Intelligent geographical origin traceability of Pu-erh tea based on multispectral feature fusion

To achieve accurate origin traceability of Pu-erh tea, this study proposes a deep learning method based on multispectral fusion. By collecting Raman and near-infrared spectral data from five major origins, an improved ECA-ResNet network structure was designed, incorporating an optimized channel attention mechanism for adaptive feature extraction and fusion. While maintaining computational efficiency, the network effectively utilizes Raman spectroscopy's sensitivity to molecular skeleton vibrations and near-infrared spectroscopy's responsiveness to functional group characteristics. Experimental results demonstrate that the feature-fused model achieves a classification accuracy of 95.05 %, surpassing single-spectral and traditional deep learning methods. Mechanism analysis reveals that the model developed functionally differentiated feature channels, achieving organic fusion of multispectral information and exhibiting robust performance in identifying latitude differences among tea origins. This study provides a reliable technical solution to traceability of Pu-erh tea origin, contributing to the standardized development of the tea industry. Future advances in correlating spectral features with specific chemical markers will further enhance the interpretability and mechanistic understanding of multispectral analytical approaches in food authentication.