MPG(...)$lcpPerimWeight contains mystery links when input patch raster has NA regions

[ecologics]

The following clip from @cxfauvelle input raster can be used to demonstrate this problem.
naErrorExample.zip

res <- raster("naErrorExample.asc")
resMPG <- MPG(res, patch=res==1)
## Red areas are NA
plot(is.na(res), col=c("white", "red"))
## Superimpose the mpgPlot and see the orphaned links
plot(res$mpgPlot, add=TRUE)

As below where red areas are NA and green areas are real patches (==1).

[achubaty]

This is a byproduct of how the NA values are handled. @samdoctolero can you take a look at this?

[achubaty]

@ecologics I spoke to Sam about this and he should have it fixed by next week.

[ecologics]

Awesome!  I am excited about this work.  Let's try to get a manuscript out early in the new year.

Paul

[achubaty]

It seems the problem is caused by the voronoi being able to spread through NA regions but links aren't permitted, so they stop short at NA boundaries.

To me, the easiest fix (and a good default i.e., conservative behaviour) is to disallow the voronoi spreading and linking through NAs. @ecologics What is the rationale for allowing spread through NA regions?

If the user wants to do something more sophisticated with the NAs they can pre-process the raster to reclassify NAs as another value (perhaps a mean or median of the neighbouring cells).

Thoughts?

[ecologics]

I agree on all these points.  The old version of grainscape did not handle NAs at all.  Indeed it routinely crashed.  So the user had to deal with them by assigning.  So making them, essentially, undefined areas on the map -- if this is easiest -- makes the most sense.

[achubaty]

Sam pushed a fix for this yesterday (137fe47). I'll take a look later today to do some tests.

[achubaty]

@samdoctolero I get an error running your updated code:

  library(igraph)
  library(raster)

  ## simplest case
  x <- 100
  y <- 100
  m <- matrix(2L, ncol = y, nrow = y)
  m[1:95, 46:55] <- NA_integer_
  m[16:35, 24:35] <- 3L
  m[16:35, 76:90] <- 3L

  r <- raster(m)
  c <- reclassify(r, rcl = cbind(c(1, 2, 3), c(10, 5, 1)))
  p <- (c == 1)

  mpg <- MPG(cost = c, patch = p)

The above fails with:

Error in data.frame(name = uniquePatches, patchId = uniquePatches, patchArea = patchArea,  : 
  arguments imply differing number of rows: 3, 2
Called from: data.frame(name = uniquePatches, patchId = uniquePatches, patchArea = patchArea, 
    patchEdgeArea = patchEdgeArea, coreArea = patchArea - patchEdgeArea)

@ecologics @cxfauvelle please hold off updating until this is resolved

[samdoctolero]

When ignoring NA values was implemented, MPG.R needed to be updated as well (59241e2).

## Get additional patch information
uniquePatches <- voronoi[] %>% unique() %>% sort()
uniquePatches <- uniquePatches[uniquePatches > 0] ## added this

I understand that the unique sets of patches had to be extracted and sorted at the same time. Since we ignored the NA values, they got included as part of the unique patches. The first element of uniquePatches had to be removed in order to have the right number of patches.

@achubaty feel free to change that line. Basically, I'm just removing the first element of the vector which will always be zero.

[achubaty]

@samdoctolero with the change in 203cff5, this is working. I have tweaked the MPG() code a bit, and built a simple set of unit tests for this issue.

The phantom links are now gone from the test raster posted by @ecologics and @cxfauvelle:

f <- 'naErrorExample.asc'
r <- raster(f)
c <- r
p <- (r == 1)

mpg <- MPG(cost = c, patch = p)

plot(is.na(r), col = c("white", "red")) ## red areas are NA
plot(mpg, add = TRUE)                   ## superimpose the mpg plot

However, note that not all patches are linked. Is this another bug? See new issue #32.