R algorithm/package for creating spatial

Tue, May 11, 2010 12:21 AM

Dear Laura,

 > Given your comments below, I was was wondering what distribution and
 > autocorrelation model you think generally best represents naturally
 > occurring environmental variables.

It depends on the underlying causes of the autocorrelation. If it's an 
intrinsic process like dispersal, something exponential is likely, or 
possibly something with a fatter tail - there's a large literature on 
dispersal kernels to explore. If it's some thing geological, geographers 
have concluded a Matern model is often the most appropriate. If it's 
some influence of man, this often has an extraordinarly steep structure 
(think clear-fell - sudden changes in the environment) that could be a 
step, or some extreme Matern. If it's topology, you're probably stuffed 
- hugely non-stationary patterns here: think long mountain chains with 
high variability, then large flat plains (we provided some code for 
simulating non-stationarity in the sup info (free) our Ecology Letters 
paper http://www3.interscience.wiley.com/journal/123241231/abstract, but 
you really need to think hard). Essentially, there's no substitue for 
thinking something about the ecology and real world that you want to 
simulate... (Why not start by measuring the actual structures and 
distributions of the covariates you want to simulate?)

Colin

Laura S wrote:

Dear Colin.

Thank you very much. I greatly appreciate your time and comments.

I am still interested in creating a landscape with environmental values 
that follow a uniform frequency distribution, but I will keep the 
caveats of strange autocorrelation in mind. I agree this representation 
of variables is highly unrealistic.


Thank you again for your time and consideration,
Laura



On Mon, May 10, 2010 at 11:16 PM, Colin Beale <cb751 at york.ac.uk 
<mailto:cb751 at york.ac.uk>> wrote:

    Hmmm, that's not a trivial problem. package geoRglm has functions to
    generate Poisson and Binomial surfaces. RandomFields is remarkably
    flexible. But there are good reasons why a uniform distribution is
    challenging to produce from a given spatial structure (think of
    'natural' autocorrelation as waves on the sea - there are only ever
    going to be a few high values and a few low values). So I don't know
    of any packages that do this 'off the shelf' - it really doesn't
    correspond to many real landscape covariates anyway, so are you sure
    you want to simulate such a landscape? If you are, I think I'd
    generate my uniform distribution using runif(), and then allocate
    values from this vector to some spatially autocorrelated structure
    generated with RandomFields using sort and order - something like:

    library(RandomFields)
    #Generate a sorted vector of uniform distribution of required length:
    my.unifdist <- sort(runif(20*20))
    #Generate a landscape of autocorrelated values:
    my.landscape <- GaussRF (1:20,1:20,param = c(0,5,1,10),
                 method = "cutoff CE", model = "exponential", grid = T)
    #Place the sorted vector in the rank order of the landscape values:
    my.uniflandscape <- matrix(my.unifdist[rank(my.landscape)], ncol
                        = 20)
    #Have a look at the landscapes:
    par(mfrow = c(1,2))
    image(my.landscape, main = "Gaussian landscape")
    image(my.uniflandscape, main = "Uniform Landscape")

    It's autocorrelated, and has a uniform distribution, but has a
    rather warped autocorrelation structure. Are you sure this is the
    sort of landscape you want to simulate?

    Colin



    Laura S wrote:

        Dear all:

        I would like to create a landscape of environmental values that
        follow a
        uniform frequency distribution and also have spatial
        autocorrelation in the
        landscape. I was wondering if there is an algorithm and/or
        package out there
        that creates autocorrelation of values that are distributed
        according to a
        non-normal frequency distribution.

        Any suggestions are greatly appreciated.

        Thank you,
        Laura


    -- 

    Dr Colin Beale
    Research Fellow
    University of York

    Mob. +255 (0) 684 059 123




-- 
" Genius is the summed production of the many with the names of the few 
attached for easy recall, unfairly so to other scientists"

- E. O. Wilson (The Diversity of Life)

Dr Colin Beale
Research Fellow
University of York

Mob. +255 (0) 684 059 123

R algorithm/package for creating spatial

Thread (2 messages)