Sampling pseudo-absences for the training and testing datasets.

paSampling performs a two-step procedure for uniformly sampling pseudo-absences within the environmental space. In the initial step, a kernel-based filter is utilized to determine the subset of the environmental space that will be subsequently sampled. The kernel-based filter calculates the probability function based on the presence observations, enabling the identification of areas within the environmental space that likely exhibit suitable conditions for the species. To achieve this, a probability threshold value is utilized to assign observations to the corresponding portion of the environmental space. These areas, deemed to have suitable environmental conditions, are excluded from the subsequent uniform sampling process conducted in the second step using the uniformSampling function, which is internally called. The bandwidth of the kernel can be automatically estimated from the presence observations or directly set by the user, providing flexibility in determining the scope and precision of the filter.

Usage

paSampling(
  env.rast = NULL,
  pres = NULL,
  thres = 0.75,
  H = NULL,
  grid.res = NULL,
  n.tr = 5,
  sub.ts = FALSE,
  n.ts = 5,
  prev = NULL,
  plot_proc = FALSE,
  verbose = FALSE
)

Arguments

env.rast: A RasterStack, RasterBrick or a SpatRaster object comprising the variables describing the environmental space.
pres: A SpatialPointsDataframe, a SpatVector or an sf object including the presence-only observations of the species of interest.
thres: (double) This value identifies the quantile value used to specify the boundary of the kernel density estimate (default thres=0.75 ). Thus, probability values higher than the threshold should indicate portions of the multivariate space likely associated with presence points.
H: The kernel bandwidth (i.e., the width of the kernel density function that defines its shape) excluding the portion of the environmental space associated with environmental conditions likely suitable for the species. It can be either defined by the user or automatically estimated by paSampling via ks::Hpi.
grid.res: (integer) resolution of the sampling grid. The resolution can be arbitrarily selected or defined using the optimRes function.
n.tr: (integer) number of pseudo-absences for the training dataset to sample in each cell of the sampling grid
sub.ts: (logical) sample the validation pseudo-absences
n.ts: (integer; optional) number of pseudo-absences for the testing dataset to sample in each cell of the sampling grid. sub.ts argument must be TRUE.
prev: (double) prevalence value to be specified instead of n.tr and n.ts
plot_proc: (logical) plot progress of the sampling, default FALSE
verbose: (logical) Print verbose

Value

An sf object with the coordinates of the pseudo-absences both in the geographical and environmental space.

Details

Being designed with species distribution models in mind, paSampling allows collectively sampling pseudo-absences for both the training and testing dataset (optional). In both cases, the user must provide a number of observations that will be sampled in each cell of the sampling grid (n.tr: points for the training dataset; n.ts: points for the testing dataset). Note that the optimal resolution of the sampling grid can be found using the optimRes function. Also, note that the number of pseudo-absences eventually sampled in each cell by the internally-called uniformSampling function depends on the spatial configuration of the observations within the environmental space. Indeed, in most cases some cells of the sampling grid will be empty (i.e., those at the boundary of the environmental space). For this reason, the number of pseudo-absences returned by paSampling is likely to be lower than the product between the number of cells of the sampling gird and n.tr(or n.ts).