Automatic interpretable identification of Lycium barbarum origin via spectro-spatial feature fusion of hyperspectral images

The compositional profile and quality of goji berries ( Lycium barbarum ) are strongly influenced by the ecological environment of their production regions. Accurate geographical origin identification is critical for quality assurance and brand integrity. To improve classification accuracy and model interpretability, a novel spectro-spatial fusion network (LSSFN) based on hyperspectral imaging is proposed. The architecture comprises dual branches for extracting spectral and spatial features, which are subsequently integrated through an adaptive fusion module. Comprehensive evaluations were conducted to assess the effects of convolutional kernel widths, spectral preprocessing methods, and image construction schemes on model performance. In addition, SHapley Additive exPlanations (SHAP) were applied to quantify feature importance and improve model transparency. Experimental results on the independent test set show that LSSFN achieves a classification accuracy of 94.47 %, outperforming the spectral-only and spatial-only models by 0.75 % and 10.02 %, respectively. The most discriminative features identified by SHAP were 623.7 nm, 991.4 nm, and the third principal component (PC3). This study offers an effective, non-invasive, and interpretable solution for origin traceability of goji berries and presents a generalizable framework for quality control in functional foods and specialty crops, while also advancing explainable deep fusion approaches in hyperspectral analytics. • Developed a rapid origin identification method for goji berries using hyperspectral imaging. • Proposed a novel spectral–spatial fusion network for deep feature extraction. • Employed SHAP analysis to reveal feature contributions and enhance interpretability. • Provided a new perspective for geographical traceability of agricultural products.