融合波段选择与无偏集成学习的大花红景天高光谱产区判别

陈冲

文章摘要

融合波段选择与无偏集成学习的大花红景天高光谱产区判别

Hyperspectral Geographical Origin Discrimination of Rhodiola crenulata by Integrating Wavelength Selection and Unbiased Ensemble Learning

投稿时间：2026-01-27 修订日期：2026-03-27

DOI：

中文关键词: 高光谱成像大花红景天产区识别极限学习机特征选择

英文关键词: spectroscopy rhodiola crenulata geographical origin discrimination extreme learning machine feature selection

基金项目:光学信息与模式识别湖北省重点实验室开放课题研究基金（课题编号：202204）；安徽省高校杰出青年科研项目（课题编号：2023AH020023）。

作者	单位	邮编
陈冲^*	安徽建筑大学	230601

摘要点击次数: 106

全文下载次数: 0

中文摘要:

大花红景天的品质受地理生态因子影响显著，其化学组分差异在可见-近红外光谱中呈现特异指纹响应。针对传统理化检测破坏性大、耗时长且难以全样覆盖的问题，本文提出一种融合高光谱特征提取、优化波段选择与集成学习的产区识别方法。实验采集400-1100 nm范围内大花红景天粉末光谱，经多元散射校正（MSC）预处理后，采用改进的MI-CARS进行特征波段选择，分析表明，所提取的波段与大花红景天内部官能团（如O-H、C-H键）的振动响应高度相关，增强了特征提取的物理可解释性。在建模阶段，引入K折OOF策略生成无偏元特征，并通过多异构ELM基学习器与顶层ELM实现分层融合与决策集成。结果表明，MI-CARS在波段选择稳定性与物理可解释性方面具有优势，所构建的ELM集成模型识别精度达97.06%，相较于单一ELM模型精度提升了5.24%，并通过对比实验和消融实验证明了该方法的有效性，为珍稀中药材质量评价提供了新的物理检测途径。

英文摘要:

The quality of Rhodiola crenulata is significantly influenced by geographic and ecological factors. These chemical differences result in specific fingerprint responses in the visible and near-infrared (Vis-NIR) spectra. Traditional physical and chemical testing methods are destructive, time-consuming, and difficult to apply to all samples. To address these issues, this study proposes an origin identification method that integrates hyperspectral feature extraction, optimized band selection, and ensemble learning. Hyperspectral data of Rhodiola crenulata powder were collected in the 400–1100 nm range. After Multiplicative Scatter Correction (MSC) preprocessing, an improved MI-CARS algorithm was used for characteristic band selection. Analysis showed that the extracted bands are highly correlated with the vibrational responses of functional groups (such as O-H and C-H bonds). This correlation enhances the physical interpretability of the feature extraction process.During the modeling stage, a K-fold Out-of-Fold (OOF) strategy was introduced to generate unbiased meta-features. Hierarchical fusion and decision integration were then achieved using multiple heterogeneous Extreme Learning Machine (ELM) base learners and a top-level ELM. The results demonstrate that MI-CARS provides advantages in band selection stability and physical interpretability. The constructed ELM ensemble model achieved an identification accuracy of 97.06%, representing a 5.24% improvement over the single ELM model. Comparative and ablation experiments further confirmed the effectiveness of this method. This study provides a new physical detection approach for the quality evaluation of rare traditional Chinese medicines.

View Fulltext 查看/发表评论下载PDF阅读器

关闭