Kriging Interpolation on Archaeological Depth Values

Question

Hi there,
  I am using interpolation to create an original surface based on depth measurements at various points across the site. There is an overall trend in the data from NW to SW, so I originally used Ordinary Kriging with a first order trend removal. Now I am wondering if I should have used Universal Kriging. Which method would be more suitable for the dataset? In terms of slope, it is a change of about 30 cm over 7m, so not very dramatic.

Also, when you are choosing which model to apply to the variogram, how do you choose? Other than trying to fit the line to the point distribution, is it wise to choose the one with the lowest RMSE value?

Any suggestions are greatly appreciated!

MarcoBoeringa · Answer

I tried the simple kriging model as you suggested, but can I assume that the mean/trend is known and constant for my data?

Thanks again!
Kirsten

This is something only you can answer with your expert and field knowledge of the terrain / research site. There are statistical tools in Geostatistical Analyst that will tell you whether some assumptions are more or less reasonable or appropriate based on statistics, but in the end it is you who needs to decide which models or assumptions are best for your data.

KirstenFalzarano · Answer

Thanks for the suggestions!
   If possible, could you explain to me the difference between simple, ordinary and universal kriging as basic concepts? I've read through all of the help file info and the statistical explanations, but I'd like to try and  relate the models to my data so I can rationalize which model I choose in my dissertation. I tried the simple kriging model as you suggested, but can I assume that the mean/trend is known and constant for my data?

Thanks again!
Kirsten

Eric6346 · Answer

Unless you have a good reason to prefer a particular model (for example, expert knowledge about the physics of these depths), a general rule is to choose the model that gives the lowest root-mean-square.  However, you need to also check the other model diagnostics too, particularly the RMS-Standardized and the Normal QQPlot. Try lots of models with different variogram models, and choose the one that looks best.

From my experience, a combination that works quite often is Simple kriging with a Normal Score Transformation and K-Bessel semivariogram.