Skip to content

What is the reason for Very high value by Universal Kriging based on Nested 3D varigram

3 messages · Bingwei Tian, Edzer Pebesma, Jon Olav Skoien

#
Dear list,

I am doing a 3D estimation of logtransfered subsurface temperature (with a
strong vertical trend) with a Nested 3D variogram, but the results show very
high value over than origin data.
This is not normal and absolutely wrong if I back transfer logged data.

I attached the 
<http://r-sig-geo.2731867.n2.nabble.com/file/n7587514/vgm.png>  and anyone
who knows what is the reason for the very high value differ from origin
data?
Or what kind of processing I should to do for the data back transfer? Thanks
in advance for any help.

Data:
summary((spdf$logt))
          Min.        1st Qu.         Median           Mean        3rd Qu. 
0.741937344729 3.487375077900 3.908014984030 3.952886346280 4.452019006490 
          Max. 
5.733988316710 

Model: Nested 3D varigram
uk.eye1  <- vgm(psill = 0.155,  model = "Gau",  range=700,  nugget=0)
uk.eye   <- vgm(psill = 0.125,  model = "Sph",  range=35000,  nugget=0, 
add.to=uk.eye1)
  model psill range
1   Nug 0.000     0
2   Gau 0.155   700
3   Nug 0.000     0
4   Sph 0.125 35000
UK: 
logt.uk <- krige(log(t)~z, spdf, grid, model = uk.eye, nmax = 20)

Result:

summary((logt.uk$var1.pred))
          Min.        1st Qu.         Median           Mean        3rd Qu. 
-1.66562650678  3.30346488250  3.76836777085  3.81376070431  4.24457939254 
          Max. 
15.05945622140 



-----
Bingwei

Ph.D. Student 
Kyoto University
C-1-2-225, Katsura Campus, Kyoto University, 
Nishikyo-ku, ?615-8530, Kyoto, Japan
--
View this message in context: http://r-sig-geo.2731867.n2.nabble.com/What-is-the-reason-for-Very-high-value-by-Universal-Kriging-based-on-Nested-3D-varigram-tp7587514.html
Sent from the R-sig-geo mailing list archive at Nabble.com.
#
You don't describe your problem completely, and you also do not provide
a reproducible example, so I can only make a wild guess: your z variable
takes an extreme value at the prediction location(s) where you see the
high predictions.
On 12/04/2014 08:19 AM, Bingwei Tian wrote:

  
    
#
It is difficult for us to know exactly what happens as you are not 
providing a reproducible example. But my guess would be that this is 
related to your use of a Gaussian variogram without nugget for the first 
part of the variogram. This can cause rather large weights (positive and 
negative) if you have observations in close vicinity of each other, 
leading to predictions far away from your range of observations. See if 
a small nugget effect (~0.01 or less) can solve the problem, or if you 
can change to a different variogram model.

Cheers,
Jon
On 12/4/2014 8:19 AM, Bingwei Tian wrote: