Abstract: In this study, visible-near infrared (Vis-NIR) spectra of Xinjiang plums at four different shelf-life stages (1, 3, 5, and 7 d after harvest) were collected. Four spectral preprocessing methods, Savitzky-Golay (SG) smoothing, standard normal variate transformation (SNV), baseline correction (BC), and min-max normalization (MN), were compared to develop shelf-life classification models using partial least squares-discriminant analysis (PLS-DA) or extreme learning machine (ELM). For the development of PLS-DA and ELM models, feature wavelength extraction methods, namely competitive adaptive reweighted sampling (CARS), variable iterative space shrinkage approach (VISSA), and successive projection algorithm (SPA), were separately combined with three spectral features, namely band ratio (BR), band difference (BD), and normalized spectral intensity difference (NSID), thereby forming combined feature sets. The results showed that the PLS-DA and ELM models based on the full spectrum exhibited limited classification performance with a validation accuracy of only 70.24%. In contrast, the MN-SPA-NSID-ELM model, which integrated MN preprocessing, SPA-based feature extraction, and NSID, achieved a validation accuracy of 97.62% using only 24 selected variables and 30 hidden layer neurons, significantly outperforming the other combined models. In addition, on an independent external test set of Xinjiang plums, this model still achieved an accuracy of 99.18% with a Kappa coefficient of 0.989, indicating improved shelf-life classification efficiency. This study provides a rapid, accurate, and nondestructive technique for the production and grading of Xinjiang plums.

Key words: Xinjiang plums; visible-near infrared spectroscopy; shelf-life; multi-strategy feature fusion; machine learning