Abstract:

Rapid and non-destructive identification of staled and fresh Torreya grandis seeds is crucial for protecting consumers' rights and interests two NIR spectral instruments，visible-near infrared （spectrometer Ⅰ） and short-medium wave NIR （spectrometer Ⅱ）， were employed to acquire spectral data from shelled Torreya grandis samples. Nine preprocessing methods were applied to the spectral data， followed by feature variable selection using four algorithms： interval combination optimization （ICO）， competitive adaptive reweighted sampling （CARS）， successive projections algorithm （SPA）， and variable combination population analysis （VCPA）. Based on optimized variables， three classification models-linear discriminant analysis （LDA）， support vector machine （SVM）， and backpropagation neural network （BP） were established. Results demonstrated that spectral feature optimization significantly enhanced model performance. For visible-NIR spectra， the CARS-SVM model achieved optimal classification， with sensitivity， specificity， and accuracy rates of 100% in the prediction set. For short-medium wave NIR spectra， the BP neural network model incorporating standard normal variate （SNV） preprocessing and VCPA-optimized variables exhibited superior performance， attaining sensitivity， specificity， and accuracy rates of 98.18%， 93.02% and 95.04%， respectively. It is aim to provide a detection method for rapid and non-destructive identification of Torreyagrandis stale seeds， effectively distinguishing their quality.

Key words: Near-infrared spectroscopy, Torreya grandis seeds, Torreya grandis stale seeds, Wavelength variable selection, Discrimination model