Rapid anatomical classification and lead contamination analysis in edible legumes using novel LIBS–deep learning frameworks

Abstract

This study presents a novel analytical approach combining Laser-Induced Breakdown Spectroscopy (LIBS) with two advanced deep learning frameworks, DLIBS-FFNet and PLSNetL, for anatomical classification and heavy metal quantification in edible legumes. The elemental composition of six bean varieties was analyzed across three anatomical components (coat, hilum, and cotyledon), revealing consistent mineral profiles rich in essential nutrients such as Ca, K, and Mg. The DLIBS-FFNet model, which integrates Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Autoencoders (AE) for feature fusion, achieved high classification accuracy of up to 96.12 % for non-contaminated beans and 99.83 % for Pb-contaminated samples. Concurrently, PLSNetL, a Partial Least Squares regression-based neural network with dynamic peak selection and adaptive feature extraction, accurately predicted lead (Pb) concentrations across the anatomical components, with R² values of 0.9924, 0.9022, and 0.8462. The combined use of LIBS with these frameworks offers a rapid, non-destructive, and robust method for compositional profiling and contaminant analysis in legumes, contributing valuable insights to food safety assessment and food composition research.