Skip to contents

Changelog

Source: NEWS.md

MLSampling 0.0.3.9000 (development version)

Work towards a CRAN submission. R CMD check --as-cran went from 1 ERROR / 5 WARNINGs / 4 NOTEs to 1 ERROR / 1 WARNING / 2 NOTEs.

Packaging

  • DESCRIPTION now uses Authors@R. The maintainer had been recorded as “Carlos”, a first name only, which CRAN rejects.

  • Removed the Remotes field. pryr is on CRAN, so the field was never needed and it was pushing pryr out of the mainstream repositories.

  • Moved 14 never-referenced packages from Imports to Suggests, leaving 12. ggplot2 and viridis stay in Imports because they are used through import() in NAMESPACE. Declared units, which is called in validate_field_boundary_geometry() but was missing.

  • Dropped quickcheck, which is archived on CRAN, and the obsolete Type and LazyData fields.

Bug fixes

Documentation

  • Removed the Rd pages for execute_udl_optimization and execute_ufn_optimization, which documented private R6 methods as if they were exported functions, and the Rd page for %||%, whose \name contained a pipe character and which is not exported.

MLSampling 0.0.3

Bug fixes

  • Random Forest classification works again. calculate_spatial_features() computed the spatial lag as weights %*% values, which fails for a factor target, and spatial_autocorr defaults to TRUE, so classification failed under the default configuration. A factor target now produces the neighbourhood class composition instead: one spatial_lag_<level> column per level holding the inverse distance weighted proportion of neighbours in that class. Numeric targets keep the single spatial_lag column and are unchanged.

Testing

  • The Random Forest property tests no longer wrap their assertions in tryCatch(), which was turning expectation failures into informational messages and reporting a passing suite. This had been masking both the classification failure above and the covariate assertion on the feature importance table.

MLSampling 0.0.2

Impact on existing results

Models fitted with version 0.0.1 were trained without any environmental covariates, and when target_variable was left unset they were trained against a spatial coordinate rather than the measured property. Any result produced with 0.0.1 should be regenerated with this version.

Bug fixes

  • Covariate extraction no longer discards the first covariate. terra::extract() returns an ID column only for SpatVector input, but every call site passes a coordinate matrix, so the unconditional [, -1, drop = FALSE] removed a real covariate instead. With a single-layer raster the feature set was emptied entirely and the models trained on no covariates at all. Fixed in the Random Forest, Bayesian Deep Learning and design comparison modules.

  • Random Forest no longer trains on the x coordinate. When target_variable was not supplied, prepare_training_data() kept the x and y columns in the sample data, so the automatic target detection selected x as the value to model.

  • Bayesian Deep Learning carried the same defect in its own prepare_training_data() and is fixed the same way. Coordinates remain available as model features through spatial_encoding; only the target selection changed.

  • Hyperparameter tuning no longer yields mtry = 0 for a single-covariate feature set, which randomForest silently reset while config_used reported the unusable value.

Behaviour changes

  • Sample data whose only numeric columns are x and y now fails with “No numeric target variable found” instead of silently training against a coordinate. Supply a column holding the measured property, or pass target_variable explicitly.

Documentation

  • Added the LICENSE file required by the MIT + file LICENSE declaration in DESCRIPTION, which was previously missing.

  • Added status badges to the README.

  • The README sampling examples now include a numeric target column, so they run as written.

MLSampling 0.0.1

  • Initial release.