| randomHabitat {secr} | R Documentation |
The Modified Random Cluster algorithm of Saura and Martinez-Millan (2000) is used to generate a mask object representing patches of contiguous ‘habitat’ cells (pixels) within a ‘non-habitat’ matrix (‘non-habitat’ cells are optionally dropped). Spatial autocorrelation (fragmentation) of habitat patches is controlled via the parameter ‘p’. ‘A’ is the expected proportion of ‘habitat’ cells.
randomDensity is a wrapper for randomHabitat that may be used as input to sim.popn.
randomHabitat(mask, p = 0.5, A = 0.5, directions = 4, minpatch = 1,
drop = TRUE, covname = "habitat", plt = FALSE, seed = NULL)
randomDensity(mask, parm)
mask |
secr mask object to use as template |
p |
parameter to control fragmentation |
A |
parameter for expected proportion of habitat |
directions |
integer code for adjacency (rook's move 4 or queen's move 8) |
minpatch |
integer minimum size of patch |
drop |
logical for whether to drop non-habitat cells |
covname |
character name of covariate when |
plt |
logical for whether intermediate stages should be plotted |
seed |
either NULL or an integer that will be used in a call to |
parm |
list of arguments for |
Habitat is simulated within the region defined by the cells of mask. The region may be non-rectangular.
The algorithm comprises stages A-D:
A. Randomly select proportion p of cells from the input mask
B. Cluster selected cells with any immediate neighbours as defined by
directions
C. Assign clusters to ‘non-habitat’ (probability 1–A) and ‘habitat’ (probability A)
D. Cells not in any cluster from (B) receive the habitat class of the majority of the <=8 adjacent cells assigned in (C), if there are any; otherwise they are assigned at random (with probabilities 1–A, A).
Fragmentation declines, and cluster size increases, as p increases up to the ‘percolation threshold’ which is about 0.59 in the default case (Saura and Martinez-Millan 2000 p.664).
If minpatch > 1 then habitat patches of less than minpatch
cells are converted to non-habitat, and vice versa. This is likely to
cause the proportion of habitat to deviate from A.
If drop = FALSE a binary-valued (0/1) covariate with the
requested name is included in the output mask, which has the same extent
as the input. Otherwise, non-habitat cells are dropped and no covariate
is added.
The argument ‘parm’ for randomDensity is a list with average density D and an optional subset of named values to override the defaults (p = 0.5, A = 0.5, directions = 4, minpatch = 1, plt = FALSE, seed = NULL). ‘rescale’ is a further optional component of ‘parm’; if ‘rescale = TRUE’ then the pixel-specific densities are adjusted upwards by the factor 1/A to maintain the same expected number of activity centres as if the nominal density applied throughout. Arguments ‘mask’ and ‘drop’ of randomHabitat are substituted automatically.
For randomHabitat –
An object of class ‘mask’. By default (drop = TRUE) this
has fewer rows (points) than the input mask.
The attribute “type” is a character string formed from paste('MRC p=',p, ' A=',A, sep='').
The RNG seed is stored as attribute ‘seed’ (see secrRNG).
For randomDensity –
A vector of cell-specific densities.
Single-linkage clustering and adjacency operations use functions
‘clump’ and ‘adjacency’ of the package raster; ‘clump’ also
requires package igraph0 (raster still uses this
deprecated version). Optional plotting
of intermediate stages (plt = TRUE) uses the plot method for
rasterLayers in raster.
A non-rectangular input mask is padded out to a rectangular rasterLayer for operations in raster; cells added as padding are ultimately dropped.
The procedure of Saura and Martinez-Millan (2000) has been followed as far as possible, but this implementation may not match theirs in every detail.
This implementation allows only two habitat classes. The parameter A is the expected value of the habitat proportion; the realised habitat proportion may differ quite strongly from A, especially for large p (e.g., p > 0.5).
Anisotropy is not implemented; it would require skewed adjacency filters (i.e. other than rook- or queen-move filters) that are not available in raster.
Gaussian random fields provide an alternative method for simulating
random habitats (e.g., rLGCP option in sim.popn).
Hijmans, R. J. and van Etten, J. (2011) raster: Geographic analysis and modeling with raster data. R package version 1.9-33. https://CRAN.R-project.org/package=raster.
Saura, S. and Martinez-Millan, J. (2000) Landscape patterns simulation with a modified random clusters method. Landscape Ecology, 15, 661–678.
## Not run:
tempmask <- make.mask(nx = 100, ny = 100, spacing = 20)
mrcmask <- randomHabitat(tempmask, p = 0.4, A = 0.4)
plot(mrcmask, dots = FALSE, col = "green")
pop <- sim.popn(10, mrcmask, model2D = "IHP")
plot(pop, add = TRUE)
# OR
plot(sim.popn(D = randomDensity, core = tempmask, model2D = "IHP",
details = list(D = 10, p = 0.4, A = 0.4, plt = TRUE)),
add = TRUE, frame = FALSE)
## plot intermediate steps A, C, D
opar <- par(mfrow = c(1,3))
mrcmask <- randomHabitat(tempmask, p = 0.4, A = 0.4, plt = TRUE)
par(opar)
## keep non-habitat cells
mrcmask <- randomHabitat(tempmask, p = 0.4, A = 0.4, drop = FALSE)
plot(mrcmask, covariate = "habitat", dots = FALSE,
col = c("grey","green"), breaks = 2)
## effect of purging small patches
opar <- par(mfrow=c(1,2))
mrcmask <- randomHabitat(tempmask, p = 0.4, A = 0.4, minpatch = 1)
plot(mrcmask, dots = FALSE, col ="green")
mrcmask <- randomHabitat(tempmask, p = 0.4, A = 0.4, minpatch = 5)
plot(mrcmask, dots = FALSE, col ="green")
par(opar)
## End(Not run)