help with r.neighbors modification

[Dana M Slaymaker]

        I am using hyperclustered TM imagery to map deciduous
 forests in New England as part of the GAP Analysis program.  
This treatment produces a large number of spectral classes which
 we then reclass into our vegetation types with "i.infer".  I  found 
that quantifying neighborhood associations increased our accuracy 
 and have used "r.neighbors" to identify the most frequently 
occurring spectral class around each pixel.  However, I need a 
way to identify the second and third most frequent neighbors as 
well and produce percentages of each if possible.  I have tried to do 
this in the MIPS sml language without much luck, being among 
the programmably challenged, and would prefer to use GRASS 
because it's in the public domain.

        Could any of you help with a modification of "r.neighbors",
 either with multilayer output or as a set of programs applied in 
sequence.  Not only would you receive the undying gratitude of my 
department, co-authors and self, but your efforts would be fully 
credited in any subsequent publication.

Dana Slaymaker
Department of Forestry and Wildlife Management
University of Massachusetts in Amherst
dana at fnr.umass.edu
(413) 545-4853