Computation of spatial data by hierarchical and objective partitioning of inputs for parallel processing

The chopin package provides a set of functions to compute on divided geospatial data.

Basic functionalities

• Distribute terra, sf, and chopin functions to parallel workers set by future or mirai

• Set parallelization strategies based on artificial grids, equal-size clusters, hierarchy, and multiple raster files

• Convenience functions for raster-vector overlay and weighted summary from vector dataset

chopin workflow

• The simplest way of parallelizing generic geospatial computation is to start from par_pad_* functions to par_grid, or running par_hierarchy, or par_multirasters functions at once.

library(chopin)
library(terra)
library(sf)
library(collapse)
library(dplyr)
library(future)
library(future.mirai)
library(future.apply)
lastpar <- par(mfrow = c(1, 1))
options(sf_use_s2 = FALSE)

Example data

• North Carolinian counties and a raster of elevation data are used as example data.

temp_dir <- tempdir(check = TRUE)

url_nccnty <-
  paste0(
   "https://raw.githubusercontent.com/",
   "ropensci/chopin/refs/heads/main/",
   "tests/testdata/nc_hierarchy.gpkg"
 )

url_ncelev <-
  paste0(
   "https://raw.githubusercontent.com/",
   "ropensci/chopin/refs/heads/main/",
   "tests/testdata/nc_srtm15_otm.tif"
 )

nccnty_path <- file.path(temp_dir, "nc_hierarchy.gpkg")
ncelev_path <- file.path(temp_dir, "nc_srtm15_otm.tif")

# download data
download.file(url_nccnty, nccnty_path, quiet = TRUE)
download.file(url_ncelev, ncelev_path, quiet = TRUE)
nccnty <- terra::vect(nccnty_path)
ncelev <- terra::rast(ncelev_path)

• To demonstrate chopin functions, we generate 10,000 random points in North Carolina

ncsamp <-
  terra::spatSample(
    nccnty,
    1e4L
  )
ncsamp$pid <- 1:nrow(ncsamp)

Creating grids

• The example below will generate a regular grid from the random point data.

ncgrid <- par_pad_grid(ncsamp, mode = "grid", nx = 4L, ny = 2L, padding = 10000)
plot(ncgrid$original)

Extracting values from raster

• Since all par_* functions operate on future backends, users should define the future plan before running the functions. multicore plan supports terra objects which may lead to faster computation, but it is not supported in Windows. An alternative is future.mirai's mirai_multisession plan, which is supported in many platforms and generally faster than plain future multisession plan.

• workers argument should be defined with an integer value to specify the number of threads to be used.

future::plan(future.mirai::mirai_multisession, workers = 2L)

• Then we dispatch multiple extract_at runs on the grid polygons.

• Before we proceed, the terra object should be converted to sf object.

pg <-
  par_grid(
    grids = ncgrid,
    pad_y = FALSE,
    .debug = TRUE,
    fun_dist = extract_at,
    x = ncelev_path,
    y = sf::st_as_sf(ncsamp),
    id = "pid",
    radius = 1e4,
    func = "mean"
  )

Hierarchical processing

• Here we demonstrate hierarchical processing of the random points using census tract polygons.

nccnty <- sf::st_read(nccnty_path, layer = "county")
nctrct <- sf::st_read(nccnty_path, layer = "tracts")

• The example below will parallelize summarizing mean elevation at 10 kilometers circular buffers of random sample points by the first five characters of census tract unique identifiers, which are county codes.

