ATR-FTIR spectroscopy combined with machine learning for classifying chia (Salvia hispanica L.) seeds by viability and genotype: a tool for ex situ germplasm conservation
Аннотация
This study explores the integration of Attenuated Total Reflectance Fourier Transform Infrared (ATR-FTIR) spectroscopy with supervised machine learning models for the rapid assessment of viability in Salvia hispanica L. (chia) seeds from two genotypes. The approach requires a small sample amount, making it particularly valuable for germplasm banks where accessions are often limited. Spectral data from both high and low frequency mid-infrared regions were used to identify molecular patterns associated with seed aging, including lipid oxidation, protein modifications, and carbohydrate alterations. Three classification algorithms (Classification Tree, Bagging Tree, and Random Forest) were evaluated under different preprocessing strategies: Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), Savitzky-Golay (SG) derivative, and their combinations. During internal model development, Random Forest combined with MSC + SG preprocessing showed the highest classification performance in both spectral regions. Independent blind validation, however, revealed differences in model transferability: the low-frequency model reached 75% accuracy for the four genotype × viability classes and 95% accuracy for viability discrimination, whereas the corresponding values for the high-frequency model were 31% and 55%, respectively. Variable importance analysis using mean decrease accuracy and mean decrease Gini identifies key spectral markers of viability loss, notably lipid-related bands in the high-frequency region (~ 2853 cm^−1), protein-related signals in the amide region, and carbohydrate and tocopherol-associated bands at ~ 1099 cm^−1 in the low-frequency region. These findings support the potential of ATR-FTIR spectroscopy combined with machine learning as a rapid approach for seed viability assessment, while emphasizing the importance of independent validation before its application in germplasm conservation workflows.