Skip to content

how to code for neighbour list editing

3 messages · James Rooney, Roger Bivand

Original
James Rooney
# 
Hi all,

Ok so I have a spatial data set with over 3000 polygons. I used poly2nb() to identify neighbour relationships and make an nb object. Of the polygons, about 30 or 40 are islands and have no neighbours according to polynb. Prior to doing Bayesian smoothing, I wish to assign neighbours to those islands (and probably weight them with a weaker weighting than true neighbours). Anyhow I've tried the edit tool for manually editing neighbours. Its basically not useful however since there are so many polygons its too hard to see the islands.

So I'd like to write some code to do something like this by manipulating the nblist:

1) Identify islands
2) For each island find nearest neighbour
3) Assign a neighbour relationship between the island and its nearest neighbour
4) Weight it appropriately (half normal weighting for example)

My problem is I don't really have any understanding of the nb object or how it stores relationships. I've tried summary() attributes() print() etc etc and can't really glean an insight into how to access or edit the inner structure of the nblist.

I was hoping someone might have dealt with this before. Or maybe there are functions for this I have yet to discover ?

Many thanks,
James
Roger Bivand
#
On Mon, 15 Jul 2013, James Rooney wrote:

            

      
      
      
    

        
Please provide example code and data to set up the problem, say US 
counties or similar:

http://www2.census.gov/geo/tiger/GENZ2010/gz_2010_us_050_00_20m.zip

library(rgdal)
usc <- readOGR(".", "gz_2010_us_050_00_20m")
library(spdep)
nb <- poly2nb(usc)

            

      
      
      
    

        
no_neighs <- which(card(nb) == 0) # returns set cardinality.

To check that you get the same result using counts of graph components:

nc <- n.comp.nb(nb)
table(nc$comp.id)
no_neighs_comps <- which(unname(table(nc$comp.id)==1))
match(no_neighs_comps, nc$comp.id)

            

      
      
      
    

        
k1nb <- knn2nb(knearneigh(coordinates(usc), k=1,
   longlat=!is.projected(usc)))

It isn't just a matter of finding the one-way neighbor, as the reverse 
should also be assigned (if i neighbours j, then j neighbours i, as in 
poly2nb()).

            

      
      
      
    

        
is.symmetric.nb(nb, force=TRUE)
nb1 <- nb

res <- k1nb[no_neighs]
nb1[no_neighs] <- res
attr(nb1, "sym") <- is.symmetric.nb(nb1, force=TRUE)
# re-assign the now incorrect symmetry attribute
nb1

nb2 <- make.sym.nb(nb1)
is.symmetric.nb(nb2, force=TRUE)
nb2

I think that some new links here have contributed to changes in the 
smaller subgraphs:

table(n.comp.nb(nb2)$comp.id)

            

      
      
      
    

        
This is harder to arrange, because of the symmetric linkages, and although 
it could be done, is unlikely to affect the result. Indeed, judicious use 
of set.ZeroPolicyOption(TRUE) to avoid having to set zero.policy=TRUE 
might be quite acceptable if there are few no-neighbour observations.

            

      
      
      
    

        
Use a smaller example to get on top of it and read the vignette in spdep 
as well as the help pages and their references. An nb object is a list of 
integer vectors, but uses an integer vector with a single value of 0 to 
signal no neighbour.

Hope this clarifies,

Roger

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      James Rooney
    

    

      
      #
    

  


  

      
Many thanks Roger,

I'll study that and see how I get on!

James