Inedible shelled Torreya grandis bad seeds will be produced during post-ripening treatment and frying, which cannot be accurately recognized and rejected manually without destroying the shells, affecting the overall quality of shelled Torreya grandis seeds. This study used two near-infrared spectrometers to collect spectral data of shelled normal and bad Torreya grandis seeds and eight spectral pre-processing methods was studied and compared. Then, a single wavelength selection method (Uninformative Variables Elimination, Competitive Adaptive Reweighted Sampling, Successive Projections Algorithm, and Subwindow Permutation Analysis) and a joint wavelength selection method were adopted to select characteristic wavelength, and Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) methods were applied to establish the identification model of Torreya grandis bad seeds. Also, the model's performance was compared to determine the better wavelength selection method for different spectrometers. The results show that for spectrometer 1, preprocessing can not improve the model performance effectively. The Successive Projections Algorithm is the optimal wavelength selection method. The sensitivity, specificity, and accuracy of the LDA and SVM models in the prediction set are 97.10%, 95.00%, 96.00% and 97.10%, 97.50%, and 97.30%, respectively, superior to the full-wavelength model. The number of modeled wavelength variables was reduced from 661 to 9, only 1.36% of the original number of wavelength variables. For spectrometer 2, baseline correction is the optimal preprocessing method, and Subwindow Permutation Analysis is the optimal feature wavelength selection method. The sensitivity, specificity, and accuracy of the prediction sets of the developed LDA and SVM models are 100.00%, 92.50%, 96.00% and 100.00%, 95.00%, and 97.30%, which are consistent with full-band model performance. The number of modeled wavelength variables was reduced from 155 to 55, which is 35.48% of the original number of wavelength variables. It can be seen that near-infrared spectroscopy can better identify the shelled bad Torreya grandis seeds, and the appropriate wavelength selection method can effectively screen the characteristic wavelengths, simplify the model, and improve the accuracy and stability of the model. It is also found that the wavelength range of 1 000 similar to 1 300 nm is related to the starch, fat, and protein content of Torreya grandis seeds, making it more suitable for identifying bad Torreya grandis seeds. This study provides a reference for the rapid and nondestructive identification of shelled Torreya grandis bad seeds.
