Predicting the composition of multiple soybean varieties from whole and ground seeds using Fourier transform...

Soybean is being increasingly included in human diets, highlighting the importance of determining its composition. Although Fourier-Transform Near-Infrared Spectroscopy (FT-NIRS) has become a promising technology, currently used models remain limited to macro-composition determination. This study investigates the feasibility to predict soybean macro-composition, fatty and amino acids, micronutrients, vitamins, phytic acid, isoflavones, trypsin inhibitor, and more than fifty Volatile Organic Compounds (VOCs) using FT-NIRS and machine learning. Our models achieved acceptable-to-highly-accurate quantification of proteins, lipids, polyunsaturated fatty acids, linoleic acid, and up to 15 amino acids, four micronutrients, five vitamins, and 14 VOCs (0.60 < R² < 0.96 and 5.3 % < NRMSE <21 %). In most cases, support vector, elastic net, and partial least square regression revealed the best regressors when combined to spectrum preprocessing filters. Our study thus provides encouraging perspectives for applications in safety and quality control, meat analog improvement, and variety development programs.