicolas Desassis
Hi Nicolas,
Thank you for your feedback. When you talk about units, is it spatial units or variable units?

Nicolas

Re: units problem for variogram automatic fit

PostPosted: Tue Mar 13, 2018 4:46 pm
by bez
Hi Nicolas,
units of the variables
A+
Note : I had sent you an email with data and script to illustrate the problem.

Re: units problem for variogram automatic fit

PostPosted: Sun Mar 18, 2018 9:57 pm
by Didier Renard
When changing the scale of the variable of interest (say dividing it by 1000), the variance (and hence the units along the variogram axis) are divided by 1,000,000.
Then model.auto procedure must follow the same scale change.
If we consider 'diff' the gap between the experimental variogram and the model; during the iterations, if 'diff' becomes smaller than 'tolstop' the iterations stop. It is considered that the convergence has been reached.
Unfortunately this parameter 'tolstop' is not scaled as a function of the sill or the variance of the data. Then the user must adapt the value of 'tolstop' to the order of magnitude of the variance of the target variable.

In your case, due to very small variability of the data, the value of 'diff' was considered as smaller than 'tolstop' after a very limited number of iterations (say 2). The program stops while the fit is not correct.

The workaround is simply to reduce the value of 'tolcode'. Put it to 1.e-10 for example.

Re: units problem for variogram automatic fit

PostPosted: Mon Mar 19, 2018 9:49 am
by bez
Hi Didier
Fair enough. I understand the problem.
I did not pay attention to the fact that tolstop was in absolute distance and not relative to the variance.
Wouldn't it be more relevant to set the tolstop limit in terms of % of the variance ?
It would then be more robust to various situations.
Nicolas