data.table internalsR/spread2.R
spread2.RdThis can be used to simulate fires, seed dispersal, calculation of iterative,
concentric, symmetric (currently) landscape values and many other things.
Essentially, it starts from a collection of cells (start, called "events")
and spreads to neighbours, according to the directions
and spreadProb with modifications due to other arguments. NOTE:
the spread function is similar, but sometimes slightly faster, but less
robust, and more difficult to use iteratively.
spread2(
landscape,
start = ncell(landscape)/2 - ncol(landscape)/2,
spreadProb = 0.23,
persistProb = NA_real_,
asRaster = TRUE,
maxSize,
exactSize,
directions = 8L,
iterations = 1000000L,
returnDistances = FALSE,
returnDirections = FALSE,
returnFrom = FALSE,
maxRetriesPerID = 10,
spreadProbRel = NA_real_,
plot.it = FALSE,
circle = FALSE,
asymmetry = NA_real_,
asymmetryAngle = NA_real_,
allowOverlap = 0,
neighProbs = NA_real_,
oneNeighbourOnly = FALSE,
skipChecks = FALSE
)Required. A RasterLayer object. This defines the possible
locations for spreading events to start and spread2 into. Required.
Required. Either a vector of pixel numbers to initiate spreading, or a
data.table that is the output of a previous spread2.
If a vector, they should be cell indices (pixels) on the landscape.
If user has x and y coordinates, these can be converted with
cellFromXY().
Numeric of length 1 or length ncell(landscape) or
a RasterLayer that is the identical dimensions as
landscape.
If numeric of length 1, then this is the global (absolute)
probability of spreading into each cell from a neighbour.
If a numeric of length ncell(landscape) or a raster,
then this must be the cell-specific (absolute)
probability of a "receiving" potential cell. Default is 0.23.
If relative probabilities are required, use spreadProbRel.
If used together, then the relative probabilities will be
re-scaled so that the mean relative probability of potential
neighbours is equal to the mean of spreadProb of
the potential neighbours.
Numeric of length 1 or RasterLayer.
If numeric of length 1, then this is the global (absolute)
probability of cell continuing to burn per time step.
If a raster, then this must be the cell-specific (absolute)
probability of a fire persisting.
Default is NA, which is the same as 0, i.e. a cell only burns
for one time step.
Logical, length 1. If TRUE, the function will return a Raster
where raster non NA values indicate the cells that were "active", and the
value is the initial starting pixel.
Numeric. Maximum number of cells for a single or all events to be spread2.
Recycled to match start length, if it is not as long as start.
This will be overridden if exactSize also provided.
See section on 'Breaking out of spread2 events'.
Numeric vector, length 1 or length(start).
Similar to maxSize, but these will be the exact
final sizes of the events. i.e., the spread2 events
will continue until they are floor(exactSize).
This will override maxSize if both provided.
See Details.
The number adjacent cells in which to look; default is 8 (Queen case). Can only be 4 or 8.
Number of iterations to spread2.
Leaving this NULL allows the spread2 to continue
until stops spreading itself (i.e., exhausts itself).
Logical. Should the function include a column with the individual cell distances from the locus where that event started. Default is FALSE. See Details.
Logical. Should the function include a column with the individual directions (in radians) from the locus where that event started. Default is FALSE.
Logical. Should the function return a column with the source, i.e, the lag 1 "from" pixel, for each iteration.
Only active if exactSize is used. This is the number of attempts
that will be made per event ID, before abandoning, therefore completing
the spread2 for that event with a size that is smaller than
exactSize. Default 10 times.
Optional RasterLayer indicating a surface of relative
probabilities useful when using neighProbs (which
provides a mechanism for selecting a specific number of
cells at each iteration).
This indicates the relative probabilities for the selection
of successful neighbours.
spreadProb will still be evaluated after
the relative probabilities and neighProbs has been
evaluated, i.e., potential cells will be identified, then
some could be rejected via spreadProb.
If absolute spreadProb is not desired,
be sure to set spreadProb = 1.
Ignored if neighProbs is not provided.
