Classification of Lupinus seeds into sweet and bitter categories using VIS-NIR spectroscopy and machine learning

The Lupinus germplasm includes sweet and bitter materials distinguished by compounds responsible for bitterness. Conventional identification is often destructive. This study assesses a non-destructive approach based on visible-near infrared (VIS-NIR) spectroscopy and machine learning to classify whole seeds from seven Lupinus species into sweet or bitter classes.

Five machine-learning algorithms were evaluated on two datasets (reflectance and absorbance) acquired with VIS-NIR spectroscopy. Analyses were conducted on raw spectra and on spectra transformed using four spectral-transformation techniques. Because classes were imbalanced, five resampling methods were compared to improve classification performance.

Performance was assessed using F1-score and ROC-AUC. On reflectance, LGR and SVC reached 92.5 and 92.0%; on absorbance, SVC and RF achieved 93.2 and 92.5%. Hybrid transformations consistently improved discrimination, and resampling reduced overfitting associated with class imbalance.

The results indicate that combining VIS-NIR spectroscopy with machine learning provides a suitable non-destructive alternative to discriminate sweet and bitter Lupinus materials/ecotypes.