A comparative ML approach to classify Lupinus species using VIS-NIR spectral data from whole seeds and various data transformation techniques and resampling methods.

The increasing interest in the cultivation and utilization of Lupinus species is driven by their nutritional value and potential for sustainable agriculture. However, the non-destructive taxonomic classification of Lupinus species from whole seeds spectral data remains scarcely explored. In this study, five machine learning classifiers were comparatively evaluated for discriminating seven Lupinus species using visible and near-infrared (VIS-NIR) spectral data, in both reflectance and absorbance modes, acquired from seeds of the official active collection of the CICYTEX Germplasm Bank. The dataset was characterized by marked class imbalance. Model performance was assessed using raw spectra and four preprocessing strategies (including three hybrid combinations) combined with six resampling methods and a no-resampling baseline. Two validation approaches were applied: an 80/20 train-test split and stratified 5-fold cross-validation. Across both spectral domains and validation schemes, Random Forest and Support Vector Classification achieved the best overall performance, with F1 scores above 94%, and AUC, accuracy and precision values above 97%. Logistic Regression also showed substantial improvements when hybrid preprocessing techniques were applied. Cross-validation confirmed the robustness and generalization ability of the best-performing models. These findings support the use of VIS-NIR spectroscopy combined with machine learning as a rapid, objective, and non-destructive framework for Lupinus species discrimination in imbalanced germplasm datasets. This approach may assist taxonomic identification, germplasm management, and the preliminary screening of promising materials for breeding programs, although further validation under independent environmental and acquisition conditions is still required.