Comparative investigation of partial least squares discriminant analysis and support vector machines for geological cuttings identification using laser-induced breakdown spectroscopy

With the hope of applying laser-induced breakdown spectroscopy (LIBS) to the geological logging field, a series of cutting samples were classified using LIBS coupled with chemometric methods. In this paper, we focused on a comparative investigation of the linear PLS-DA method and non-linear SVM method. Both the optimal PLS-DA model and SVM model were built by the leave-one-out cross-validation (LOOCV) approach with the calibration LIBS spectra, and then tested by validation spectra. We show that the performance of SVM is significantly better than PLS-DA because of its ability to address the non-linear relationships in LIBS spectra, with a correct classification rate of 91.67% instead of 68.34%, and an unclassification rate of 3.33% instead of 28.33%. To further improve the classification accuracy, we then designed a new classification approach by the joint analysis of PLS-DA and SVM models. With this method, 95% of the validation spectra are correctly classified and no unclassified spectra are observed. This work demonstrated that the coupling of LIBS with the non-linear SVM method has great potential to be used for on-line classification of geological cutting samples, and the combination of PLS-DA and SVM enables the cuttings identification with an excellent performance. (C) 2014 Elsevier B.V. All rights reserved.