Precise prediction of amino acid content and geographical origin tracing of lotus seeds based on multimodal data...

This study established a three-tiered framework of feature characterization, information coupling and modeling mapping to correlate multi-source coupled information with the geographical origin and quality of lotus seeds. Near-infrared spectroscopy (NIRS), gray level co-occurrence matrix (GLCM) texture features and amino acid profiles were acquired, and Bayesian optimization was adopted. Dual-modal (amino acid + NIRS; amino acid + GLCM) and tri-modal (amino acid + NIRS + GLCM) models were constructed. These multimodal combinations outperformed single-modal counterparts in most cases instead of universally, and Bagged Trees achieved the optimal overall performance balance. Key NIRS-GLCM features for amino acid prediction were screened via feature-target association visualization. For L-carnosine, γ-aminobutyric acid and phosphorylethanolamine, the amino acid-free NIRS-GLCM non-destructive model yielded R²p ≥ 0.92209. A MATLAB-based GUI integrating preprocessing, multi-model prediction and result analysis was built as a conceptual prototype to illustrate the workflow for quality assessment of lotus seeds and other foods.