Skip to content

scale semivariance befor fit.variogram

4 messages · Samuel Turgeon, Paul Hiemstra, Edzer Pebesma

Original
1 day later
Paul Hiemstra
#
On 07/17/2010 06:19 PM, Samuel Turgeon wrote:
Hi Sam,

It is possible, probably the covariance matrix has a lot of zeros 
because of rounding of your small values. Could this be linked to using 
single precision floats in gstat (edzer?)? I think there is no problem 
in 'backtransforming' your values, i.e. you only change the unit of the 
variable you are using (e.g. from milligram to microgram). Take care 
though that when predicting whilst using the fitted variogram model, 
your observations also need to be multiplied by this constant. My 
suggestion would be to multiply the data at the very beginning of the 
script, do all you analysis and at the very end do the division. You 
could also skip the division and present your results in the new unit.

regards,
Paul

  
    

      
      
    

  


  


      

    

    

      
      

        
      

    

    

      
      

        


  

        

      Edzer Pebesma
    

    

      
      #
    

  


  

      
The problem is not gstat using floats (it doesn't), it's roughly using
the normal equations to solve a regression problem -- this calls for a
sum of squared values matrix X'X, if the values then are too far apart,
as it seems in your case, even double arithmetic runs into an explosion
of numeric errors.

        
On 07/19/2010 05:44 PM, Samuel Turgeon wrote: