LoginSign Up
← Back to all sparks
W

waywiser

ANALYTICS
Velocity0.0

Ergonomic tooling to assess models of spatial data

github.com/ropensci/waywiser ↗

Spatial model assessment that spent the last year on cross-platform arithmetic and CRAN rules.

spatial-statisticsmodel-assessmenttidymodelscran-compliancer-package
◆Current state
waywiser provides spatial model assessment metrics in a tidymodels idiom — spatial autocorrelation measures, area of applicability, and multi-scale assessment of predictions. The substantive work landed in 0.3.0 through 0.5.0, and the recent releases are consolidation: 0.6.0 made metric functions return NA everywhere they previously returned NaN, because macOS disagreed with every other platform, and taught ww_multi_scale() to handle classification and class probability metrics correctly when given rasters. The three releases since are entirely CRAN policy compliance — no internet downloads during checks, no writing to directories, no syntax that would raise the R version floor.
◆Where it's heading
The package has reached the point where the interesting bugs are cross-platform and cross-package rather than statistical. Its main function, ww_multi_scale(), has been the focus of nearly every release since 0.4.0, working through units handling, aggregation ordering, raster inputs and metric-type dispatch. The dependency on vip and the tidymodels metric machinery means a share of releases exist only to track breaking changes elsewhere.
◆Prediction
Expect the next substantive release to continue on ww_multi_scale() edge cases, given that it has absorbed most of the fixes in this window. The recent run of CRAN-compliance patches suggests no feature work is currently in flight.

◆Recent moves

  1. 1y ago

    Stops downloading data during CRAN checks

    Removes network access from the CRAN check path. The third consecutive release spent on check-environment policy rather than the package's own behaviour.

    View source ↗
  2. 1y ago

    Vignettes no longer write to CRAN directories

    Sets sf_proj_search_paths() in vignettes so builds do not write outside permitted locations. Packaging hygiene with no user-visible effect.

    View source ↗
  3. 1y ago

    Keeps the R version floor below 4.1

    Drops the shorthand lambda syntax so the package still installs on R before 4.1, and updates tests for new dependency versions. Compatibility maintenance only.

    View source ↗
  4. 2y ago

    NaN results become NA; raster metrics dispatch correctly

    Metric functions now return NA wherever they previously returned NaN, resolving a genuine cross-platform inconsistency where macOS results differed from everywhere else. ww_multi_scale() also converts truth and estimate to factors for classification metrics and errors rather than guessing when class and probability metrics are mixed.

    View source ↗
  5. 2y ago

    Guards against ignored grid arguments; faster on sf data

    Adds warnings and errors for two cases where user arguments were silently ignored — a crs passed through to st_make_grid(), and extra arguments supplied alongside an explicit grid list — and speeds up the sf path. Turning silent no-ops into errors is the right trade for an assessment tool where a wrong grid means wrong numbers.

    View source ↗
  6. 2y ago

    Fixes wrong observation counts and ignored grid units

    Two real correctness bugs: truth and estimate counts always came back as 1 for sf data because they were computed after aggregation, and grid cellsizes carrying units were ignored. Both produced plausible-looking output that was wrong, which is the worst failure mode for a metrics package.

    View source ↗