Integration of hyperspectral imaging, non-targeted metabolomics and machine learning for vigour prediction of naturally and accelerated aged sweetcorn seeds

Understanding and predicting the storage stability of sweetcorn seeds is critical for effective supply chain management, however, prediction ability relies heavily on accelerated ageing (AA) studies and this is not always directly applicable to natural ageing (NA). In this study, hyperspectral imaging (HSI) and non-targeted metabolomics (LC-MS/MS) were integrated using PLS-R, SVM-R and OPLS-DA to predict loss of seed vigour in NA seeds, using data based on AA seeds. The inconsistencies in the pattern of spectral variation between seeds undergoing AA and NA were first identified. AA-based vigour prediction models were then built using all wavelengths and effective wavelengths (EWs) selected by regression coefficients. These models were externally validated by independent AA and NA seed datasets, respectively. The results yielded satisfactory predictions for AA seeds (R2 ≥ 0.814), but low precision for NA seeds (R2 ≤ 0.696). Metabolome analysis identified 54 differential metabolites, containing a large proportion of amino acids, dipeptides and their derivatives, which were important substances reflecting discrepancies between the ageing mechanisms of AA and NA seeds. Subsequently, N-H bond-related wavebands were deemed to be a possible interference factor in the models' practicability. After removing the N-H bond-related EWs, the AA-based models achieved better performance on NA seeds, with R2v-2 value increasing from 0.696 to 0.720 for Lvsechaoren and from 0.668 to 0.727 for Zhongtian 300. In summary, coupling HSI, LC-MS/MS and machine learning was shown as an appropriate approach for non-destructive monitoring and predicting the vigour of stored sweetcorn seeds.