topic Re: Kriging Problem! in ArcGIS GeoStatistical Analyst Questions

Kriging Problem!

jieding — Sat, 30 Apr 2011 16:14:04 GMT

I used the same rainfall data to do ordinary kriging in arcmap 9.3 and arcmap10.
In arcmap 9.3, the default parameters give a better predition map than the optimized parameters in arcmap10.
Does anyone know why this happens?

Re: Kriging Problem!

SteveLynch — Mon, 02 May 2011 11:35:17 GMT

How do you know that the predictions are better?

Thanks
Steve

Re: Kriging Problem!

Fevziakbas — Mon, 30 May 2011 10:59:30 GMT

Hi Steve,

Could you please a brief explanation about "optimize model " in Arcgis 10 geostatistical analysis.

Is it sufficient to choose this option to provide best map to the user.

Re: Kriging Problem!

Fevziakbas — Mon, 30 May 2011 12:08:40 GMT

Hi Steve;
here is an example.
Surface soil clay content semivariogram modelling and cross validation result.
Model :spherical
Model -------------     Mean Pre. Err.-------- Root Mean. Sq. -----    Root M sq Stand.
1 (optimized by program)     0.03453 --------   7.521 ------------ 1.036
2 (my semivariogram model)   0.0556 --------- 7.437 ------------ 1.0741

Which one I should choose?

Re: Kriging Problem!

EricKrause — Tue, 31 May 2011 14:14:24 GMT

Hi Steve,

Could you please a brief explanation about "optimize model " in Arcgis 10 geostatistical analysis.

Is it sufficient to choose this option to provide best map to the user.

In kriging, the optimize button uses weighted least-squares. The algorithm iterates through the range and fits the nugget and sill using weight least-squares for each iteration. It then calculates the combination with the minimum root-mean-square. Because the covariance matrix between the three parameters is not estimated, this process will not necessarily find the global root-mean-square minimum. If you play with the parameters enough, you may find a better root-mean-square.

Looking at your cross-validation statistics, you can justify either model. Your model has a better root-mean-square, but it has a worse mean prediction error and rmse-standardized.