If TRUE, then plot the raster at every iteration,
so one can watch the spread2 event grow.
Logical. If TRUE, then outward spread2 will be by equidistant rings,
rather than solely by adjacent cells (via directions arg.).
Default is FALSE.
Using circle = TRUE can be dramatically slower for large problems.
Note, this will likely create unexpected results if spreadProb < 1.
A numeric or RasterLayer indicating the ratio of the
asymmetry to be used. i.e., 1 is no asymmetry; 2 means that the
angles in the direction of the asymmetryAngle are 2x the
spreadProb
of the angles opposite tot he asymmetryAngle Default is
NA, indicating no asymmetry. See details. This is still experimental.
Use with caution.
A numeric or RasterLayer indicating the angle in degrees
(0 is "up", as in North on a map),
that describes which way the asymmetry is.
numeric (logical will work for backwards compatibility).
See details. Default is 0, i.e., no overlapping.
An optional numeric vector, whose sum is 1.
It indicates the probabilities that an individual spread iteration
will spread to 1, 2, ..., length(neighProbs) neighbours, respectively.
If this is used (i.e., something other than NA), circle and
returnDistances will not work currently.
Logical. Default is FALSE. If TRUE, then this
spread algorithm will allow exactly one neighbour to be
spread to (not fewer or more). This could be used, e.g.,
for an animal moving. If this is TRUE, then allowOverlap
will be set to 2 if it is 0 or 1.
Logical. If TRUE, the argument checking (i.e., assertions) will be skipped. This should likely only be used once it is clear that the function arguments are well understood and function speed is of the primary importance. This is likely most useful in repeated iteration cases i.e., if this call is using the previous output from this same function.
Either a data.table (asRaster=FALSE) or a RasterLayer
(asRaster=TRUE, the default).
The data.table will have one attribute named spreadState, which
is a list containing a data.table of current cluster-level information
about the spread events.
If asRaster=TRUE, then the data.table (with its spreadState
attribute) will be attached to the Raster as an attribute named pixel as it
provides pixel-level information about the spread events.
The RasterLayer represents every cell in which a successful spread2 event occurred.
For the case of, say, a fire this would represent every cell that burned.
If allowOverlap is TRUE, the return will always be a data.table.
If asRaster is FALSE, then this function returns a
data.table with 3 (or 4 if returnFrom is TRUE) columns:
initialPixels | the initial cell number of that particular
spread2 event. |
pixels | The cell indices of cells that have
been touched by the spread2 algorithm. |
state | a logical indicating whether the cell is active (i.e., could still be a source for spreading) or not (no spreading will occur from these cells). |
from | The pixel indices that were the immediately preceding
"source" for each pixels, i.e., the lag 1 pixels.
Only returned if returnFrom is TRUE |
The attribute saved with the name "spreadState" (e.g., attr(output, "spreadState"))
includes a data.table with columns:
id | An arbitrary code, from 1 to length(start) for each "event". |
initialPixels | the initial cell number of that particular
spread2 event. |
numRetries | The number of re-starts the event did because it got
stuck (normally only because exactSize was used
and was not achieved. |
maxSize | The number of pixels that were provided as inputs via
maxSize or exactSize. |
size | The current size, in pixels, of each event. |
and several other objects that provide significant speed ups in iterative calls to
spread2. If the user runs spread2 iteratively, there will likely be significant
speed gains if the data.table passed in to start should have the attribute
attached, or re-attached if it was lost, e.g., via
setattr(outInput, "spreadState", attr(out, "spreadState")), where out is the
returned data.table from the previous call to spread2, and outInput is
the modified data.table. Currently, the modified data.table
must have the
same order as out.
There are 2 main underlying algorithms for active cells to "spread" to
nearby cells (adjacent cells): spreadProb and neighProb.
Using spreadProb, every "active" pixel will assess all
neighbours (either 4 or 8, depending on directions), and will "activate"
whichever neighbours successfully pass independent calls to
runif(1,0,1) < spreadProb.
