Latest Contributions by EricKrause

There's a typo in that pdf that I just noticed.  For average standard error, the formula is missing a square.  You can find the correct formulas here: http://resources.arcgis.com/en/help/main/10.1/index.html#//00300000000z000000 "Average Standard Error" is the only formula that is different than you might expect.  It might be better called "Root-Mean-Variance."  We used this formula instead of a simple average because this formula is more directly comparable to the RMS. ... View more

http://dusk.geo.orst.edu/gis/geostat_analyst.pdf The formulas for the crossvalidation summary statistics can be found in Appendix A on page 273 (page 279 of the pdf, since the first six pages aren't numbered). ... View more

If you really need a surface, you might try IDW or Natural Neighbors.  I definitely would not try to do kriging.  However, know that no matter which interpolator you use, the results will probably have large errors. ... View more

I took a look at your data, but I wasn't able to fit a good model.  There are several issues I saw that make it difficult to fit a model.  First, it looks like the variable to interpolate (NReclass) is a categorical classification.  There are many repeat values: more than half the data has NReclass = 10, and another quarter of the data has NReclass = 15.  With so many repeated values, the distribution is extremely right-skewed, and none of our transformations can handle such extreme skewness and repeated values. Also, Moran's I with a Fixed Distance Band gives a p-value of just over .05, which means that the spatial autocorrelation is fairly weak. If you have the original data (before the reclassification), it may give you better results. ... View more

I talked this over with a couple people, and we're not completely sure why this is happening.  It is probably a data-specific phenomenon where adding more points does not actually add more statistical information.  Because we work with searching neighborhoods, you may be filling up the neighborhood with duplicate information and losing relevant information beyond the search window. In general, more data produces better predictions, but you seem to have hit a rare exception to that rule. ... View more

With the exception of areal interpolation in ArcGIS 10.1, geostatistical layers can only predict to points.  If you supply a polygon feature class to predict to, it will predict the value at the polygon centroid. To predict the value in a polygon, convert the geostatistical layer to raster (via GA Layer to Grid gp tool), then use Zonal Statistics as Table.  You'll then need to join the table back to the polygons. ... View more

Update: The topic has been updated with a better example of when a lower RMS does not indicate a better model: http://resources.arcgis.com/en/help/main/10.1/index.html#//0031000000q0000000 ... View more

http://dusk.geo.orst.edu/gis/geostat_analyst.pdf The formulas for the semivariogram models are in Appendix A, page 256 (page 262 of the pdf).  Each model corresponds to a different underlying physical process. K-Bessel and Stable often produce the best results because they take an additional shape parameter that allows the models to change curvature while still maintaining the same nugget, range, and sill. ... View more

Despite its name, MWK is not actually an interpolation method, so you can't perform crossvalidation.  It is designed to calculate local semivariogram parameters (nugget, range, sill) at locations you specify.  There is an optional prediction raster output, but that is using Local Polynomial Interpolation.  The model source parameter is typically a Kriging layer that you created in the Geostatistical Wizard.  It needs this layer so that it can pull things like semivariogram model type (Spherical, Exponential, etc).  You can learn how to make a geostatistical layer in our tutorial. However, in ArcGIS 10.1, we have Empirical Bayesian Kriging.  While it's not exactly a moving window, it builds local, overlapping models, so it has many of the advantages of MWK. ... View more

Also, by absolute, just ignore the negative integers right. Sorry, I missed this earlier.  In the Field Calculator, there's a function called Abs().  That takes the absolute value of a field.  Negative values become positive, and positive values don't change.  For example, if you had three values, (-2 , 5, -7), taking the absolute value would result in (2 , 5 , 7). ... View more

That table looks correct now. About iterating the workflow many times, you cannot automate the Geostatistical Wizard, so there's no way to fully automate this workflow.  You could certainly write a Python script tool or a model in Model Builder to calculate the MAE once you've created the kriging layer. ... View more

That table has the correct fields, but all the error values of zero are a bit unusual.  Did you predict back to the same data that you used to build the model?  Validation really only makes sense when you validate on points that were not used in the interpolation.  That is why Subset Features creates two different outputs: one for interpolating (the "training" features) and one for validating (the "test" features).  "Take the average of those values in the 'error' column (the 5 originally omitted), repeat a 100 times say, to give a distribution of MAE? " Remember to take the absolute value of the errors before taking the average.  You want the mean absolute error, not the mean error. ... View more

That table has the correct fields, but all the error values of zero are a bit unusual.  Did you predict back to the same data that you used to build the model?  Validation really only makes sense when you validate on points that were not used in the interpolation.  That is why Subset Features creates two different outputs: one for interpolating (the "training" features) and one for validating (the "test" features). ... View more

The problem is that trend, autocorrelation, and anisotropy can all look the same.  Even if you see something that looks like a trend, you may be able to account for it through autocorrelation or anisotropy.  The path you choose should be based on your knowledge of the data and model diagnostics like cross-validation. As for whether to use the same methodology for the different soil layers, that's a tough question to answer.  If you use a fundamentally different model (for example, remove trend on one layer and do anisotropic corrections on another), you're implying that the underlying physics of the soil surfaces are different.  The physics of the soil surfaces may well be different, but that's a decision that you need to make as an expert in your domain. ... View more

What you're referring to is called validation.  Here's how you do it: 1.  Use Subset Features to randomly partition your data into a training set and a testing set (in the tool, you specify how many points will in each subset).  2.  Perform kriging in the Geostatistical Wizard on the training features. 3.  Use the kriging surface as input to GA Layer to Points. Predict to the testing features, and specify the field to validate on (the filed you used to interpolate).  This will create a new point feature class with validation statistics.  4.  To calculate the mean absolute error, you'll use the "Error" field in the new feature class.  Make a new field and calculate the absolute value of the error.  Then take the average of these absolute values.  To calculate the average standard error, take the average of the "Standard Error" field. ... View more