Unsupervised Classification of Essential Information in Hyperspectral Imaging for Intelligent Identification of Trace Adulterants in Food Matrix

Abstract

In the contribution, a new and intelligent strategy was developed to identify trace adulterants in food matrix. The strategy was based on the hierarchical agglomerative clustering (HAC) analysis of essential information (EI) selected by interesting features finder (IFF) as well as uniform manifold approximation and projection (UMAP) from hyperspectral imaging (HSI) data (named IFF-UMAP-HAC). Four Raman HSI datasets and four NIR HSI datasets were utilized to verify the accuracy and reliability of the new strategy and a comparison between it and UMAP-HAC as well as t-distributted stochastic neighborhood embedding (t-SNE) without or with IFF and HAC was made. When the adulterant with high levels was presented in food matrix (i.e. ≥0.1% mass melamine in milk powder), all of methods provided accurate identification and well separation of adulterants from HSI data. However, only IFF-UMAP-HAC provided satisfactory results for analysis of trace adulterants in food matrix such as samples adulterated with 0.014% melamine in milk powder, ≤0.05% melamine in wheat gluten and trace ternary adulterants in milk powder. Moreover, the new strategy did require no prior knowledge of HSI data structures and class information. Overall, the synergy of HSI with IFF-UMAP-HAC was feasible and had the advantages of sample preparation-free, nondestructive, real-time and in situ, paving the way to quickly detect trace adulterants in food matrix.

Keywords: Essential information, IFF, HSI, UMAP, t-SNE