Non-destructive detection of internal mold in Torreya grandis seeds using NIR spectroscopy and deep learning with integrated preprocessing and variable selection

Torreya grandis seeds are valued for their nutritional and culinary benefits, but internal mold can affect seed quality without being visible. This study explores the use of near-infrared spectroscopy for rapid, non-destructive detection of internal mold in these seeds. By analyzing spectral differences between healthy and moldy seeds, various preprocessing techniques were applied to optimize the data. Key wavelengths were selected using both individual and hybrid variable selection methods. Linear discriminant analysis (LDA), convolutional neural networks (CNN), and multilayer perceptron (MLP) models were developed for classification. Shapley additive explanations (SHAP) visualized the impact of key wavelengths, while t-distributed Stochastic Neighbor Embedding (t-SNE) was used for model visualization. The results show that CNN and MLP models outperform the traditional LDA model. The Savitzky-Golay (SG)-Baseline-Single Nucleotide Variant (SNV) preprocessing method is the most effective for CNN and MLP models. Competitive Adaptive Reweighted Sampling (CARS)-Iterative Channel Optimization (ICO) and Univariate Variable Elimination (UVE)-CARS-ICO were optimal for variable selection in CNN and MLP, respectively. The accuracy, sensitivity, and specificity of the CARS-ICO-CNN and UVE-CARS-ICO-MLP models were 97.22%, 96.15%, 97.83% and 97.22%, 92.31%, 100%, respectively. These findings highlight the potential of NIR spectroscopy and deep learning models for quality control in seed processing. ● Developed a method for discriminating internal mold in Torreya grandis seeds. ● Identified the optimal preprocessing method using a combined strategy. ● Achieved high accuracy with an optimized CNN discrimination model. ● Used a combined variable selection strategy to optimize the classification model. ● Applied SHAP and t-SNE to interpret model features and visualize class separation.