Alligator - Calibration Methods
The following quantitative calibration methods are available in Alligator:
MLR - multiple linear regression
MLR is a well known regression method selecting single wavelengths by correlation based on Gaussian principles. Multiple wavelengths are selected if more than one term is selected in the options. The automatic selection of additional wavelengths is stoped if the maximal number of terms is reached OR the F value of the next wavelength is lower than the confirgured limit.
There are sub-version to MLR with respect to how to add an additional wavelength:
- add additional wavelengths manually (i.e. from the literature)
- select each wavelength by hand (step-up)
- select additional wavelengths fully automatical (forward)
- select additional wavelengths fully automatical but check whether any existing eavelength shall be replace by another before an additional wavelength is added (stepwise)
The Norris regression uses the fraction of the absorption values at two wavelengths, either by hand or automatically.
A full search for the best triplet of wavelengths is available as well. (The best two wavelengths can be found by using stepwise with a limit of two terms.)
The advantage of MLR models being well interpretable models is hampered by its sensitivity to multicollinearity in the spectra. Typically all spectra from scanning instruments (monochromator, Diode Array or FT instruments) are highly multilinear.
MLR models are BLUE - best linear unbiased estimators - for a given dataset if ALL Gaussian prerequsites are fulfilled.
PLS - partical least square regression
PLS compresses the spectra into a PLS score space, which is created by maximizing covarinace between the spectra and the reference values in the data set. PLS is by far the most used algorithm in practical NIRS applications outside science. Following the compression the regression is calculated scores versus reference values.
There are sub-versions to PLS regression:
- nPLS (next PLS) an identical twin of modified PLS by Mark Westerhaus, maximizing correlation between spectra and reference values to create the score space
- PCA regression, maximizing covariance in spectra space to create the score space
- BVS PLS: backward variable selection/elimination PLS
- iPLS: interval PLS
- rPLS: recursive PLS
LOCAL
LOCAL is based on an existing database of spectra with reference values. In LOCAL for any new spectrum to be predicted a subset from the database is selected and an individual PLS model is calculated and the unknown spectrum is predicted. The number of samples as well as key parameters for the PLS modeling are set in the options.
LOCAL can also be calculated in PLS or PCA score space.
SVM - support vector machines
A modern artificial intelligence method will optimize regression especially for non-linear problems.
ANN - Artifical neural networks
As a qualitative method PCA is available.
PCA - principal component analysis
It is implemented that way that no constituent value is predicted, but H values to determine whether a sample be longs to a population or not. Global and neighborhood H values are predicted.