The algorithm will iterate again and again, each time starting from the newly
"activated" cells. Several built-in decisions are as follows.
no active cell can activate a cell that was already activated by
the same event (i.e., "it won't go backwards"). 2. If allowOverlap is
FALSE, then the previous rule will also apply, regardless of which
"event" caused the pixels to be previously active.
This function can be interrupted before all active cells are exhausted if
the iterations value is reached before there are no more active
cells to spread2 into. The interrupted output (a data.table) can be passed
subsequently as an input to this same function (as start).
This is intended to be used for situations where external events happen during
a spread2 event, or where one or more arguments to the spread2 function
change before a spread2 event is completed.
For example, if it is desired that the spreadProb change before a
spread2 event is completed because, for example, a fire is spreading, and a
new set of conditions arise due to a change in weather.
asymmetry here is slightly different than in the spread function,
so that it can deal with a RasterLayer of asymmetryAngle.
Here, the spreadProb values of a given set of neighbours around each active pixel
are adjusted to create adjustedSpreadProb which is calculated maintain the
following two qualities: $$mean(spreadProb) = mean(ajustedSpreadProb)$$ and
$$max(spreadProb)/min(spreadProb) = asymmetry$$ along the axis of
asymmetryAngle. NOTE: this means that the 8 neighbours around an active
cell may not fulfill the preceeding equality if asymmetryAngle is not
exactly one of the 8 angles of the 8 neighbours. This means that
$$max(spreadProb)/min(spreadProb)$$ will generally be less than
asymmetry, for the 8 neighbours. The exact adjustment to the spreadProb
is calculated with:
$$angleQuality <- (cos(angles - rad2(asymmetryAngle))+1)/2$$
which is multiplied to get an angle-adjusted spreadProb:
$$spreadProbAdj <- actualSpreadProb * angleQuality$$
which is then rescaled:
$$adjustedSpreadProb = (spreadProbAdj - min(spreadProbAdj)) * par2 + par1$$,
where par1 and par2 are parameters calculated internally to make the 2 conditions above true.
If maxSize or exactSize are used, then spreading will continue and stop
before or at maxSize or at exactSize, respectively.
If iterations is specified, then the function will end, and the returned data.table
may (if maxSize) or will (if exactSize) have at least one active
cell per event that did not already achieve maxSize or exactSize.
This will be very useful to build new, customized higher-level wrapper functions that
iteratively call spread2.
exactSize may not be achieved if there aren't enough cells in the map.
Also, exactSize may not be achieved because the active cells are "stuck",
i.e., they have no inactivated cells to move to; or the spreadProb is low.
In the latter two cases, the algorithm will retry again, but it will only
retry from the last iteration's active cells.
The algorithm will only retry 10 times before quitting.
Currently, there will also be an attempt to "jump" up to four cells away from
the active cells to try to continue spreading.
A common way to use this function is to build wrappers around this, followed
by iterative calls in a while loop. See example.
spread2 eventsThere are 3 ways for the spread2 to "stop" spreading.
Here, each "event" is defined as all cells that are spawned from each unique
start location.
The ways outlined below are all acting at all times, i.e., they are not
mutually exclusive.
Therefore, it is the user's responsibility to make sure the different rules
are interacting with each other correctly.
spreadProb | Probabilistically, if spreadProb is low enough,
active spreading events will stop.
In practice, this number generally should be below 0.3
to actually see an event stop. |
maxSize | This is the number of cells that are "successfully" turned
on during a spreading event. spreadProb will still
be active, so, it is possible that the end size of each event
is smaller than maxSize, but they will not be greater
than maxSize |
exactSize | This is the number of cells that are "successfully" turned
on during a spreading event. This will override an event that
stops probabilistically via spreadProb, but forcing
its last set of active cells to try again to find neighbours.
It will try maxRetriesPerID times per event, before giving up.
During those maxRetriesPerID times, it will try to "jump" up to
4 cells outwards from each of the active cells, every 5 retries. |
iterations | This is a hard cap on the number of internal iterations to
complete before returning the current state of the system
as a data.table. |
This function can be used iteratively, with relatively little overhead compared to using
it non-iteratively. In general, this function can be called with arguments set as user
needs, and with specifying e.g., iterations = 1. This means that the function will spread
outwards 1 iteration, then stop. The returned object will be a data.table or
RasterLayer that can be passed immediately back as the start argument into a subsequent
call to spread2. This means that every argument can be updated at each iteration.
