Creating unique sets randomly distributed values for many defined areas in a raster?

Question

I have a large approximately circular study area comprised of 5000+ polygons which represent forest parcels. For each of these polygons I have up to 4 different species. These species occur in an approximately random distribution (not much clumping of like species in tropical forests) E.g. one polygon may have 40% cover of Species 3 and 60% cover of species 4. For the purposes of creating an input for a separate computer model I need to create a raster that represents the unique distribution of tree canopy within each polygon. Each 10 X 10 meter cell will represent the canopy of one species. I can do this by assigning a random distribution of numbers between 0 and 1 to each polygon, and then reclassing the cells in each polygon to represent the percent cover of each species. e.g. if the cover of a given polygon is 40% species 3 and 60% species 4, then I would reclass 0-0.4 to be ‘3’ and 0.40001 to 1 to be ‘4’. My problem is that I can’t figure out how to do this on a large scale, other than manually (which would be impractical).

Is there a way to do it with ModelBuilder?

Answer 1

Let me first remark on the nature of your proposal. It's a good one, but we must understand that any realization of this random process will rarely have exactly the intended species proportion in each polygon: it will only attain those proportions on average (as statistical expectations). The actual proportions will vary from their expectations from one polygon to another.

If you want to achieve exactly the intended proportions, then you are need a random sample that is stratified by the polygons. For approaches to that (harder) problem, please see the thread at Stratified Random Sampling with large dataset in ArcGIS?.

Data preparation

A convenient way to represent the polygon-specific species distribution would be to create four grids, one for each species. Each cell gives the chance of that species being there. The grids may contain only non-negative values (and no NoData values inside any polygons) and they must sum to 1 at each polygon cell.

Convert these grids to cumulative distribution grids by taking their cumulative sum. The order doesn't really matter, so to illustrate let's suppose these species-probability grids are called "P1", "P2", "P3", and "P4". Then one way to create the cumulative sum grids is to compute

C0 = 0
C1 = "P1"
C2 = "C1" + "P2"
C3 = "C2" + "P3"

Simulation

With this preprocessing accomplished, the actual simulation is remarkably simple and efficient. As proposed in the question, create a grid of uniform random values in the interval [0,1]. Let's call this "U". The key to this solution is to use the Less Than Frequency operation. When applied to "U" and the set {"C0", "C1", "C2", "C3"}, it will produce a grid with values in the set {0, 1, 2, 3, 4}--but it should have no zero values, because every value in "U" will exceed zero (the values in "C0"). These numbers encode the randomly selected species: 1 for the species represented in "P1", 2 for the species in "P2", and so on. That is the desired reclassification.

Analysis of the algorithm

To see why this method works, consider the chance that the Less Than Frequency grid equals 2 (say). This will happen when exactly two of {"C0", ... "C3"} are less than "U". Because the values of the "C" grids are always in ascending order, this will happen precisely when "U" lies between "C1" and "C2". The chance of this occurring is "C2" - "C1" = "P2", as intended. Because the "C" grids are actually grids, and not just numbers, their values may vary from one cell to the next: that is how they accomplish this cell-specific reclassification operation.