An effective multi-source information fusion method for electronic nose and hyperspectral to identify the spring tea quality at different harvesting periods

Multimodal information fusion technology enhances the reliability and stability of representing the comprehensive characteristics of sample. In this work, a Global and Local Feature Fusion Network (GLFNet) is proposed to adaptively fuse information from the electronic nose (e-nose) and hyperspectral systems to identify the spring tea quality at different harvesting periods. The fusion unit dynamically integrates the global and local features of gas and spectral information, while the reinforcing unit strengthens the interconnections among these fused features. The e-nose captures macroscopic quality characteristics of spring tea through gas information, whereas hyperspectral analysis reveals microscopic quality features by exploiting the distinct light absorption properties of various chemical functional groups. By integrating these two modalities, a more comprehensive characterization of tea quality is achieved. GLFNet outperforms other state-of-the-art information fusion methods, including both machine learning and deep learning models, achieving the highest classification performance with an accuracy of 99.58%, a precision of 98.65%, a recall of 98.08%, and an F1-score of 97.95%. By fusing e-nose and hyperspectral non-destructive testing technologies, GLFNet provides an effective method for identifying spring tea quality at different harvesting periods.