kernelUD: Estimation of Kernel Home-Range
In adehabitat: Analysis of Habitat Selection by Animals

Description Usage Arguments Details Value Author(s) References See Also Examples

kernelUD is used to estimate the utilization distribution (UD) of animals monitored by radio-tracking, with the classical kernel method.
getvolumeUD and kernel.area provide utilities for home-range size estimation.
getverticeshr stores the home range contour as objects of class area in a list of class kver, with one component per animal.

kernelUD(xy, id = NULL, h = "href", grid = 40, same4all = FALSE,
         hlim = c(0.1, 1.5), kern = c("bivnorm", "epa"), extent = 0.5)
## S3 method for class 'khr'
print(x, ...)
## S3 method for class 'khr'
image(x, axes = FALSE, mar = c(0,0,2,0),
          addcontour = TRUE, addpoints = TRUE, ...)
plotLSCV(x)
getvolumeUD(x)
kernel.area(xy, id, h = "href", grid = 40,
            same4all = FALSE, hlim = c(0.1,1.5), kern = "bivnorm",
            levels = seq(20,95, by = 5),
            unin = c("m", "km"),
            unout = c("ha", "km2", "m2"), extent = 0.5)
getverticeshr(x, lev = 95)

`xy`	a data frame with two columns (x and y coordinates of the animal relocations)
`id`	an optional factor giving the animals identity associated to `xy`
`h`	a character string or a number. If `h` is set to `"href"`, the ad hoc method is used for the smoothing parameter (see details). If `h` is set to `"LSCV"`, the least-square cross validation method is used. Note that `"LSCV"` is not available if `kern = "epa"`. Alternatively, `h` may be set to any given numeric value
`grid`	a number giving the size of the grid on which the UD should be estimated. Alternatively, this parameter may be an object of class `asc`, or a list of objects of class `asc`, with named elements corresponding to each level of the factor id (see examples)
`same4all`	logical. If `TRUE`, the same grid is used for all animals. If `FALSE`, one grid per animal is used
`hlim`	a numeric vector of length two. If `h = "LSCV"`, the function minimizes the cross-validation criterion for values of h ranging from `hlim[1]href` to `hlim[2]href`, where `href` is the smoothing parameter computed with the ad hoc method (see below)
`kern`	a character string. If `"bivnorm"`, a bivariate normal kernel is used. If `"epa"`, an Epanechnikov kernel is used.
`extent`	a value indicating the extent of the grid used for the estimation (the extent of the grid on the abscissa is equal to `(min(xy[,1]) + extent * diff(range(xy[,1])))`).
`x`	an object of class `khr` returned by `kernelUD`.
`axes`	logical. Whether the axes are to be plotted
`mar`	the margin parameter (see `help(par)`)
`addcontour`	logical. If `TRUE`, contours are drawn on the graphics
`addpoints`	logical. If `TRUE`, the animal relocations are drawn on the graphics
`levels`	a vector of percentage levels for home-range size estimation
`unin`	the units of the relocations coordinates. Either `"m"` for meters (default) or `"km"` for kilometers
`unout`	the units of the output areas. Either `"m2"` for square meters, `"km2"` for square kilometers or `"ha"` for hectares (default)
`lev`	the percentage level for home range contour estimation.
`...`	additionnal parameters to be passed to the generic functions `print` and `image`

The Utilization Distribution (UD) is the bivariate function giving the probability density that an animal is found at a point according to its geographical coordinates. Using this model, one can define the home range as the minimum area in which an animal has some specified probability of being located. The functions used here correspond to the approach described in Worton (1995).

The kernel method has been recommended by many authors for the estimation of the utilization distribution (e.g. Worton, 1989, 1995). The default method for the estimation of the smoothing parameter is the ad hoc method, i.e. for a bivariate normal kernel

h = Sigma*n^(-1/6)

where

Sigma = 0.5*(sd(x)+sd(y))

which supposes that the UD is bivariate normal. If an Epanechnikov kernel is used, this value is multiplied by 1.77 (Silverman, 1986, p. 86). Alternatively, the smoothing parameter h may be computed by Least Square Cross Validation (LSCV). The estimated value then minimizes the Mean Integrated Square Error (MISE), i.e. the difference in volume between the true UD and the estimated UD. Note that the cross-validation criterion cannot be minimized in some cases. According to Seaman and Powell (1998) "This is a difficult problem that has not been worked out by statistical theoreticians, so no definitive response is available at this time" (see Seaman and Powell, 1998 for further details and tricky solutions). plotLSCV allows to have a diagnostic of the success of minimization of the cross validation criterion (i.e. to know whether the minimum of the CV criterion occurs within the scanned range). Finally, the UD is then estimated over a grid.

The default kernel is the bivariate normal kernel, but the Epanechnikov kernel, which requires less computer time is also available for the estimation of the UD.

The function getvolumeUD modifies the UD component of the object passed as argument, so that the contour of the UD displayed by the functions contour and image.khr corresponds to the different percentage levels of home-range estimation (see examples). In addition, this function is used in the function kernel.area, to compute the home-range size. Note, that the function plot.hrsize (see the help page of this function) can be used to display the home-range size estimated at various levels.