When using this function iteratively, there are several things to keep in mind.
The output will likely be sorted differently than the input (i.e., the
order of start, if a vector, may not be the same order as that returned).
This means that when passing the same object back into the next iteration of the
function call, maxSize or exactSize may not be in the same order.
To get the same order, the easiest thing to do is sort the initial start
objects by their pixel location, increasing.
Then, of course, sorting any vectorized arguments (e.g., maxSize) accordingly.
NOTE: the data.table or RasterLayer should not be altered
when passed back into spread2.
allowOverlapIf 1 (or TRUE),
then individual events can overlap with one another, i.e., allow
overlap between events. If 2 (or NA), then each pixel
is essentially independent, allowing overlap between and within
events. This likely requires a user to intervene as it is possible
to spread back onto itself. If 3 (did not exist previously),
individual events can overlap, and there can be overlap within an
event, but only within an iteration, i.e., once an iteration is
finished, and a pixel was activated, then the spreading will not
return onto these pixels. If 0 (or FALSE), then once a
pixel is activated, it cannot be re-activated, within or between event.
This allows events to not interfere with one another i.e.,
they do not interact (this is slower than if
allowOverlap = FALSE). Default is 0. In the case of 2 or 3,
this would be, perhaps, useful for dispersal of,
say, insect swarms.
spread() for a different implementation of the same algorithm.
spread is less robust but it is often slightly faster.
library(terra)
origDTThreads <- data.table::setDTthreads(2L)
origNcpus <- options(Ncpus = 2L)
a <- rast(ext(0, 10, 0, 10), res = 1)
sams <- sort(sample(ncell(a), 3))
# Simple use -- similar to spread(...)
out <- spread2(a, start = sams, 0.225)
if (interactive()) {
terra::plot(out)
}
# Use maxSize -- this gives an upper limit
maxSizes <- sort(sample(1:10, size = length(sams)))
out <- spread2(a, start = sams, 0.225, maxSize = maxSizes, asRaster = FALSE)
# check TRUE using data.table .N
out[, .N, by = "initialPixels"]$N <= maxSizes
#> [1] TRUE TRUE TRUE
# Use exactSize -- gives an exact size, if there is enough space on the Raster
exactSizes <- maxSizes
out <- spread2(a, start = sams, spreadProb = 0.225,
exactSize = exactSizes, asRaster = FALSE)
out[, .N, by = "initialPixels"]$N == maxSizes # should be TRUE TRUE TRUE
#> [1] TRUE TRUE TRUE
# Use exactSize -- but where it can't be achieved
exactSizes <- sort(sample(100:110, size = length(sams)))
out <- spread2(a, start = sams, 1, exactSize = exactSizes)
# Iterative calling -- create a function with a high escape probability
spreadWithEscape <- function(ras, start, escapeProb, spreadProb) {
out <- spread2(ras, start = sams, spreadProb = escapeProb, asRaster = FALSE)
while (any(out$state == "sourceActive")) {
# pass in previous output as start
out <- spread2(ras, start = out, spreadProb = spreadProb,
asRaster = FALSE, skipChecks = TRUE) # skipChecks for speed
}
out
}
set.seed(421)
out1 <- spreadWithEscape(a, sams, escapeProb = 0.25, spreadProb = 0.225)
set.seed(421)
out2 <- spread2(a, sams, 0.225, asRaster = FALSE)
