Calibration in non-linear near infrared reflectance spectroscopy: a comparison of several methods Original Research Article

Principal component regression (PCR) and partial least-squares regression (PLSR) are the two calibration procedures most frequently used in quantitative applications of near infrared diffuse reflectance spectroscopy (NIRRS). Some systems, however, exhibit a non-linear relationship that neither methodology can model. Frequently, the main culprit of such non-linearity is the multiplicative effect arising from non-uniform particle sizes or diameters in the samples. In this work, we tested various approaches to minimizing the non-linearity resulting from the multiplicative effect of differences in particle size or sample thickness, using the determination of linear density in acrylic fibres as physical model. The approaches tested involve the prior linearizing of data by logarithmic conversion and/or the use of non-linear calibration systems; in this context, the results of applying stepwise polynomial PCR (SWP-PCR) and PLSR (SWP-PLSR), and those provided by a neural network based on the scores of the PCR model (PC-ANN), were compared. The PC-ANN approach was found to provide the best results with linear density data. On the other hand, the SWP-PLSR approach performed on par with the previous one when the variable was linearized by conversion of its values into decimal logarithms.