Train pre-trained PLSR models from an OSSL library

Iterates over properties and fits one PLSR model per target against the OSSL spectra in ossl_library$Xr, with internal cross-validation to pick the optimal number of components per property. The returned list is a drop-in replacement for the ossl_models argument of predict_ossl_pretrained and fill_from_spectra.

Usage

train_pls_from_ossl(
  ossl_library,
  properties = c("clay_pct", "sand_pct", "silt_pct", "cec_cmol", "ph_h2o", "oc_pct"),
  ncomp_max = 20L,
  validation = c("CV", "LOO", "none"),
  segments = 10L,
  preprocess = "snv+sg1",
  min_n = 50L,
  verbose = TRUE
)

Arguments

ossl_library

A list with two named elements: Xr (numeric matrix of training spectra) and Yr (data.frame keyed by property name, one row per training spectrum). See ossl_library_template.

properties

Character vector of column names in ossl_library$Yr to train models for. Defaults to the six core soil properties exposed by OSSL.

ncomp_max

Integer. Upper bound on the number of PLS components to consider during cross-validation. Defaults to 20.

validation

One of "CV" (default, k-fold), "LOO" (leave-one-out, slow), "none" (uses ncomp_max components without selection).

segments

Number of CV segments when validation = "CV". Default 10.

preprocess

Pre-processing label passed to preprocess_spectra. Stored on the trained models so predict_from_spectra can reapply it.

min_n

Minimum number of valid training samples (after dropping rows with non-finite y or X). Properties below this threshold are skipped with a warning. Default 50.

verbose

If TRUE (default), prints a per-property summary on completion.

Value

A named list of soilKey_pls_model objects, one per successfully trained property. Carries trained_at, soilKey_version and preprocess attributes for provenance.

Details

Spectra are pre-processed inside the function (default "snv+sg1"); the same preprocessing is used downstream by predict_from_spectra so the user does not have to remember which transform was applied at training time.

Examples

if (FALSE) { # \dontrun{
lib <- download_ossl_subset(region = "south_america")
models <- train_pls_from_ossl(lib,
                               properties = c("clay_pct", "ph_h2o"))
result <- predict_from_spectra(my_pedon, models = models)
} # }