Simultaneous variable selection for the classification of near infrared spectra

Abstract

This paper addresses the problem of wood wastes recycling automation. We propose variable selection methods based on near infrared spectroscopic data to select a set of wavebands that captures the main spectral peaks of wood materials to improve the sorting performances. The spectra are first jointly modeled as linear combinations of explanatory variables drawn from a collection of Gaussian-shaped functions. The aim is to select a common subset of wavebands shared by several spectra. The variable selection is then formulated as an unconstrained simultaneous sparse approximation problem in which the coefficients related to different spectra are encouraged to be piecewise constant, i.e. the coefficients associated to successive spectra should have comparable magnitudes. We also investigate the case where the coefficients are constrained to be nonnegative. These problems are solved using the fast iterative shrinkage-thresholding algorithm. The proposed approaches are illustrated on a dataset of 290 spectra of wood wastes; each spectrum is composed of 1647 wavelengths. We show that the selected variables lead to better classification performances as compared to standard approaches.