The class khr is a class grouping three sub-classes, khrud, kbbhrud and khrudvol: kernelUD returns a list of the class khrud. This list has one component per animal (named as the levels of argument id). Each component is itself a list, with the following sub-components:

`UD`	an object of class `asc`, with the values of density probability in each cell of the grid
`h`	if `LSCV` is not used, the value of the smoothing parameter. if `LSCV` is used, a list with three components: `CV` the results of the cross-validation procedure. The first column contains the sequence of values tested for the smoothing parameter, and the second column contains the value of the cross-validation criterion. `convergence` TRUE if the LSCV succeeds (i.e. if the optimum smoothing parameter have been found by the procedure), FALSE otherwise. `h` the value of the smoothing parameter used in UD estimation.
`locs`	The relocations used in the estimation procedure.
`hmeth`	The argument `h` of the function kernelUD

getvolumeUD returns a list of class khrvol, with the same components as lists of class khrud.

kernel.area returns a data frame of subclass hrsize, with one column per animal and one row per level of estimation of the home range.

getverticeshr returns an object of class kver.

Clement Calenge clement.calenge@oncfs.gouv.fr

Silverman, B.W. (1986) Density estimation for statistics and data analysis. London: Chapman \& Hall.

Worton, B.J. (1989) Kernel methods for estimating the utilization dirstibution in home-range studies. Ecology, 70, 164–168.

Worton, B.J. (1995) Using Monte Carlo simulation to evaluate kernel-based home range estimators. Journal of Wildlife Management, 59,794–800.

Seaman, D.E. and Powell, R.A. (1998) Kernel home range estimation program (kernelhr). Documentation of the program.

asc for additionnal informations on objects of class asc, mcp for estimation of home ranges using the minimum convex polygon, and for help on the function plot.hrsize. kver for information on objects of class kver, kernelbb for an alternative approach of the kernel estimation for trajectory data.

data(puechabon)
loc <- puechabon$locs[, c("X", "Y")]
id <- puechabon$locs[, "Name"]

## Estimation of UD for the four animals
(ud <- kernelUD(loc, id))

image(ud) ## Note that the contours
          ## corresponds to values of probability density
udvol <- getvolumeUD(ud)
image(udvol)
## Here, the contour corresponds to the
## home ranges estimated at different probability
## levels (i.e. the contour 90 corresponds to the 90 percent
## kernel home-range)
## udvol describes, for each cell of the grid,
## the smaller home-range to which it belongs 

## Calculation of the 95 percent home range
ver <- getverticeshr(ud, 95)
elev <- getkasc(puechabon$kasc, "Elevation") # Map of the area
image(elev)
plot(ver, add=TRUE)
legend(696500, 3166000, legend = names(ver), fill = rainbow(4))


## Example of estimation using LSCV
udbis <- kernelUD(loc, id, h = "LSCV")
image(udbis)

## Compare the estimation with ad hoc and LSCV method
## for the smoothing parameter
(cuicui1 <- kernel.area(loc, id)) ## ad hoc
plot(cuicui1)
(cuicui2 <- kernel.area(loc, id, h = "LSCV")) ## LSCV
plot(cuicui2)

## Diagnostic of the cross-validation
plotLSCV(udbis)


## Use of the same4all argument: the same grid
## is used for all animals
udbis <- kernelUD(loc, id, same4all = TRUE)
image(udbis)

## Estimation of the UD on a map
## (e.g. for subsequent analyses on habitat selection)
elev <- getkasc(puechabon$kasc, "Elevation")
opar <- par(mfrow = c(2, 2), mar = c(0, 0, 2, 0))
cont <- getcontour(elev)

for (i in 1:length(udbis)) {
   image(elev, main = names(udbis)[i], axes = FALSE)
   points(udbis[[i]]$locs, pch = 21, bg = "white", col = "black")
}


## Measures the UD in each pixel of the map
udbis <- kernelUD(loc, id, grid = elev)
opar <- par(mfrow = c(2, 2), mar = c(0, 0, 2, 0))
for (i in 1:length(udbis)) {
  image(udbis[[i]]$UD, main = names(udbis)[i], axes = FALSE)
  box()
  polygon(cont[, 2:3])
}
par(opar)




## Estimation of the UD with a list of objects of class "asc" passed as
## argument grid (useful for large datasets)

## For example, consider the following limits:
lim <- rbind(c(697901,701061,3160198,3162604),
             c(698936,701089,3159969,3162518),
             c(698461,701928,3157362,3160427),
             c(698265,701369,3157219,3162661))

gro <- lapply(1:4, function(i) {
              subsetmap(elev, xlim = lim[i,1:2], ylim=lim[i,3:4])
})
names(gro) <- levels(id)

## show the data:
opar <- par(mfrow=c(2,2), mar=c(0.1,0.1,2,0.1))
lapply(1:4, function(i) {
  image(gro[[i]], main=names(gro)[i], axes=FALSE)
  points(loc[id==names(gro)[i],])
  box()
})
gro

## The map has been subset to fit the relocations.
## Now, estimate the UD:
ud.one.per.grid <- kernelUD(loc, id, grid = gro)
image(ud.one.per.grid)


## The UD can then be matched to habitat maps