Powered by FOSS

Non-linear calibration models for NIR spectroscopy

A comprehensive comparative study of different nonlinear calibration techniques for NIR spectroscopic data analysis


By Nirperformance staff


A featured paper by Wangdong Ni, Lars Nørgaard and Morten Mørup recently published in Analytica Chimica Acta, compares the different NIR calibration techniques available for spectroscopic applications showing nonlinear behavior. The study is carried out as an evaluation of three publically available NIR data sets.

The paper presents a comprehensive comparison of nonlinear methods for spectroscopic calibrations using PLS as the linear benchmark. The methods compared are:
Kernel PLS (KPLS), support vector machines (SVM), least-squares SVM (LS-SVM), relevance vector machines (RVM), Gaussian process regression (GPR), artificial neural network (ANN), and Bayesian ANN (BANN).

The methods are compared with respect to performance on the three selected NIR data sets reflecting both non-linear, weak non-linear and linear relations between the spectra and the dependent variable. Different aspects of the various approaches are discussed including computational time, model interpretability, potential over-fitting using the non-linear models on linear problems, robustness to small or medium sample sets, and robustness to pre-processing.

With appropriate preprocessing, both linear and nonlinear calibration techniques can generate similar predictive performances for the datasets with weak nonlinearity. Although none of these calibration techniques are always the best on all datasets, considering different aspects, the study finds GPR and BANN to be the most promising in terms of low prediction error and thus highly recommended for spectroscopic calibrations of small sample sets. This is due to the consistency in predictive performance for cases with and without preprocessing and effective handling of spectroscopic datasets with both strong and weak nonlinearity.

The conclusions of the study suggest that while the LS-SVM is attractive due to its good predictive performance for both linear and nonlinear calibrations, GPR and BANN are the most powerful and promising methods for handling linear as well as nonlinear systems, especially when the data sets are moderately small.

Get the full paper from Analytica Chimica Acta here

Do you have any colleagues or connections who might benefit from this research? Send them an email or share this with your followers and connections on Twitter and LinkedIn.

Leave a Reply

  • (will not be published)