[R-sig-Geo] Odd behavior of dismo's extract function

[Dan Warren]

This is not an error per se so much as just something very weird that I
have noticed with a project I've been working on recently.  I'm wondering
if anyone here has any insight as to what may be causing this behavior.  I
haven't yet been able to duplicate it with simulated rasters (more info on
that below), but it appears very reliably with real environmental data
including the PC rasters for Cuba I have hosted here:

https://github.com/danlwarren/ENMTools/tree/master/test/testdata

What's happening is this: if I go to extract data from those rasters using
occurrence points, the amount of time it takes increases very rapidly up to
exactly 250 points, and falls dramatically after that.  So dramatically
that it takes over two minutes to extract data for 250 points but just
under three seconds for 251.  I've established that it's not a question of
the points themselves being wonky, because it happens with random points as
well.


extract.test <- function(env, N){
      extract(env, randomPoints(env, N))
}

env.files <- list.files(path = "testdata/", pattern = "pc", full.names =
TRUE)
env <- stack(env.files)

system.time(extract.test(env, 250))

   user  system elapsed
  2.807   0.084   2.891

system.time(extract.test(env, 251))

   user  system elapsed
124.562   0.516 125.061

numpoints,time
1,1.54
5,3.93
10,6.764
50,29.939
100,61.431
150,79.295
200,110.283
250,120.118
251,2.748
252,2.756
254,2.767
500,2.876
1000,3.153

The data being extracted looks perfectly reasonable in all cases.  It's not
just these layers, either.  Although (as I mentioned above) I have yet to
come up with simulated rasters that show this behavior, I see this behavior
for both of the sets of rasters for real environmental data that I've
tried.  The results above are from a PCA on Worldclim data for Cuba, but I
just tried them on some Climond data I've got for Australia and I get the
same behavior.  Those rasters are much larger, though, and a result the
times are longer; 251 points took about 43 seconds, whereas I just had to
give up and stop the 250 point extraction after about 30 minutes.

As for those simulated rasters, I've tried the following:

Plain grids of sequential numbers
As above, but with a bunch of NAs added
Filling the Cuban rasters with sequential numbers
Filling the Cuban rasters with random numbers from a uniform (0,1)
distribution

None of those show this issue.  Anyone have any thoughts about what might
be going on here?

	[[alternative HTML version deleted]]