• Fixed bug where passing a polygon to spatial_nndm_cv() forced leave-one-out
  CV, rather than the intended sampling of prediction points from the polygon.

• spatial_block_cv() now adds an expand_bbox attribute to the resulting rset
  for compatibility with rsample::reshuffle_rset()

• autoplot.spatial_block_cv() now plots the proper grid (using the new
  expand_bbox attribute).

• spatial_block_cv() gains an argument, expand_bbox, which represents the
  proportion a bounding box should be expanded by (each corner of the bounding
  box is expanded by bbox_corner_value * expand_bbox).

  • This is a breaking change for data in planar coordinate reference
    systems. Set to 0 to obtain previous behaviors.
  • Data in geographic coordinates was already having its bounding box expanded
    by the default 0.00001.
  • This makes it so that regularly spaced data is less likely to fall precisely
    along grid lines (and therefore fall into two assessment sets) and so that
    geographic data falls is more likely to fall within the constructed grid.
  • Thanks to Nikos on StackOverflow for reporting this behavior:
    https://stackoverflow.com/q/77374348/9625040

• spatial_block_cv() will now throw an error if observations are in multiple
  assessment folds (caused by observations, or observation centroids, falling
  precisely along grid polygon boundaries).

• In spatial_nndm_cv(), passing a single polygon (or multipolygon) to the
  prediction_sites argument will result in prediction sites being sampled from
  that polygon, rather than from its bounding box.

• get_rsplit() is now re-exported from the rsample package. This provides a
  more natural, pipe-able interface for accessing individual splits;
  get_rsplit(rset, 1) is identical to rset$splits[[1]].

• spatial_nndm_cv() is a new function for nearest neighbor distance matching
  cross-validation, as described in Milà et al. 2022
  (doi: 10.1111/2041-210X.13851). NNDM was first implemented in CAST
  (https://cran.r-project.org/package=CAST).

Breaking changes

• spatial_clustering_cv() no longer accepts non-sf objects. Use
  rsample::clustering_cv() for these instead (#126).

• spatial_clustering_cv() now uses edge-to-edge distances, like the rest of
  the package, rather than centroids (#126).

New features

• All functions now have a repeats argument, defaulting to 1, allowing for
  repeated cross-validation (#122, #125, #126).

• spatial_clustering_cv() now has a distance_function argument, set by
  default to as.dist(sf::st_distance(x)) (#126).

Minor improvements and fixes

• Outputs from spatial_buffer_vfold_cv() should now have the correct radius and buffer attributes (#110).

• spatial_buffer_vfold_cv() now has the correct id values when using repeats (#116).

• spatial_buffer_vfold_cv() now throws an error when repeats > 1 && v >= nrow(data) (#116).

• The minimum sf version required is now >= 1.0-9, so that unit objects can be passed to cellsize in spatial_block_cv() (#113; #124).

• autoplot() now handles repeated cross-validation properly (#123).

• Mike Mahoney is taking over as package maintainer, as Julia Silge (who remains
  a package author) moves to focus on ModelOps work.

• Functions will now return rsplits without out_id, like most rsample
  functions, whenever buffer is NULL.

• spatial_block_cv(), spatial_buffer_vfold_cv(), and buffering now support
  using sf or sfc objects with a missing CRS. The assumption is that data in an
  NA CRS is projected, with all distance values in the same unit as the
  projection. Trying to use alternative units will fail. Set a CRS if these
  assumptions aren't correct.

• spatial_buffer_vfold_cv() and buffering no longer support tibble or
  data.frame inputs (they now require sf or sfc objects). It was not easy to
  use these to begin with, but should have always caused an error: use
  rsample::vfold_cv() instead or transform your data into an sf object.

• spatial_buffer_vfold_cv() has had some attribute changes to match rsample:

  • strata attribute is now the name of the column used for stratification,
    or not set if there was no stratification.
  • pool and breaks have been added as attributes
  • radius and buffer are now set to 0 if they were passed as NULL.

New features

• spatial_buffer_vfold_cv() is a new function which wraps
  rsample::vfold_cv(), allowing users to add inclusion radii and exclusion
  buffers to their vfold resamples. This is the supported way to perform
  spatially buffered leave-one-out cross validation (set v to nrow(data)).

• spatial_leave_location_out_cv() is a new function with wraps
  rsample::group_vfold_cv(), allowing users to add inclusion radii and
  exclusion buffers to their vfold resamples.

• spatial_block_cv() is a new function for performing spatial block
  cross-validation. It currently supports randomly assigning blocks to folds.

• spatial_clustering_cv() gains an argument, cluster_function, which
  specifies what type of clustering to perform. cluster_function = "kmeans",
  the default, uses stats::kmeans() for k-means clustering, while
  cluster_function = "hclust" uses stats::hclust() for hierarchical
  clustering. Users can also provide their own clustering function.

• spatial_clustering_cv() now supports sf objects! Coordinates are inferred
  automatically when using sf objects, and anything passed to coords will
  be ignored with a warning. Clusters made using sf objects will take
  coordinate reference systems into account (using sf::st_distance()),
  unlike those made using data frames.

• All resampling functions now support spatial buffering using two arguments.
  radius lets you specify an inclusion radius for your test set, where any
  data within radius of the original assessment set will be added to the
  assessment set. buffer specifies an exclusion buffer around the test set,
  where any data within buffer of the assessment set (after radius is
  applied) will be excluded from both sets.

• autoplot() now has a method for spatial resamples built from sf objects.
  It works both on rset objects and on rsplit objects, and has a special
  method for outputs from spatial_block_cv().

• boston_canopy is a new dataset with data on tree canopy change over time in
  Boston, Massachusetts, USA. It uses a projected coordinate reference system
  and US customary units; see ?boston_canopy for instructions on how to
  install these into your PROJ installation if needed.

Documentation

• The "Getting Started" vignette has been revised to demonstrate the new
  features and clustering methods.

• A new vignette has been added walking through the spatial buffering process.

Dependency changes

• R versions before 3.4 are no longer supported.

• glue, sf, and units have been added to Imports.

• ggplot2 has been moved to Imports. It had been in Suggests.

• covr, gifski, lwgeom, and vdiffr are now in Suggests.

• rlang now has a minimum version of 1.0.0 (was previously unversioned).

• First release of spatialsample