Tiled analog search for memory-constrained queries — tiled_analog

Performs analog searches on large raster datasets by dividing the focal region into tiles and processing each tile separately. This reduces memory usage at the cost of increasing compute time. Works with any analog_*() function.

Usage

tiled_analog_search(
  x,
  pool,
  n_tiles,
  fun,
  geog,
  env = NULL,
  y = NULL,
  x_cov = NULL,
  weight = NULL,
  cell_area_weight = "auto",
  ...,
  output_file = NULL,
  progress = TRUE
)

Arguments

x: SpatRaster with focal locations (points to find analogs for).
pool: SpatRaster with reference locations (potential analog pool).
n_tiles: Approximate number of tiles. The function will find a grid close to this number that creates square-ish tiles. Choosing larger values for n_tiles will reduce memory usage, but will also reduce computational efficiency. Choose the smallest n_tiles that fits your memory constraints.
fun: An analog_* function to apply to each tile (e.g., analog_velocity, analog_impact).
geog: A kernel() object giving the geographic search treatment. Its max (the geographic search radius) is required and finite: it both sizes the tile halos (so edge focals still see all in-range analogs) and is forwarded to fun. Tiling is only beneficial when the geographic search is bounded, hence geog with a finite max is required.
env: Optional kernel() object giving the environmental search treatment, forwarded to fun. NULL (default) means no environmental kernel is passed.
y: Optional SpatRaster with values to aggregate across analogs. Must have spatial properties matching pool.
x_cov: Optional SpatRaster with covariates for focal points. Must have spatial properties matching x.
weight: Optional single-layer SpatRaster of per-pool-cell weights. Must have the same CRS and extent as pool. See analog_search() for details.
cell_area_weight: Controls cell-area weighting. One of "auto" (default; on for raster pools, which is always the case here), TRUE, or FALSE. Unlike a non-tiled query, where weights are normalized to mean 1 over whatever pool is passed, here weights are normalized to mean 1 over the full pool raster once at the outer level, and the resulting raster is sliced per tile. This keeps absolute magnitudes of weighted stats (e.g. sum_weights) consistent across tiles.
...: Additional arguments passed to fun (e.g. select, stat, k). May include covariates as a SpatRaster matching pool's CRS and extent (cropped per tile alongside y). May also include normalize for helpers that support it (e.g. analog_density()); when normalization is requested, the global mean cell area is computed once over the full pool and propagated to each tile so per-tile D_max values are consistent.
output_file: Optional filename for disk-based output. If specified and fun returns a SpatRaster, tiles are written to temporary files during processing and merged to output_file at the end. This is useful when results are too large to fit in memory. Ignored for data.frame results.
progress: Logical indicating whether to show progress bar.

Value

Same type as fun returns (SpatRaster or data.frame). If output_file is specified, returns a disk-backed SpatRaster.

Details

Tiled analog searches work by splitting x into a number of smaller tiles and calling the requested analog function on each tile, using an analog pool that is the size of the tile buffered by the geographic radius (geog's max). This buffer is necessary for correctness but increases compute time, particularly if the radius is large. The results for each tile are temporarily written to disk, and are merged into a single results raster once all tiles have processed.

The function requires geog to have a finite max, as tiling is only beneficial when geographic distance constraints limit the reference pool size for each focal point. The function will warn if the radius is so large that tiling provides minimal memory benefit.

If index_res is specified in ..., all tiles will use the same lattice resolution. If index_res is not specified, each tile will independently auto-tune its lattice resolution based on local data characteristics. This adaptive behavior is generally fine and can even be beneficial when environmental distributions vary substantially across the landscape (e.g., mountains vs plains).