Research on the identification of Ziziphi Spinosae semen and its adulterants based on the fusion of hyperspectral technology and transfer learning.

Ziziphi Spinosae semen is a high-value food ingredient with medicinal properties, but it is increasingly subject to adulteration using cheaper substitutes such as Ziziphi Mauritianae semen and Hovenia acerba seeds. Conventional detection methods are often time-consuming, labor-intensive, and reliant on complex instrumentation, making them unsuitable for rapid screening. To address this challenge, this study proposes a multimodal adulteration detection framework that integrates transfer learning with hyperspectral technology. First, an efficient modeling strategy combining transfer learning fine-tuning and feature extraction is designed based on three mainstream network architectures: ConvNeXt, inception, and EfficientNet. Secondly, preprocessing and feature variable selection methods are applied to process hyperspectral data, identifying key discriminative wavelengths to significantly reduce feature dimensions. This optimization improves model computational efficiency and stability while maintaining or enhancing classification accuracy. Finally, image features and processed hyperspectral data are fused at the multimodal data layer, enabling collaborative learning through the deep complementarity of image morphological features and spectral information. This effectively overcomes the limitations of single-modal perception. The results show that the model based on transfer learning and hyperspectral data fusion outperforms single-modal methods, achieving a classification accuracy of 98 % or higher, and demonstrates excellent stability. This study provides a rapid, efficient, and highly accurate method for the identification of Ziziphi Spinosae semen adulteration and offers technical support for the quality control and intelligent supervision of seed-based food products.