Page 1 of 1

[SOLVED] Strange outputs for kriging

PostPosted: Fri Jun 16, 2017 8:43 am
by dimitridor
Hello,

I'm trying to krige my multivariate data set with the following command:

grid <- kriging(dat2,grid,vario.model.2,moving.neigh)

But I receive a bunch of this kind of outputs:

[24,] -5.462 -1.744 -1.302 -2.224 -6.357 -1.937 -2.565 -1.700 -1.186 -1.476 -1.386 -1.268 -4.691 -12.732 -2.008 -2.159 -3.164 -5.543 -11.743 -60.502 -23.969 -6.524 -5.465 269.194 -54.054 -4.084 -6.767 -3.450 -1.891 -1.762 -0.470 -0.564 -0.880 -1.448 -2.751 0.153 0.098 -0.387 -0.012 0.362 0.288 0.284 -0.051 -0.243 -0.664 0.227 -0.398 -0.486 -1.359 -2.107 -0.184 -0.099 -1.070 0.157 0.095 -0.030 0.038 0.125 -1.192 -0.395 -0.651 -0.301 -0.415 -0.483 -1.806 -1.203 -0.898 -0.923 -0.068 -0.127 -0.072 0.110 -0.065 0.057 0.188 0.184 0.193 0.134 0.127 0.122 0.194 0.248 0.128 0.173 0.151 -0.586 -0.287 -0.434 0.109 0.149 0.154 0.127 0.065 -0.029 -0.077 -0.806 -1.822 -1.398 -0.569... <truncated>

At the end of some lines, I get:
Error in matrix inversion (rank=1023) : Pivot #63 is null

Nicolas suggested me it could be due to a singularity problem and I should search for duplicates. I did it with:

dat2=duplicate(dat, dist=100.0)

But I have still the problem, even when I choose large values for dist.

Did I made a mistake somewhere? What is the meaning of the strange outputs?

Thank you for your help,

Dimitri

Re: Strange outputs for kriging

PostPosted: Mon Jun 19, 2017 8:33 pm
by Didier Renard
As we discussed over a private mail, the problem here is that the model is not correct:
your variables are linked (they are probably compositional).
Then it is well-known that the kriging matrix is singular.

You must use a specific algorithm to tackle this type of variables, by considering N-1 variables for example.

Re: Strange outputs for kriging

PostPosted: Wed Jun 21, 2017 4:40 pm
by dimitridor
Thank you, Didier.
My variables are Sand, Silt and Clay contents in soils and are compositional of course.
Now, is there any possibility to estimate e.g. Sand content only using Silt as covariable? I didn't found out how to do that.
Many thanks for your help,

Dimitri

Re: Strange outputs for kriging

PostPosted: Wed Jun 28, 2017 9:30 pm
by Didier Renard
Hi Dimitry

Your proposal corresponds to a special type of Factorial Kriging Analysis applied in the domain of compositional variables, where the target variable is estimated starting from all the other variables.
As a matter of fact, you avoid the problem of singular matrices (linked to the closure of the variables) ... but:
- it seems strange to ignore the information of the target variable (but this is not an issue in the isotropic case, as knowing all other variables is equivalent to knowing the target one)
- it breaks the symmetry between all variables.
- it needs a precise set of equations to deal with the universality conditions...

So, unfortunately, there is no function provided which codes this (strange) cokriging. But it could be easily coded if the technique is demonstrated to solve the problem of compositional variables.
On my side, I prefer an algebraic solution allowing the solve a matrix, even if it singular: i.e. partial inversion (SVD,...)

Re: Strange outputs for kriging

PostPosted: Mon Jul 08, 2019 9:58 am
by tme
Hello,

I also encounter this strange behavior. Generally speaking, does this problem occur from the very beginning of the data modelling or later on at the variogram step ? From what I understood, the variogram is computed based on the z values in a db. I only have one variable in my case. But many duplicate cases, as events occur in the same place (but not at the same time). Is this singular matrix problem due to duplicates ?

My R session is totally frozen while kriging.

Thanks for your feedback

Re: Strange outputs for kriging

PostPosted: Tue Jul 09, 2019 7:52 am
by Fabien Ors
Dear,
If you have many duplicates (several samples are located at the same coordinates), you must use the 'duplicate' function as suggested above.
Hope this helps