Spatially Thin Occurrence Data — borg

Reduces spatial clustering by ensuring a minimum distance between all retained observations. Uses iterative nearest-neighbor removal: at each step, the point with the smallest nearest-neighbor distance is removed until all pairwise distances exceed min_dist.

Usage

borg_thin(data, coords = NULL, min_dist = NULL, verbose = FALSE)

Arguments

data: A data frame, sf object, or terra::SpatVector.
coords: Character vector of length 2. Coordinate column names. Required for data frames; ignored for sf/SpatVector.
min_dist: Numeric. Minimum distance between retained points. Units match the coordinate system (degrees for lat/lon, meters for projected CRS). If NULL, uses 10 percent of the nearest-neighbor interquartile range as a default.
verbose: Logical. Print thinning progress. Default: FALSE.

Value

The input data filtered to retained rows. Has attribute "thinned_idx" with the original row indices of kept observations, and "n_removed" with the count of removed points.

Details

Spatial thinning is standard practice in species distribution modelling to reduce sampling bias. Clustered occurrences inflate apparent model performance and bias spatial CV fold sizes.

For geographic coordinates (lat/lon), distances are computed using the Haversine formula (meters). For projected coordinates, Euclidean distance is used.

Examples

# Thin clustered points to 5-unit minimum spacing
set.seed(42)
d <- data.frame(
  x = c(rnorm(50, 0, 1), rnorm(50, 10, 1)),
  y = c(rnorm(50, 0, 1), rnorm(50, 10, 1)),
  species = "A"
)
d_thin <- borg_thin(d, coords = c("x", "y"), min_dist = 1)
nrow(d_thin)  # fewer rows