feat: add probability prediction to predict_spatial()

[be-marc]

Closes #88
Closes #89

Summary

Reimplementation of #89 on current main. predict_spatial() now supports probability predictions. Instead of adding a predict_type argument, the prediction type is taken from the $predict_type of the learner, so a mismatch between argument and learner state is impossible.

• LearnerClassifSpatial mirrors the wrapped learner's $predict_type and fills a full-size probability matrix with NaN rows for incomplete observations. Only the column names are carried over, which fixes the dimension error with masked rasters reported in Support for plotting probabilities in predict_spatial #88.
• For rasters, the target raster is initialized with one layer per class and terra::writeValues() receives the full probability matrix, so multiclass tasks are supported. Layers are named after the class levels, and the categorical-labels step is skipped for probability output.
• For sf vectors, one probability column per class is returned. The vector branch now uses $predict_newdata() instead of $predict() because predicting probabilities directly on an unsupervised task fails mlr3's prediction-data check.
• The internal spatial learner wrappers now use the id of the wrapped learner instead of the hardcoded classif.ranger / regr.ranger.

Verification

library(mlr3spatial)
library(terra, exclude = "resample")

task_train = tsk("leipzig")
learner = lrn("classif.rpart", predict_type = "prob")
learner$train(task_train)

stack = rast(system.file("extdata", "leipzig_raster.tif", package = "mlr3spatial"))
pred = predict_spatial(stack, learner, chunksize = 1L)
pred # SpatRaster with layers forest, pasture, urban, water
plot(pred)

🤖 Generated with Claude Code