MLSP 0.2.0
New features
ml_f() now compares every combination of preprocessing
and calibration method in one call. By default it runs all 26 distinct
preprocessing methods (raw reflectance, standard normal variate,
absorbance, and Savitzky-Golay smoothing, first and second derivatives)
with PCR, PLSR, LASSO, ‘Cubist’, and random forest, and returns the top
10 combinations. New arguments models, top_n,
rank_by, and verbose;
smoother_selection accepts "all", method
names, or a function.- 26 preprocessing methods are available and compared by default: raw
reflectance, SNV, log10(1/R), and 23 Savitzky-Golay filters (linear and
quadratic smoothers, quadratic and cubic first derivatives, and
quadratic second derivatives with 3- to 11-point windows).
- The 14 Savitzky-Golay settings of version 0.1.0 that reproduce
another method exactly (cubic smoothers, linear first derivatives, cubic
second derivatives, and the 3-point quadratic smoother) were removed;
requesting one of them names the identical method to use instead.
- New
preprocess_methods() lists the methods and explains
how window size, polynomial order, and derivative order work. The
methods and naming follow Clingensmith et al. (2019),
doi:10.1111/ejss.12753.
- New
spectral_preprocess() applies any method by
name.
merge_of_lab_and_spectrum() accepts any number of soil
property columns and gains soil_id, spec_id,
n_splits, and cal_prop arguments.type_of_soil accepts a column name as well as an
index.
Bug fixes
- Savitzky-Golay preprocessing failed with current versions of
‘gsignal’ (named spectra are no longer passed to
gsignal::sgolayfilt()).
- Quadratic smoothing with 9- and 11-point windows returned the wrong
object.
- PCR and PLSR predictions used all components at once, giving invalid
metrics; the number of components is now chosen by cross-validation
(one-standard-error rule) and reported in the
LV
column.
- Calibration metrics are now cross-validated for every model
(out-of-bag for random forest); the
cal-mse and
cal-rpiq columns are renamed cv-mse and
cv-rpiq.
- Random forest results were mislabelled by number of trees;
rf_ntree (default 500) now sets the number of trees.
msd.comp() no longer warns when predictions are
constant.
Other changes
- Maintainer changed to Katsutoshi Mizuta.
- All examples run on simulated spectra.
- Added a ‘testthat’ test suite covering every preprocessing method
and the bug fixes above.