Evaluation of Gualouzi quality using hyperspectral imaging technology and explainable artificial intelligence

The adulteration of aged and fresh Gualouzi (GLZ) poses risks to product quality and consumer health. This study aimed to develop a rapid, nondestructive, and high-throughput method for evaluating GLZ quality by integrating hyperspectral imaging (HSI) with explainable artificial intelligence (XAI). First, spectral data were collected from two sets of samples, Wanlou No. 20 (GLZ-20) and Wanlou No. 9 (GLZ-9), stored for varying durations. The preprocessed full-wavelength spectral data were used to construct classification models. Shapley additive explanations (SHAP) was then applied to the models for global interpretability analysis to select feature wavelengths. The results demonstrated that SHAP-based feature selection outperformed conventional methods. Models using the top 16 and 23 SHAP-selected wavelengths achieved prediction accuracies of 97.2% and 98.3% for GLZ-20 and GLZ-9, respectively, in the prediction sets. In validation set I, the accuracies reached 97.7% and 98.9%. Both varieties showed prediction errors below 3.3% in validation set II. Furthermore, SHAP-based model interpretation revealed the intrinsic relationship between feature wavelengths and sample characteristics, thereby enhancing the reliability of discrimination results. This study confirms the feasibility of integrating HSI and XAI for GLZ quality evaluation and provides a reference framework for assessing seed-type nuts. • Integrating HSI with XAI to differentiate aged and fresh Gualouzi • Models exhibited <3.3% prediction errors in adulteration detection • Applying SHAP to screen feature wavelengths and interpret models • Achieving quality assessment based on single-seed single-sided analysis • Providing a reference for quality assessment of seed-type nuts