Rapid classification of plastics by laser-induced breakdown spectroscopy (LIBS) coupled with partial least squares discrimination analysis based on variable importance (VI-PLS-DA)

TLDR: In this article, an extension of Partial Least Squares Discrimination Analysis (PLS-DA) that uses variable importance to select input variables was presented, which has the highest classification accuracy and shortest classification time.

Abstract: With the extensive use of plastic products, the recycling and reuse of plastics raise more concerns. Laser-induced breakdown spectroscopy (LIBS) and chemometric methods have been applied to classify plastics. However, the methods are prone to fall into over-fitting when predicting unknown samples. Variable importance is the impact of input variables to classification results. Selecting input variables by variable importance can be used to avoid over-fitting, which has been used for improving model performance based on random forest (RF). However, the progress of optimizing the parameters of RF model is complex. Partial least squares discrimination analysis (PLS-DA), most widely used in spectral data, is a simple and stable method in multivariate analysis. To avoid over-fitting phenomenon and acquire stable results, this paper presents an extension of PLS-DA that uses variable importance to select input variables, namely VI-PLS-DA. In order to validate the classification ability of VI-PLS-DA for plastics, VI-PLS-DA was compared with PLS-DA, RF, and VI-RF. VI-PLS-DA has the highest classification accuracy (99.55%) and shortest classification time (0.096 ms), which indicated a good classification performance for plastics analysis.