Changelog • soilcnn

soilcnn 0.1.0

CRAN release: 2026-10-21

First release.

Patch store from points and aligned rasters (dsm_prepare(), dsm_load()), with a subsample by whole spatial blocks, stratified by region on request.
Resampling specs for spatial blocks, kNNDM, random folds, a holdout and groups (spatial_cv(), knndm_cv(), random_cv(), holdout_cv(), region_cv()), with buffers that keep training patches from touching validation and test patches, and an optional calibration set carved beside the test set by the same criterion (calibration_frac).
Tuning of a dual-branch convolutional network over the plan (dsm_train()), a seed noise floor and the one-standard-error rule (seed_noise_floor(), one_se()), and random forest and multilayer perceptron baselines on the same folds; caret models through caret_spec().
A frozen test set scored only after the choice (freeze_selection(), score_test_grid()).
The selected configuration refitted under several seeds (dsm_final()), and maps of any extent (dsm_predict()): ensemble median and spread, the mean by smearing (smear(), smear_map(), smearing_check()), and the area of applicability (aoa_reference(), dissimilarity_index()).
Conformal intervals calibrated three ways, every one checked on the test set by dsm_final() – overall, by group, inside and outside the area of applicability, and by level: on the cross-validated residuals; split conformal on the calibration set; and CV+ with the tuning run’s fold models at each pixel (conformal_calibrate(), conformal_scaled_calibrate(), conformal_scale_fit(), cv_plus_calibrate(), cv_plus_interval()). Each at a constant width or one that grows with the level and the dissimilarity index, and with points or whole groups weighing alike.
Variable importance by permutation, SHAP (expected gradients, integrated gradients, kernel), SAGE, refits without each variable and accumulated local effects (dsm_importance() and its methods), with maps of SHAP values.
Intervals for clustered test points from whole blocks of equal area (block_bootstrap(), equal_area_blocks(), spatial_correlogram()).
A small synthetic landscape for examples and tests (example_landscape(), example_run()).
Every final model and every map records the soilcnn version that made it – with its GitHub commit, when installed from there – beside the versions of torch and R (terra, for a map) and the commit of the project it was written into.