Package index • soilcnn

The chain

Points and rasters to a patch store, the store to a tuned network, the chosen configuration to a final model under several seeds, the model to maps – and what the model learned.

dsm_prepare(): Build a patch store from a point table and a folder of aligned rasters.
dsm_load(): Load everything a run needs, and check that it fits together.
dsm_train(): Fit and tune a model over a resampling plan.
dsm_final(): Refit the configuration a tuning run selects, under N seeds.
dsm_predict(): Predict a map from a final model.
dsm_importance(): Variable importance of a fitted model, by the method you choose.
dsm_report_final(): The declaration of a final model that already exists.

Example data

A small synthetic landscape, and a run fitted on it once a session.

example_landscape(): A small synthetic landscape, for the examples and the tests.
example_run(): A small fitted run on the example landscape, made once a session.

Who trains and who scores

Resampling designs, the fold plans they cut – with a test set and an optional calibration set carved by the same criterion – and the checks of a plan.

spatial_cv(): Spatially blocked k-fold: whole blocks of ground go to one fold.
knndm_cv(): Folds matched to where the map will be predicted (kNNDM).
random_cv(): Random k-fold. Ignores geography by construction.
holdout_cv(): A single train/validation/test split.
region_cv(): Leave-region-out, on a grouping that already exists (biome, catchment, ...).
resolve_resampling(): Turn a spec into a fold plan against real points.
spatial_folds(): Spatially blocked folds of a point table.
knndm_folds(): Fold plan from k-fold Nearest Neighbour Distance Matching.
random_folds(): Random k-fold over the train+validation pool.
holdout(): A single train/validation/test split of a point table.
region_folds(): Fold by a grouping column: no group is split across folds.
prediction_sample(): A regular sample of the prediction area, as kNNDM's predpoints.
check_fold_plan(): Check a plan's folds are well formed, and describe them.
fold_leakage_report(): How much does each fold still leak?

Models, grids and devices

list_models(): Every registered model, as a table.
get_model(): Fetch a registered model.
model_spec(): Describe a model the framework can fit.
register_model(): Add a model to the registry.
caret_spec(): Turn a caret method into a model_spec.
caret_available(): Every caret method that could be borrowed here.
make_tune_grid(): Generate a random hyperparameter grid (like caret's tuneLength).
make_manual_tune_grid(): Build a manual grid from explicit lists of values per parameter.
target_transform_spec(): The forward and inverse functions for a named target transform.
setup_torch_device(): Configure torch threads and select compute device.
set_torch_threads(): Set torch's thread pools.

Choosing a configuration

Repetitions summarised, the noise floor of the seeds, the one-standard-error rule, and a test set scored only after the choice is frozen.

summarise_resamples(): Aggregate a unit-level comparison table into one row per config.
seed_noise_floor(): How much does a metric move when ONLY the seed changes?
print_noise_floor(): Print a noise-floor report in the terms it should be read in.
one_se(): Pick the simplest config within one standard error of the best.
print_one_se(): Say what one_se() did, including when it did nothing.
paired_family_test(): Paired comparison between two model families.
freeze_selection(): Record which config was chosen, and when.
score_test_grid(): Score every trained unit of a tuning run on the held-out test set.
selected_config_id(): The configuration a final run deployed.
latest_run_dir(): The most recent run under base, by time, and only if it finished.

Intervals

Conformal intervals calibrated on cross-validated residuals, on a calibration set (split conformal) or with every fold’s model (CV+), at a constant width or one that follows the level and the dissimilarity index, with points or whole groups weighing alike – and their coverage.

conformal_calibrate(): Calibrate a conformal interval from held-out residuals.
conformal_interval(): Turn predictions into intervals.
conformal_scale_fit(): Fit the scale of an interval: |residual| on covariates that say how wrong.
conformal_scaled_calibrate(): Calibrate an interval whose width is a fitted scale.
conformal_scaled_interval(): Turn predictions into intervals with a fitted scale.
cv_plus_calibrate(): Calibrate a CV+ interval from the folds of a cross-validation.
cv_plus_interval(): Turn the fold models' predictions into CV+ intervals.
conformal_cv(): Calibrate on the folds of a resampling run, then report coverage.
cv_residuals(): Cross-validated residuals from a tuning run, for calibration.
cv_residuals_for_config(): Cross-validated residuals of a configuration, from any tuning run.
picp_report(): Coverage, and whether it holds where it is needed.

The mean surface

The conditional mean of a log-scale model, by Duan’s smearing.

smear(): The conditional MEAN surface, from transform-space predictions.
smear_map(): The mean surface of a median map already made.
smearing_factor(): The smearing factor, from held-out residuals in transform space.
smearing_from_run(): The smearing factor from a tuning run's out-of-fold predictions.
smearing_check(): The smearing factors on held-out points: which one unbiases the mean surface.

Where the map may be believed

The dissimilarity index and the area of applicability (Meyer and Pebesma 2021).

aoa_reference(): The dissimilarity reference of a fitted model.
aoa_di(): The DI of new rows of RAW predictor values, QC'd and scaled as the reference was.
aoa_threshold(): The DI threshold that separates the area of applicability.
inside_aoa(): Is each row inside the area of applicability?
print_aoa(): Report what an AOA covers.
di_reference(): Summarise a training set for dissimilarity-index computation.
dissimilarity_index(): The dissimilarity index of new data.

What the model learned

Importance by permutation, SHAP, SAGE, refits and the patch’s context; accumulated local effects; SHAP over the map.

permutation_importance(): Permutation importance: how much a fitted model relies on each variable.
shap_importance(): SHAP importance: how each variable pushes each prediction, up or down.
sage_importance(): SAGE: how much of the model's skill each variable carries, shared fairly.
refit_importance(): Refit importance: what the model loses when trained without a variable.
context_importance(): Context importance: how far from the point, and at which scale, the model reads.
ale_effect(): ALE: how the prediction changes along each variable's range.
compare_importance(): Several importances, side by side.
importance_groups(): Which channels the importance treats as one variable.
importance_points(): Points of the map to explain: a regular grid of the model's raster cells.
importance_map(): Maps of SHAP values, from an importance computed at points of the map.
importance_weights(): One weight per channel, for the area of applicability, from an importance.
occlusion_report(): What each part of the patch is worth to one trained unit of a tuning run.
plot(<dsm_importance>): Plot a variable importance.
plot(<importance_comparison>): Plot importances side by side.
plot(<importance_map>): Plot maps of SHAP values.

Errors on clustered points

Intervals for a difference in error, drawn from whole blocks of equal area.

block_bootstrap(): A metric, or two models' difference in it, with an interval from whole blocks.
block_bootstrap_by_size(): The interval as the blocks grow.
equal_area_blocks(): Square blocks of equal area, in kilometres.
spatial_correlogram(): Moran's I by distance class: how far apart two points stop resembling each other.
plot(<block_bootstrap_by_size>): Plot the interval as the blocks grow.
plot(<spatial_correlogram>): Plot a spatial correlogram.

Metrics

calc_metrics(): Compute all regression metrics for one obs/pred pair.
ccc(): Lin's Concordance Correlation Coefficient.