kernelUD: Estimation of Kernel Home-Range

Description

The function kernelUD estimates the UD of one or several animals.

plotLSCV allows to explore the results of the least-square cross-validation algorithm used to find the best smoothing value.

image allows a graphical display of the estimates.

getvolumeUD and kernel.area provide utilities for home range and home-range size estimation.

getverticeshr stores the home range contour as an object of class SpatialPolygonsDataFrame (package sp), with one row per animal.

estUDm2spixdf can be used to convert the result into an object of class SpatialPixelsDataFrame

as.data.frame.estUD can be used to convert an object of class estUD as a data frame.

Usage

kernelUD(xy, h = "href", grid = 60,
         same4all = FALSE, hlim = c(0.1, 1.5),
         kern = c("bivnorm", "epa"), extent = 1,
         boundary = NULL)
# S3 method for estUDm
print(x, ...)
# S3 method for estUD
image(x, ...)
# S3 method for estUDm
image(x, ...)
# S3 method for estUD
as.data.frame(x, row.names, optional, ...)
plotLSCV(x)
getvolumeUD(x, standardize = FALSE)
kernel.area(x, percent = seq(20, 95, by = 5),
            unin = c("m", "km"),
            unout = c("ha", "km2", "m2"), standardize = FALSE)
estUDm2spixdf(x)

Value

The function kernelUD returns either: (i) an object belonging to the S4 class estUD (see ?estUD-class) when the object

xy passed as argument contains the relocations of only one animal (i.e., belong to the class SpatialPoints), or (ii) a list of elements of class estUD when the object

xy passed as argument contains the relocations of several animals (i.e., belong to the class SpatialPointsDataFrame).

The function getvolumeUD returns an object of the same class as the object passed as argument (estUD or estUDm).

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

SpatialPolygonsDataFrame.

estUDm2spixdf returns an object of class

SpatialPixelsDataFrame.

Arguments

xy: An object inheriting the class SpatialPoints containing the x and y relocations of the animal. If xy inherits the class SpatialPointsDataFrame, it should contain only one column (factor) corresponding to the identity of the animals for each relocation.
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 inheriting the class SpatialPixels, that will be used for all animals. For the function kernelUD, it may in addition be a list of objects of class SpatialPixels, with named elements corresponding to each level of the factor id.
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 controlling the extent of the grid used for the estimation (the extent of the grid on the abscissa is equal to (min(abscissa.relocations) + extent * diff(range(abscissa.relocations))), and similarly for the ordinate).
same4all: logical. If TRUE, the same grid is used for all animals. If FALSE, one grid per animal is used. Note that when same4all = TRUE, the grid used for the estimation is calculated by the function (so that the parameter grid cannot be a SpatialPixels object).
boundary: If, not NULL, an object inheriting the class SpatialLines defining a barrier that cannot be crossed by the animals. There are constraints on the shape of the barrier that depend on the smoothing parameter h (***see details***)
x: an object of class estUD (UD for one animal) or estUDm (UD for several animals). For the function estUDm2spixdf, an object of class estUDm only. For the function as.data.frame.estUD, an object of class estUD only.
percent: for kernel.area, a vector of percentage levels for home-range size estimation. For getverticeshr, a single value giving the percentage level for home-range estimation.
standardize: a logical value indicating whether the UD should be standardized over the area of interest, so that the volume under the UD and *over the area* is equal to 1.
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)
row.names: unused argument here
optional: unused argument here
...: additionnal parameters to be passed to the generic functions print and image

Author

Clement Calenge clement.calenge@ofb.gouv.fr

Details

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 $$h = \sigma n^{- \frac{1}{6}}$$ where $$\sigma^2 = 0.5 (var(x)+var(y))$$ 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: that the pixel values of the resulting object are equal to the percentage of the smallest home range containing this pixel. 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 parameter boundary allows to define a barrier that cannot be crossed by the animals. When this parameter is set, the method described by Benhamou and Cornelis (2010) for correcting boundary biases is used. The boundary can possibly be defined by several nonconnected lines, each one being built by several connected segments. Note that there are constraints on these segments (not all kinds of boundary can be defined): (i) each segment length should at least be equal to 3*h (the size of "internal lane" according to the terminology of Benhamou and Cornelis), (ii) the angle between two line segments should be greater that pi/2 or lower that -pi/2. The UD of all the pixels located within a band defined by the boundary and with a width equal to 6*h ("external lane") is set to zero.

