Old advice on multivariate calibration: still in force, but not always followed. A tutorial

A growing number of multivariate calibration protocols use deep learning for estimating target properties from spectra, without considering simpler alternatives such as partial least-squares (PLS) regression. In addition, rational analyses are lacking on the following issues: (1) the real need of mathematically pre-processing the spectra before model building, (2) the proper statistical comparison of root mean square prediction errors for different models, and (3) the impact of the uncertainty associated to the primary reference technique used to train the models.

To illustrate the above general concepts, a specific experimental example is analyzed: a publicly available data set regarding the estimation of dry matter content in fruits from UV-visible-near infrared spectra. Literature results using various shallow and deep learning methods, and those presently achieved by global and local PLS models are compared, concluding that the latter may be preferred due to its simplicity, accuracy, and prediction errors which are close to the uncertainty of the reference technique. The implications regarding general multivariate calibration protocols are highlighted in view of these results.

Multivariate calibration reports will benefit from the application of parsimonious concepts, producing more robust and reliable procedures, with prediction errors that are in reasonable agreement with those associated to the reference values in the calibration target variables.