Latest Contributions by EricKrause

If it gives better cross-validation, then that's a good justification for using the model.  Ultimately, you need to justify that the model fits your data the best, and cross-validation is a good way to do that. ... View more

You can compare against Universal kriging if you want, but personally, I would just use Simple kriging.  Build two Simple kriging models, one with trend removal and the other without.  See if removing the trend gives better cross-validation. ... View more

If you do your trend removal well, you shouldn't see a big difference between Universal and Ordinary/Simple kriging.  As a general rule, we suggest using Simple kriging because it supports Normal Score Transformations. The difference between Ordinary/Simple and Universal is subtle and a bit confusing.  The trend removal page is essentially just changing your input data by removing the trend that you fit.  Then the semivariogram is fit to the detrended data.  The difference between Ordinary/Simple and Universal is that Universal goes back and re-fits a global trend model to the data that has already been detrended (it has to do this for the Universal kriging equations to work).  If you did your trend removal well, there should be very little global trend left to re-estimate, and Ordinary/Simple and Universal kriging should give very similar results. ... View more

I've experienced this slow-down when doing data frame clipping.  The issue for me was that the polygon had way too many vertices. If you use the Simplify Polygon tool on the clipping polygon, it should start working quickly again. ... View more

We do not have spatio-temporal kriging in ArcGIS, so you won't be able to model the temporal correlation directly in the kriging. However, we have the Densify Sampling Network tool that suggests new locations of monitoring stations based on (among other things) kriging standard errors.  The tool also can take an optional weight raster where you can weight certain locations higher than others.  Though I don't have any specific recommendations about how you would do it, you may want to try to model this weight raster based on temporal trends in the data. ... View more

Once you have the fishnet polygons, you can use the Subset Features gp tool to get a random sample of a particular size. Note: Subset Features will make a new feature class of the randomly sampled polygons.  Depending what you want to do with the sample, this may be an advantage or a disadvantage of the tool. ... View more

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. ... View more

The software will automatically ignore polygons with missing data.  If the polygons are coded correctly, you shouldn't have to do anything. Also, be careful with extrapolation.  Areal interpolation is based on simple kriging, so it is inherently bad at extrapolation.  As you move away from the source polygons, the predictions will converge to the mean value of the polygons. Read the two help topics we have to areal interpolation.  The workflow topic shows a Rate (Binomial) example, but the workflow is basically the same for all three data types: http://resources.arcgis.com/en/help/main/10.1/index.html#//0031000000q8000000 http://resources.arcgis.com/en/help/main/10.1/index.html#/Using_areal_interpolation_to_perform_polygon_to_polygon_predictions/0031000000qm000000/ ... View more

The formula for Binomial (Rate) areal interpolation is in this paper (it's free to download): http://www.sciencedirect.com/science/article/pii/S1878029611000053 It accounts for varying population sizes by doing a correction to the empirical semivariogram so that polygons with larger populations exert more influence on the model. ... View more

The best analysis method depends on a lot of things, particularly what kind of data you have.  My first thought is Areal Interpolation, which is available now in ArcGIS 10.1.  However, if you have a lot of covariates, you'll probably have better luck with a regression model. ... View more

These are the two courses designed to help you learn interpolation with Geostatistical Analyst (Dan already posted the first one): http://training.esri.com/gateway/index.cfm?fa=catalog.webCourseDetail&courseid=2052 http://training.esri.com/gateway/index.cfm?fa=catalog.webcoursedetail&courseid=2128 If you do those two courses in that order, you should have a pretty good idea how to perform interpolation in ArcGIS.  There's only so much we can teach in 3 hour web courses, but we got in quite a lot. You might also find this one useful: http://training.esri.com/gateway/index.cfm?fa=catalog.webCourseDetail&courseid=2053 ... View more

Aye, this is embarrassing.  It's been so long since I used Indicator/Probability kriging that I forgot exactly how they work. For binary data, it doesn't make sense to do Probability kriging, and there's a bit more to Indicator kriging than I remembered.  Set the Threshold value to 0.5, then after creating your surface, areas with a predicted value below 0.5 should be classified as a "0" and any value above 0.5 should be classified as a "1".  You can do this by symbolizing the geostatistical layer, or you can convert it to raster and use the Con tool in Spatial Analyst. ... View more

If you've done the Geostatistical Analyst tutorial, you shouldn't have any problems performing Indicator or Probability kriging.  The process is the same.  As for kriging on ordinal data, I know there has been research done in this area, but I'm not up-to-date, and we don't have that capability in Geostatistical Analyst. ... View more

Indicator kriging and Probability kriging are both designed to work with binary data.  Choose kriging and supply your binary dataset on the first page of the Geostatistical Wizard.  On the second page, choose Indicator or Probability from the kriging types, then proceed as usual.  The prediction surface will represent the probability that the binary variable is a 1. http://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#//00310000004n000000.htm http://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#/Understanding_probability_kriging/00310000004r000000/ ... View more

http://getthegistofit.blogspot.com/2012/04/pollution-exposure-risk-in-washington.html Give that a read.  They start with Kernel Density, then they do post-processing on the results.  I can't really comment on the legitimacy of the methodology, but it's a good place to start. ... View more