Select to view content in your preferred language

Exploring Data & Selecting a Model

2357
2
05-07-2011 11:20 AM
MathieuCain
Occasional Contributor
I have been through ESRI's tutorials on the use of their Geostatistical Analyst tool and have done some basic online search, but am still not clear on understanding the Explore Data tools (particularly the Trend Analysis, and the Semivariogram Cloud) so as to select the appropriate model parameters.

1) Histogram only showed close to normal distribution with Log transformation
2) Normal QQPlot only showed points lying close to line with Log transformation
3) Trend Analysis?
4) Semivariogram Cloud?

How then to apply these "revelations" to a Kriging interpolation?

Without really understanding what I am doing, I appear to have gotten the best RMS value (423.6) with a Radial Basis Function.

I read from an external source that I should aim for the following:
- Mean prediction error --> close to 0
- RMS value --> smaller the better
- RMS standardized --> close to 1

My models are not looking good.

I have attached my basic outputs from Exploring the Data in the hopes someone might be able to help me understand this better. Note: I actually want to understand the decision making process here so would greatly appreciate your experience and advice
0 Kudos
2 Replies
EricKrause
Esri Regular Contributor
We just added a web course that goes through the ESDA tools in preparation for interpolation (the course is less than a week old), and in the next month or two, we'll be adding a course specifically about how to use the results from ESDA to choose the best model.

Exploring Spatial Patterns in Your Data Using ArcGIS 10:
http://training.esri.com/Courses/GADataPatterns10_0/player.cfm?c=315


Looking at your pdf, the log transformation looks appropriate, but your semivariogram cloud doesn't appear to show much spatial autocorrelation.  You should make a new field and calculate the log of the variable you want to interpolate. If the data doesn't have spatial autocorrelation, no interpolation method is going to be any good.  You should run the Spatial Autocorrelation (Moran's I) gp tool in Spatial Statistics toolbox.  Though I would suggest you first make a new field and calculate the log of the data you're analyzing, then run the semivariogram cloud and Spatial Autocorrelation with the transformed data.
0 Kudos
MathieuCain
Occasional Contributor
Thank you for your reply Eric.

Since I use ArcGIS 9 (not 10), I took the "Introduction to ArcGIS 9 Geostatistical Analyst" free virtual campus course (http://training.esri.com/acb2000/showdetl.cfm?DID=6&Product_ID=808). How does the course you suggested (albeit for a version I do not have) compare?

In your previous message you said that the semivariogram cloud in my pdf didn't appear to show much spatial autocorrelation. How did you arrive at this conclusion? What would the cloud look like if there was spatial autocorrelation? Would there be spatial autocorrelation if for instance most of the data values lay along the bottom x-axis (around y = 0), while at the same time, a number of values hovered along the top of the graph.

Thanks again for your thoughts and patience.
0 Kudos