Hyperspectral Imaging and Machine Learning for Assessing the Germination of Hydroprimed Lettuce Seeds

Hyperspectral imaging (HSI) provides high-resolution spectral information across broad wavelength ranges and has emerged as a promising non-destructive approach for assessing seed quality. This study validated the use of HSI combined with machine learning to identify spectral features associated with germination performance in hydroprimed lettuce seeds. Seeds of the Roxa and Vanda genotypes were subjected to hydropriming, dried, and stored prior to evaluation. Hyperspectral images were acquired using 300 spectral bands spanning 384.7–1020.7 nm for each treatment, and mean spectra were used as predictors in classification models. Machine learning analysis was performed primarily with Random Forest to discriminate seed performance categories, including normal seedlings, abnormal seedlings, non-germinated seeds, and dead seeds. Feature-importance results indicated that wavelengths near 390 nm were among the most informative predictors of germination performance, whereas bands at 473–475 nm contributed comparatively less. The Random Forest models showed high sensitivity and specificity across the defined classes, supporting the feasibility of spectral-based categorization of seed outcomes. In complementary analyses, Ridge regression reinforced the relevance of peripheral regions of the spectrum, highlighting 384.73 nm and 874.57 nm as important for classification and underscoring the value of targeted band selection to improve model interpretability and efficiency. Hydropriming affected both spectral signatures and physiological responses, with genotype- and lot-dependent outcomes. In Roxa, selected seed lots exhibited superior performance in germination and vigor tests after hydropriming and storage. In Vanda, specific lots displayed greater uniformity and improved emergence indices. Overall, the integration of HSI and machine learning captured strong associations between spectral data and seed performance metrics, with high canonical correlations (R² = 0.91–0.99). These results demonstrate that HSI coupled with predictive modeling is an effective, rapid, and non-destructive strategy to assess the physiological quality of hydroprimed lettuce seeds and to support decision-making in seed lot classification.