Identification of growth years of Kudzu root by hyperspectral imaging combined with spectral–spatial feature tokenization transformer

Abstract

As a traditional Chinese medicinal plant, Kudzu root contains a variety of beneficial substances for human health. The growth year of Kudzu root has an important influence both in terms of its economic value and its suitability for applications in medicine and food. Hence, accurately and quickly identifying of Kudzu root growth years holds profound significance for producers, consumers, and regulatory bodies alike. In this study, hyperspectral imaging with a spectral range of 948.72−2512.97nm was used to acquire spectral images of Kudzu samples from three different growth years. To harness the wealth of spectral and spatial information, a powerful spectral–spatial feature tokenization transformer (dubbed SSFTT) was applied. By integrating spectral and spatial information using attention mechanisms, SSFTT could learn joint representations of features. In comparison to classical machine learning methods using only spectral information and representative deep learning methods, the SSFTT model demonstrated outstanding performance, achieving a mean accuracy surpassing 97%. Our study highlights that integration of hyperspectral imaging and deep learning provides an effective and feasible solution for accurately classifying Kudzu root samples from different growth years, holding promising prospects for real-world applications.