Hyperspectral System Coupled With a Global Spectral Feature Classification Network to Identify the Origin of Mung Bean

Abstract

Addressing the challenge of geographical origin fraud in mung beans is pivotal, as accurate traceability is fundamental to preserving their quality-specific attributes, protecting branded origins, and ensuring supply chain integrity. In this study, a proposed Global Spectral Feature Classification Network (GSFC-Net) coupled with a hyperspectral system is proposed to identify the quality differences of mung beans from different origins. Using the hyperspectral system, spectral information of mung beans from six different origins is collected and preprocessed. We propose the Global Spectral Feature Calculation Module (GSFCM), which integrates convolution, self-attention, and residual connection to extract full-band deep features that effectively represent the original spectral information. The GSFC-Net is designed to self-adaptively establish a nonlinear mapping between the original spectral information and the origin labels. The rationality of GSFC-NET is verified through structural optimization and ablation experiments. Compared with the state-of-the-art spectral information classification methods, GSFC-Net achieves outstanding classification performance and stability, with an accuracy of 98.10%, precision of 98.09%, recall of 98.55%, and an F1-score of 98.32%. In conclusion, this study presents an effective analytical method for identifying mung bean origin, providing a technical method to ensure product authenticity and support fair trade in the market.