sensitivity analysis

GRASS GRASS user
1

grassu-list'ers

Here is an interesting problem that I would appreciate some input on
before I re-invent the wheel, so to speak. I need to do a sensitivity
analysis on a spatial model to test its robustness to
mis-classifications in the underlying data layers. To do this, I need
to randomly select a fixed percentage of polygons within each map
layer, and reclassify them to random categories within the set of
possible categories for that map.

For example, suppose a layer called forest.soils has 50 categories in
400 polygons, and I want to reclassify 5 percent of them. I would then
need to randomly select 20 polygons from the set of 400, and reclassify
each to a random value between 1 and 50 (I'm excluding NO_DATA from the
present analysis). This new, perturbed map layer would then be used as
input to the spatial model.

Any thoughts on the best way to proceed? Is there a pre-existing GRASS
facility for doing random reclassifications (I suspect not...)? How
might I string together other GRASS tools to do this?

Any help would be appreciated.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Michael Camann camann@dial.pick.uga.edu
Department of Entomology camann@phoenix.cs.uga.edu
University of Georgia (706) 542-1388
Athens, GA 30602 (706) 542-2276

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2

mike camann (camann@pick.uga.edu) writes on 25 Jun 94:

mis-classifications in the underlying data layers. To do this, I need
to randomly select a fixed percentage of polygons within each map
layer, and reclassify them to random categories within the set of
possible categories for that map.

when you say "polygons," are you talking about vector data? If so,
you might want to copy the vector map and consecutively
label each polygon (1-400). This could be done fairly quickly
with a little awk (on the dig_att file, assuming all polygons
are already labeled).

For raster data, you might look at r.clump to make contiguous
areas separate categories (1-400).

each to a random value between 1 and 50 (I'm excluding NO_DATA from the

You'll probably have to write a little 10-20 line program to do this.
What distribution should the random values have? Uniform? Normal?
There's probably enough in the source for s.rand and s.perturb to
get you started with this.

First, select 20 random values between 1-400 (uniformly, without
replacement, I suppose). Then, generate 20 random values and you're
done.

depending upon how critical this work is (e.g., are peoples' lives
involved?), you may want to be careful when selecting a RNG. The
appended code uses one distribution with the OS (which is not
known to be super-good).

--Darrell

#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#define RANDOM(lo,hi) ((float)rand()/RAND_MAX*(hi-lo)+lo)

main()
{
  int i;
  for(i=0;i<20;i++)
    printf("%d\n", RANDOM(1,400));
}