A Cross-Attention Spectral–Spatial Fusion Network for Geographical Origin Identification of Goji Berries (Lycium barbarum) Using Hyperspectral Imaging

Goji berries (Lycium barbarum) are widely consumed as functional foods and medicinal–food products, and their quality and authenticity are closely associated with geographical origin. Rapid and reliable origin identification is therefore essential for quality control, authenticity assessment, and traceability. In this study, a hyperspectral imaging (HSI)-based deep learning framework was developed for the non-destructive geographical origin identification of goji berries by integrating spectral and spatial information. Spectral features were extracted using a one-dimensional convolutional neural network equipped with a spectral–channel attention module (SCAM), which adaptively enhances informative wavelengths and discriminative spectral representations. Spatial features were learned via HSI, using DenseNet to characterize complementary morphological and textural information. A cross-attention fusion module was further introduced to establish bidirectional interactions between spectral and spatial features, enabling adaptive integration of complementary information and reducing feature redundancy. The proposed framework achieved a classification accuracy of 98.24%, outperforming conventional machine learning methods and representative deep learning models. These results demonstrate that cross-attention-based spectral–spatial fusion can effectively exploit the complementary characteristics of HSI data, providing a rapid, non-destructive, and reliable strategy for goji berry authentication and geographical traceability, with potential applications in the quality control of functional foods and other agricultural products.