Build a reusable nearest-neighbor structure

`gwr_neighbors()` computes nearest neighbors once and stores them in the target-contiguous layout used by the C++ fitting kernels. Reusing this object avoids repeating spatial searches across models and tuning parameters.

Usage

gwr_neighbors(
  coords,
  query_coords = NULL,
  k,
  include_self = is.null(query_coords),
  algorithm = c("kd_tree", "cover_tree", "CR", "brute")
)

Arguments

coords

Numeric coordinate matrix with observations in rows, or an `sf` point object. Optional matrix `"crs"` metadata is retained.

query_coords

Optional matrix or `sf` point coordinates at new target locations. When supplied, neighbors are selected from `coords` for every row of `query_coords`.

k

Number of retained neighbors, including the target itself when `include_self = TRUE`.

include_self

Whether each observation is inserted as its own first neighbor.

algorithm

Nearest-neighbor algorithm passed to [FNN::get.knn()].

Value

An object of class `gwrs_neighbors`.

Examples

quake <- datasets::quakes[1:30, ]
coords <- as.matrix(quake[, c("long", "lat")])

neighbors <- gwr_neighbors(coords, k = 18)
neighbors
head(neighbors$index[, 1:3])

query <- coords[1:3, , drop = FALSE] + 0.01
cross_neighbors <- gwr_neighbors(
  coords, query_coords = query, k = 12, include_self = FALSE
)
dim(cross_neighbors$index)