# The one with high escape probability is larger (most of the time)
NROW(out1) > NROW(out2) ## TODO: not true
#> [1] FALSE
## Use neighProbs, with a spreadProb that is a RasterLayer
# Create a raster of different values, which will be the relative probabilities
# i.e., they are rescaled to relative probabilities within the 8 neighbour choices.
# The neighProbs below means 70% of the time, 1 neighbour will be chosen,
# 30% of the time 2 neighbours.
# The cells with spreadProb of 5 are 5 times more likely than cells with 1 to be chosen,
# when they are both within the 8 neighbours
sp <- rast(ext(0, 3, 0, 3), res = 1, vals = 1:9) #small raster, simple values
# Check neighProbs worked
out <- list()
# enough replicates to see stabilized probabilities
for (i in 1:100) {
out[[i]] <- spread2(sp, spreadProbRel = sp, spreadProb = 1,
start = 5, iterations = 1,
neighProbs = c(1), asRaster = FALSE)
}
out <- data.table::rbindlist(out)[pixels != 5] # remove starting cell
table(sp[out$pixels])
#> lyr.1
#> 1 2 3 4 6 7 8 9
#> 3 6 8 7 16 15 22 23
# should be non-significant -- note no 5 because that was the starting cell
# This tests whether the null model is true ... there should be proportions
# equivalent to 1:2:3:4:6:7:8:9 ... i.e,. cell 9 should have 9x as many events
# spread to it as cell 1. This comes from sp object above which is providing
# the relative spread probabilities
keep <- c(1:4, 6:9)
chisq.test(keep, unname(tabulate(sp[out$pixels]$lyr.1, 9)[keep]),
simulate.p.value = TRUE)
#>
#> Pearson's Chi-squared test with simulated p-value (based on 2000
#> replicates)
#>
#> data: keep and unname(tabulate(sp[out$pixels]$lyr.1, 9)[keep])
#> X-squared = 56, df = NA, p-value = 1
#>
## Example showing asymmetry
sams <- ncell(a) / 4 - ncol(a) / 4 * 3
circs <- spread2(a, spreadProb = 0.213, start = sams,
asymmetry = 2, asymmetryAngle = 135,
asRaster = TRUE)
## ADVANCED: Estimate spreadProb when using asymmetry, such that the expected
## event size is the same as without using asymmetry
# \donttest{
if (interactive()) {
# Still using `raster` as it works more easily with parallelism due to not using pointers
# This will updated at a later release
if (requireNamespace("raster", quietly = TRUE)) {
ras <- raster::raster(a)
ras[] <- 1
n <- 100
sizes <- integer(n)
for (i in 1:n) {
circs <- spread2(ras, spreadProb = 0.225,
start = round(ncell(ras) / 4 - ncol(ras) / 4 * 3),
asRaster = FALSE)
sizes[i] <- circs[, .N]
}
goalSize <- mean(sizes)
if (requireNamespace("DEoptim", quietly = TRUE)) {
library(parallel)
library(DEoptim)
# need 10 cores for 10 populations in DEoptim
cl <- makeCluster(pmin(10, detectCores() - 2))
parallel::clusterEvalQ(cl, {
library(SpaDES.tools)
library(terra)
library(raster)
library(fpCompare)
})
objFn <- function(sp, n = 20, ras, goalSize) {
sizes <- integer(n)
for (i in 1:n) {
circs <- spread2(ras, spreadProb = sp, start = ncell(ras) / 4 - ncol(ras) / 4 * 3,
asymmetry = 2, asymmetryAngle = 135,
asRaster = FALSE)
sizes[i] <- circs[, .N]
}
abs(mean(sizes) - goalSize)
}
aa <- DEoptim(objFn, lower = 0.2, upper = 0.23,
control =
DEoptim.control(
cluster = cl, NP = 10, VTR = 0.02,
# imposing itermax simply for example; should let go to completion
itermax = 5,
initialpop = as.matrix(rnorm(10, 0.213, 0.001))),
ras = ras, goalSize = goalSize)
# The value of spreadProb that will give the
# same expected event sizes to spreadProb = 0.225 is:
sp <- aa$optim$bestmem
circs <- spread2(ras, spreadProb = sp, start = ncell(ras) / 4 - ncol(ras) / 4 * 3,
asymmetry = 2, asymmetryAngle = 135, asRaster = FALSE)
stopCluster(cl)
}
}
}
# clean up
data.table::setDTthreads(origDTThreads)
options(Ncpus = origNcpus)
# }