Changes in version 0.2.0 (2026-10-04)                  

First CRAN release.

  - Model settings are now given per model with params, using Earth
    Engine's argument names, for example params = list(gbt =
    list(shrinkage = 0.01)). This replaces the separate arguments
    n_trees, shrinkage, svm_cost, knn_k and the rest, which applied to
    several models at once and could silently override values equal to
    their defaults. Misspelt settings are now an error.
  - Every model uses Earth Engine's defaults, except the tree counts
    Earth Engine requires (rf 500, gbt 150) and knn, which uses 15
    neighbours because Earth Engine's single neighbour gives a two-value
    map. The SVM is now Earth Engine's default linear classifier, and
    boosted trees use Earth Engine's learning rate, which gives
    better-calibrated probabilities.
  - Fixed classification SVMs (C_SVC, NU_SVC) returning inverted scores.
    Their probability depended on the order of the training rows, which
    Earth Engine does not preserve; the training table is now sorted
    before fitting.
  - generate_pseudo_absences() now reads pseudo-absences from the same
    years as the presences, in the same proportions, instead of the
    latest embedding year. Pass aoi_year to place them all in one year.
  - Models are trained on their rows in a fixed order, so retraining
    (which Earth Engine does for every map tile) gives the same model.
    Random forests are now exactly reproducible; boosted trees vary by
    at most about 0.005 inside Earth Engine.
  - Simpler arguments: balance_trees is gone (use bg_ratio, default 1,
    or NULL for every absence), as are persist_classifier, async and
    options from evaluate_models(), which now requires predict_coords
    with a year column. generate_map() maps the records' most common
    year by default.
  - sdm_gee_status() is now gee_tasks() and sdm_clean_assets() is
    gee_clean_assets(). sdm_verbose() and the exported calculate_cbi()
    are gone: use suppressMessages(), and the cbi from
    calculate_classifier_metrics().
  - Removed naive Bayes, which discards negative values and so cannot
    use the signed embeddings.
  - The vignette now tests models on the following year's records.
  - Maps download straight from Earth Engine in tiles sized to its 32 MB
    download limit, fetched two at a time and stitched into one GeoTIFF
    per model. Google Drive is used only for maps Earth Engine will not
    compute tile by tile.
  - generate_map() computes the ensemble on Earth Engine and accepts aoi
    = "bbox".
  - The Earth Engine Python client is declared with
    reticulate::py_require(), so reticulate installs it when needed;
    AlphaSDM no longer installs Python itself and no longer depends on
    'rgee'.
  - The saved project ID lives in tools::R_user_dir("AlphaSDM",
    "config"). clear_gee_credentials() removes it and asks before
    signing out of Earth Engine.
  - Connecting to Earth Engine no longer deletes old temporary assets;
    use gee_clean_assets().
  - A getting-started vignette, vignette("AlphaSDM"), maps saguaro
    around Tucson from GBIF records.
  - Removed unused code, including the remains of built-in
    cross-validation.