Evaluation of relevance vector machine as a chemometric method for NIR spectroscopic quantitative analysis and classification

Recently, a relevance vector machine (RVM), a machine learning technique employing Bayesian inference to obtain parsimonious solutions for probabilistic classification and regression, has drawn an attention in the field of NIR spectroscopy. It is basically similar to support vector machine (SVM), but it is based on the Bayesian formulation with an appropriate prior. In comparison with SVM, RVM could be potentially valuable since it avoids several drawbacks of SVM, such as kernel function limitation and parameter tuning complexity, as reported. Therefore, it is worthwhile to evaluate RVM as an alternative multivariate calibration method for analysis of NIR spectroscopic data. First, it was tested for discriminating geographical origin of diverse agricultural samples using NIR spectroscopy. For this purpose, NIR datasets of 10 different agricultural samples (such as sesame and red pepper) were obtained. All of the NIR spectra were collected with a Foss NIRSystems Model 6500 spectrometer equipped with a quartz halogen lamp and PbS detector. All of the samples were ground into powders (20-mesh) for the collection of diffuse reflectance spectra. For the purpose of comparison, PLS-DA and SVM were also performed using the same datasets and the resulting discrimination accuracies were compared with those acquired using RVM. Second, RVM was used to determine the concentrations of major inorganic acids (H3PO4, H2SO4, HNO3 and HCl) in etchant solutions using their transmission NIR spectra. NIR spectra were collected using an ABB Bomem FT-NIR spectrometer equipped with a tungsten-halogen source and a germanium detector. The measurement was largely based on the recognition of the varied spectral feature of water absorption bands by their presence in the samples, rather than direct NIR absorption by themselves. As a result, the resulting spectral variations according the concentration change were relatively indistinct and unforeseen spectral features under the water bands were necessary to be recognized for multivariate quantitative analysis. The resulting accuracies in the determination of component concentrations using RVM were compared with those using PLS. In final, the advantages and challenges of RVM for analysis of NIR spectroscopic data are profoundly discussed.