• This example demonstrates the hierarchy can be defined from any given polygons if the unique identifiers are suitably formatted for defining the hierarchy.

px <-
  par_hierarchy(
    # from here the par_hierarchy-specific arguments
    regions = nctrct,
    regions_id = "GEOID",
    length_left = 5,
    pad = 10000,
    pad_y = FALSE,
    .debug = TRUE,
    # from here are the dispatched function definition
    # for parallel workers
    fun_dist = extract_at,
    # below should follow the arguments of the dispatched function
    x = ncelev,
    y = sf::st_as_sf(ncsamp),
    id = "pid",
    radius = 1e4,
    func = "mean"
  )

Multiraster processing

• Here we demonstrate multiraster processing of the random points using multiple rasters.

ncelev <- terra::rast(ncelev_path)
tdir <- tempdir(check = TRUE)
terra::writeRaster(ncelev, file.path(tdir, "test1.tif"), overwrite = TRUE)
terra::writeRaster(ncelev, file.path(tdir, "test2.tif"), overwrite = TRUE)
terra::writeRaster(ncelev, file.path(tdir, "test3.tif"), overwrite = TRUE)
terra::writeRaster(ncelev, file.path(tdir, "test4.tif"), overwrite = TRUE)
terra::writeRaster(ncelev, file.path(tdir, "test5.tif"), overwrite = TRUE)
rasts <- list.files(tdir, pattern = "tif$", full.names = TRUE)

pm <-
  par_multirasters(
    filenames = rasts,
    fun_dist = extract_at,
    x = NA,
    y = sf::st_as_sf(ncsamp)[1:500, ],
    id = "pid",
    radius = 1e4,
    func = "mean",
    .debug = TRUE
  )
par(lastpar)

Function selection guide for par_*()

We provide two flowcharts to help users choose the right function for parallel processing. The raster-oriented flowchart is for users who want to start with raster data, and the vector-oriented flowchart is for users with large vector data.

In raster-oriented selection, we suggest four factors to consider:

• Number of raster files: for multiple files, par_multirasters is recommended. When there are multiple rasters that share the same extent and resolution, consider stacking the rasters into multilayer SpatRaster object by calling terra::rast(filenames).

• Raster resolution: We suggest 100 meters as a threshold. Rasters with resolution coarser than 100 meters and a few layers would be better for the direct call of exactextractr::exact_extract().

• Raster extent: Using SpatRaster in exactextractr::exact_extract() is often minimally affected by the raster extent.

• Memory size: max_cells_in_memory argument value of exactextractr::exact_extract(), raster resolution, and the number of layers in SpatRaster are multiplicatively related to the memory usage.

For vector-oriented selection, we suggest three factors to consider:

• Number of features: When the number of features is over 100,000, consider using par_grid or par_hierarchy to split the data into smaller chunks.

• Hierarchical structure: If the data has a hierarchical structure, consider using par_hierarchy to parallelize the operation.

• Data grouping: If the data needs to be grouped in similar sizes, consider using par_pad_balanced or par_pad_grid with mode = "grid_quantile".

Caveats

Why parallelization is slower than the ordinary function run? Parallelization may underperform when the datasets are too small to take advantage of divide-and-compute approach, where parallelization overhead is involved. Overhead here refers to the required amount of computational resources for transferring objects to multiple processes. Since the demonstrations above use quite small datasets, the advantage of parallelization was not as noticeable as it was expected. Should a large amount of data (spatial/temporal resolution or number of files, for example) be processed, users could find the efficiency of this package. A vignette in this package demonstrates use cases extracting various climate/weather datasets.

Notes on data restrictions

chopin works best with two-dimensional (planar) geometries. Users should disable s2 spherical geometry mode in sf by setting sf::sf_use_s2(FALSE). Running any chopin functions at spherical or three-dimensional (e.g., including M/Z dimensions) geometries may produce incorrect or unexpected results.

See also

Useful links:

• https://docs.ropensci.org/chopin/

• https://github.com/ropensci/chopin

• Report bugs at https://github.com/ropensci/chopin/issues

Author

Maintainer: Insang Song geoissong@gmail.com (ORCID)

Authors:

• Kyle Messier (ORCID) [contributor]

Other contributors:

• Alec L. Robitaille (Alec reviewed the package version 0.6.3 for rOpenSci, see <https://github.com/ropensci/software-review/issues/638>) [reviewer]

• Eric R. Scott (Eric reviewed the package version 0.6.3 for rOpenSci, see <https://github.com/ropensci/software-review/issues/638>) [reviewer]