PLS, ANN or LOCAL Calibration?
FOSS NIR solutions are capable of handling calibrations based on different mathematical principles, each offering unique advantages. PLS calibration is a well established method that uses a linear combination of the wavelengths for calibration development, rather than one single wavelength in ordinary linear regression.
PLS is the obvious choice when the number of samples is limited, for systems where the parameter of interest has a linear relationship to the spectra and where the range of the parameter is restricted. Some natural sample types vary a lot, both in origin and composition, leading to highly on-linear relationships between the parameter of interest and the spectra. In such cases, ANN calibrations are available and offer a good solution because the core of the ANN is based on non-linear mathematics.
ANN can be used to make broad calibration models for a wide selection of products without sacrificing accuracy. The ANN technique requires a large number of samples to work properly, typically more than 1000. This fact explains why ANN calibrations are highly robust and transferable, and consequently are very cost effective to support over time. On the other side, this requirement precludes the use of ANN for smaller datasets. PLS and ANN are static models that remain unchanged until a calibration update is made.
In contrast, LOCAL calibrations are dynamic. With LOCAL calibrations the spectrum of the unknown sample is compared with spectra in a huge database, and the prediction is based on those of the spectra that most closely resemble the unknown. The unique advantage with LOCAL calibrations is that updates are straightforward. New samples are simply added to the database. However, being dynamic also makes the model sensitive to how well the database represents the new samples measured.