[R-sig-eco] [R-sig-Geo] Need an algorithm to identify spatial clusters

[Dan Putler]

Hi Mark,

Is your data on a raster grid? If yes, could it be converted to a vector
point set? If you can get it into a vector point set then DBSCAN might
solve your problem. The remaining issue is the choice of a cluster
validation method to determine the number of clusters. I have developed
one that seems to work well for two dimensional data (which is your
case), but is still under development. If you have an interest, let me
know, I can send code along.

Dan

On Tue, 2009-04-14 at 10:26 -0600, Mark Andersen wrote:
> Hello, all,
> 
> I am currently interested in simulating the initiation and spread of 
> forest fires in simulated landscapes consisting of 2 tree species (a 
> native and an invasive). I'm simulating fire seasons of fixed length; 
> fires can start at any time and in either tree species during this fire 
> season, and spread through the simulated forest according to a simple 
> percolation-like process. I have developed the algorithms for this in 
> Matlab, and don't anticipate any problems converting them to R if I need 
> to.
> 
> However, it's the next task that has me at a loss. At the end of each 
> fire season, there will be several burned patches on the simulated 
> landscape; the number of these patches won't necessarily correspond to 
> the total number of fires that season, because fires may coalesce. I 
> need to count how many of these burned patches there are, and find the 
> size of each. It seems that there should be an algorithm available to do 
> this, but I haven't found one, possibly because I haven't been using the 
> right search criteria. I used RSiteSearch("spatial cluster"), and got 
> many references to the spatclus package. However, from the spatclus 
> documentation it appears that this is not quite what I'm looking for. 
> Spatclus looks like it picks out potential clusters of occurrence of 
> some phenomenon (such as earthqukes or cancer); I need something that 
> will delineate clusters of contiguous pixels with the same value (i.e., 
> burned in my simulations). Could anyone point me towards such an 
> algorithm, or help me get started finding one?
> 
-- 
Dan Putler
Sauder School of Business
University of British Columbia