[GRASSLIST:5172] Re: re-classifying raster values

[Roger Bivand]

On Mon, 13 Dec 2004, Dylan Beaudette wrote:

> Hi!
> 
> I was wondering if anyone has tried to re-classify the values of a continuous 
> raster image with a method known as "natural breaks' in the ESRI world, or 
> any other statistical technique for that matter. I was hoping to try some of 
> the methods talk about on this website;
> http://www.geovista.psu.edu/grants/dg-qg/classing_epi/summary.html
> 
> Would it be possible to computer the class intervals in GRASS, or would this 
> have to be done in a stats package like R?

I've some results on this in a paper written in March using R:

http://spatial.nhh.no/papers/aag04.pdf

where the interesting bits are using stats in R to try out natural breaks
- references in the paper. The best solution to date is the R bclust()  
function for bagged clustering in package e1071, because it combines the
strength of kmeans() and hiearchical clustering, allowing a post-fitting
choice of number of classes. The same method is used in the course notes
on this page: http://spatial.nhh.no/geo304/modules-h04.html - look under
week 39. The interesting geovista reference is: Armstrong, M.~P., Xiao,
N., Bennett, D.~A., 2003. {Using genetic algorithms to create
multicriteria class intervals for choropleth maps}. {Annals, Association
of American Geographers}, 93 (3), {595--623}.

Straight-in coded natural breaks run into the combinatorical problem that 
there are millions of possible maps for each number of classes (above 
trivially small numbers like 1 or 2).

I've found bclust() very good - if there is too little data, the trick in 
the lecture notes of including a jittered replicate of the original data 
usually helps. I also like displaying an empirical cumulative distribution 
function to "see" the class intervals. R startup is getting better too. 
With a little pushing, making a tcltk display in R run from GRASS is 
doable to help choose class intervals.

A discussion on some of these cognitive issues woulkd be valuable,

Roger

> 
> Many thanks,
> 
> 

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93
e-mail: Roger.Bivand at nhh.no