Two-Step Partial Least Squares-Discriminant Analysis Modeling for Accurate Classification of Edible Sea Salt Products Using Laser-Induced Breakdown Spectroscopy

Abstract

Laser-induced breakdown spectroscopy (LIBS) has been widely applied to material classification in various fields, and partial least squares-discriminant analysis (PLS-DA) is one of the frequently used classical multivariate statistics to construct classification models based on the LIBS spectra. However, classification accuracy of the PLS-DA model is sensitive to the number of classes and their similarities. Considering this characteristic of PLS-DA, we suggest a two-step PLS-DA modeling approach to improve the classification accuracy. This strategy was demonstrated for a six-class problem in which six commercial edible sea salts produced in Japan, South Korea, and France are classified using their LIBS spectra. At the first step, test spectra were sorted into four classes and one extended class, composed of the two other most confusing classes, and then the test spectra in the extended class were further classified into each of the two constituent classes which were modeled separately from the other four classes. This two-step classification has been found to remarkably improve the PLS-DA classification accuracy by maximizing the difference between the confusing classes in the second-step modeling.