Functional principal component analysis for near-infrared spectral data: a case study on Tricholoma matsutakeis

Abstract

Abstract Tricholoma matsutakeis (TM) is the most expensive edible fungi in China. Given its price and exclusivity, some dishonest merchants will sell adulterated TM by combining it with cheaper fungi in an attempt to earn more profits. This fraudulent behavior has broken food laws and violated consumer trust. Therefore, there is an urgent need to develop a rapid, accurate, and nondestructive tool to discriminate TM from other edible fungi. In this work, a novel detection algorithm combined with near-infrared spectroscopy (NIR) and functional principal component analysis (FPCA) is proposed. Firstly, the raw NIR data were pretreated by locally weighted scatterplot smoothing (LOWESS) and multiplication scatter correction (MSC). Then, FPCA was used to extract valuable information from the preprocessed NIR data. Then, a classifier was designed by using the least-squares support-vector machine (LS-SVM) to distinguish categories of edible fungi. Furthermore, the one-versus-one (OVO) strategy was included and the binary LS-SVM was extended to a multi-class classifier. The 166 samples of four varieties of fungi were used to validate the proposed method. The results show that the proposed method has great capability in near infrared spectra classification, and the average accurate of FPCA-LSSVM is 97.3% which is greater than that of PCA-LSSVM (93.5%).