Dual-stream feature-decoupling network for rapid and robust rice traceability using handheld Raman spectroscopy

Rice classification and traceability are critical for ensuring food safety, preventing origin fraud, and enhancing agricultural product quality management. Low-cost handheld Raman spectroscopy offers a non-destructive and rapid method for on-site analysis. However, practical application is severely limited because the raw spectral data obtained by these portable instruments often suffer from severe fluorescence interference, random noise, and poor spectral utilization, which collectively lead to reduced analytical accuracy and robustness in real-world settings.

To address these challenges, we propose the Raman-Fluorescence Dual-stream Network (RFDNet) for robust, portable Raman-based rice origin identification. The framework first employs a wavelet decomposition combined with an autoencoder to effectively separate and refine the Raman signal from the background fluorescence and noise. This is followed by a dual-branch feature extraction: a Raman branch designed to enhance fine-grained peak gradients and a fluorescence branch built to model the low-frequency background morphology. We integrate multi-head attention and morphological convolution modules to adaptively capture both local and global spectral dependencies. Experimental results on rice origin datasets demonstrate that the RFDNet achieves a high accuracy of 91%, which represents an improvement of 2.25% compared to the established baseline models.

This dual-stream approach provides a robust and generalizable solution for non-destructive, rapid on-site rice traceability. By decoupling and utilizing complementary information from the Raman signal and fluorescence background, RFDNet effectively overcomes the limitations of signal interference and spectral under-utilization in portable Raman instruments. This highlights its significant potential for reliable authentication and quality management across a broader range of agricultural products.