Geographical origin discrimination of Ophiopogonis Radix using hyperspectral imaging with multi-scale 3D convolution and transformer

Abstract

Ophiopogonis Radix (OPR) is a plant with significant medicinal and edible value, and its medicinal efficacy varies with its geographical origin. This study employed hyperspectral imaging (HSI) technology in conjunction with a proposed deep learning (DL) network to accurately detect the geographical origin of OPR. Hyperspectral images in the wavelength range of 400–1000 nm were collected from OPR samples originating from four different geographical locations. To fully exploit the spatial and spectral information while reducing redundancy, three key modules were developed. First, the spectral–spatial attention (SSA) mechanism was designed to adaptively update weights from both spectral and spatial dimensions, enabling focused learning of effective features while suppressing irrelevant ones. Second, the multi-scale three-dimensional convolution (M3DC) module was employed to effectively extract fine-grained multi-scale features by dividing the spectral bands into multiple subsets and hierarchically connecting them. Third, depthwise separable convolution (DSC) was utilized to reduce model complexity, and the Transformer module was employed to effectively address the inherent long-range dependency problem in hyperspectral image data, further compensating for the limitations of fixed receptive fields in convolution operations. Notably, the proposed model achieved an optimal classification accuracy of 98.73%, with precision, recall, and F1-score all reaching 98.74%, outperforming various representative algorithms. Additionally, ablation experiments confirmed the effectiveness of each module in improving performance. The encouraging results reveal the potential of combining HSI with advanced deep learning techniques as an efficient, non-destructive solution for the quality monitoring of OPR medicinal materials and other-related traditional medicinal and food materials.