VIS–NIR–SWIR Hyperspectral Imaging and Advanced Machine and Deep Learning Algorithms for a Controlled Benchmark of Bean Seed Identification and Classification

Renan Falcioni, Nicole Ghinzelli Vedana, Caio Almeida de Oliveira, João Vitor Ferreira Gonçalves, Marcelo Luiz Chicati, José Alexandre M. Demattê, Marcos Rafael Nanni

Reliable seed accession identification underpins germplasm conservation, traceability and breeding; however, conventional assays remain destructive, labour-intensive and difficult to scale. Here, visible–near-infrared–shortwave infrared (VIS–NIR–SWIR) hyperspectral imaging (HSI; 449.54–2399.17 nm; 563 bands) was used to classify 32 grain–legume accessions (n = 3200 seeds; 100 seeds per accession), comprising 30 common bean (Phaseolus vulgaris L.) landraces plus two outgroup legumes (Vigna angularis (Willd.) Ohwi & Ohashi and Cajanus cajan (L.) Huth). Each seed was represented by one ROI-averaged spectrum obtained from mean representative pixels within a standardised 10 × 10 pixel window at the centre of each seed. A fixed stratified 70:30 seed-level training:test partition was used, with 70 seeds per accession (n = 2240) reserved for fully independent training and 30 seeds per accession (n = 960) reserved as a fully independent test set. Principal component analysis (PCA) captured 97.42% of the spectral variance in the first three components (PC1 = 63.34%, PC2 = 23.78%, and PC3 = 10.31%). One-versus-rest wavelength association mapping revealed a maximum R2 of 0.775 at 461.37 nm, and ReliefF concentrated the strongest reduced-band signal within 449.54–456.30 nm and 577.02–597.54 nm. In the original ReliefF-selected 16-band benchmark, the subspace discriminant reached 68.25% macro-F1 and 68.54% balanced accuracy; after edge-band trimming, the alternative 16-band configuration decreased to 60.67% and 60.94%, respectively. With respect to the full-spectrum sensitivity benchmark, linear discriminant analysis achieved 96.35% balanced accuracy, followed by linear SVM (94.17%). Deep learning trained directly on the full 563-band spectra reached 84.90% test accuracy, 84.47% macro-F1, 86.27% precision and 84.90% recall, with MLP_Wide outperforming the convolutional, recurrent and attention-based alternatives. Overall, under controlled laboratory conditions, this benchmark shows that accession discrimination is driven mainly by visible-domain contrasts in the most compact representations, whereas the full spectral context remains important for the most confusable accessions and for cautious future sensor design. The reduced-band findings should therefore be interpreted as exploratory guidance for sensor design rather than as a validated deployment-ready specification.