Artificial intelligence-assisted hyperspectral image analysis for geographical traceability and terpene content prediction in wenyujin rhizoma concisum

Wenyujin Rhizoma Concisum (PJH), the dried rhizome of Curcuma wenyujin Y.H. Chen et C. Ling, is a high-value protected geographical indication (PGI) industrial-medicinal crop rich in terpenes. The PJH market is plagued by inferior substitutes from non-PGI regions, while traditional detection methods are incompetent for industrial-scale quality control. This study aimed to construct a customized hyperspectral imaging -artificial intelligence platform for PJH’s rapid geographical traceability and quality assessment. A customized deep learning model (DeepGTNet) integrating 1D convolution, adaptive structure and lightweight modular design was proposed, specifically designed for hyperspectral data, achieved 99.29% recognition accuracy and a Kappa coefficient of 99.11% for geographical traceability, significantly outperforming conventional machine learning methods. Robust random forest and partial least squares regression models were established through optimized wavelength selection and feature fusion, enabling non-destructive prediction of six key terpenes with high performance (R² > 0.8, RPD ≥ 2.0). Additionally, the structure-spectrum relationship between terpene functional groups and HSI bands was clarified. This research provides a reliable and comprehensive solution for origin traceability and quality evaluation of PJH, and promotes intelligent sustainable development of the industrial crop processing industry. • Hyperspectral image database of Pianjianghuang from different geographical regions. • Terpenes’ functional groups correlate with characteristic spectral bands. • DeepGTNet model enables geographical authentication from hyperspectral data. • Terpene content can be quantitatively predicted from intact samples.