Visible-Near-Infrared Hyperspectral Imaging Enables Nondestructive Identification of Bean Accessions via 1D Spectral...

Reliable seed accession identification via sensors underpins germplasm conservation, traceability, and breeding, yet conventional assays are often destructive, labor-intensive, and difficult to scale. Here, visible-near-infrared (VNIR) hyperspectral imaging (Headwall Photonics, 825 bands) was used to obtain one ROI-averaged 1D reflectance spectrum per seed and to classify 32 grain-legume accessions (N = 3200 seeds; 100 seeds per accession), comprising 30 common bean (Phaseolus vulgaris L.) landraces and two locally co-classified outgroup legumes (Vigna angularis (Willd.) Ohwi and H. Ohashi and Cajanus cajan (L.) Huth). The seed-level reflectance signatures exhibited a strong accession structure (PERMANOVA; F = 201.40, p < 0.001). Principal component analysis (PCA) captured 94.48% of the total spectral variance in the first three components (PC1 = 67.17%, PC2 = 21.80%, PC3 = 5.51%), indicating that discrimination is dominated by smooth, wavelength-contiguous modes consistent with color-related absorption and scattering-mediated optics. Wavelength-resolved one-versus-rest association identified accession-specific informatdive bands, with the strongest single association reaching r ² = 0.736 at 424.79 nm, whereas supervised ReliefF ranking concentrated multiclass informativeness in a narrow green window (562.85-584.65 nm; peak ≈ 581.74 nm), providing direct guidance for reduced-band multispectral screening; however, reduced-band tests indicate that this window alone is insufficient for the full 32-class task. In full-spectrum multiclass modeling, classical learners performed best, with a linear support vector machine (Linear SVM) achieving 88.59% accuracy on the independent 20% test split (weighted F1-score = 88.52%) and exhibiting structured confusions among spectrally adjacent accessions. One-dimensional deep learning models were evaluated only as a secondary benchmark; the best network (MLP_1D) reached 82.03% test accuracy (weighted F1-score = 81.52%) but did not exceed the strongest classical model under this mean spectrum, moderate-data regime. The green window is interpreted as an optical leverage region linked to seed coat color and pattern and surface microstructure, although no targeted chemical assays were performed in this study. Because only one seed lot per accession was analyzed under controlled laboratory conditions, robustness to storage conditions, seed moisture content, and harvest time remains to be tested. Overall, VNIR hyperspectral imaging coupled with interpretable 1D spectral reflectance analysis provides a practical, mechanistically plausible route to accession-level seed identification and can inform reduced-band sensor designs for scalable screening in food and agriculture.