References

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

Worton, B.J. (1989) Kernel methods for estimating the utilization distribution 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.

Benhamou, S. and Cornelis, D. (2010) Incorporating Movement Behavior and Barriers to Improve Biological Relevance of Kernel Home Range Space Use Estimates. Journal of Wildlife Management, 74, 1353--1360.

Examples

Run this code


## Load the data
data(puechabonsp)
loc <- puechabonsp$relocs

## have a look at the data
head(as.data.frame(loc))
## the first column of this data frame is the ID


## Estimation of UD for the four animals
(ud <- kernelUD(loc[,1]))

## The UD of the four animals
image(ud)

## Calculation of the 95 percent home range
ver <- getverticeshr(ud, 95)

## and display on an elevation map:
elev <- puechabonsp$map
image(elev, 1)
plot(ver, add=TRUE, col=rainbow(4))
legend(699000, 3165000, legend = names(ud), fill = rainbow(4))

## Example of estimation using LSCV
udbis <- kernelUD(loc[,1], h = "LSCV")
image(udbis)


## Compare the estimation with ad hoc and LSCV method
## for the smoothing parameter
(cuicui1 <- kernel.area(ud)) ## ad hoc
plot(cuicui1)
(cuicui2 <- kernel.area(udbis)) ## LSCV
plot(cuicui2)

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



## Use of the same4all argument: the same grid
## is used for all animals
## BTW, we indicate a grid with a fine resolution:
udbis <- kernelUD(loc[,1], same4all = TRUE, grid = 100)
image(udbis)


## Estimation of the UD on a map
## (e.g. for subsequent analyses on habitat selection)
## Measures the UD in each pixel of the map
udbis <- kernelUD(loc[,1], grid = elev)
image(udbis)


##########################################
##
## Estimating the UD with the presence of a barrier
## The boars are located on the plateau of Puechabon (near
## Montpellier, France), and their movements are limited by the
## Herault river.

## We first map the elevation:
image(elev)

## Then, we used the function locator() to identify the limits of the
## segments of this barrier. BEWARE! The boundary should satisfy the two
## constraints: (i) segment length > 3*h, (ii) no angle lower than pi/2
## between successive segments. We choose a smoothing parameter of 100
## m, so that no segment length should be less than 300 m.
## Because the resolution of the map is 100 m, this means that no
## segment should cover less than 3 pixels. We have used the function
## locator() to digitize this barrier and then the function dput to
## have the following limits:

bound <- structure(list(x = c(701751.385381925, 701019.24105475,
                        700739.303517889,
                        700071.760160759, 699522.651915378,
                        698887.40904327, 698510.570051342,
                        698262.932999504, 697843.026694212,
                        698058.363261028),
                        y = c(3161824.03387414,
                        3161824.03387414, 3161446.96718494,
                        3161770.16720425, 3161479.28718687,
                        3161231.50050539, 3161037.5804938,
                        3160294.22044937, 3159389.26039528,
                        3157482.3802813)), .Names = c("x", "y"))

lines(bound, lwd=3)

## We convert bound to SpatialLines:
bound <- do.call("cbind",bound)
Slo1 <- Line(bound)
Sli1 <- Lines(list(Slo1), ID="frontier1")
barrier <- SpatialLines(list(Sli1))

## estimation of the UD
kud <- kernelUD(loc[,1], h=100, grid=100, boundary=barrier)

## Result:
image(kud)

## Have a closer look to Calou:
kud2 <- kud[[2]]
image(kud2, col=grey(seq(1,0,length=15)))
title(main="Home range of Calou")
points(loc[slot(loc,"data")[,1]=="Calou",], pch=3, col="blue")
plot(getverticeshr(kud2, 95), add=TRUE, lwd=2)
lines(barrier, col="red", lwd=3)

Run the code above in your browser using DataLab