Scholarlink

<p>Multimodal data complementarity can make use of different data feature advantages, solving single-modal problems of one-sided information and weak anti-interference, and improving quantitative analysis accuracy. This study proposed lotus seed origin traceability using hyperspectral imaging (HSI) and Bayesian-optimized convolutional neural network-long short-term memory network (CNN-LSTM) cross-modal fusion. Results showed its 97.22% classification accuracy (highest, better than HSI gray level co-occurrence matrix (GLCM) feature splicing and single modality). The gradient-weighted class activation mapping (GradCAM) visualized the model’s attention area, revealing links between discrimination basis and origin environmental conditions, and providing intuitive evidence for the morphological interpretation of the origin of lotus seeds. As above, the combination of HSI and Bayesian-optimized CNN-LSTM cross-modal fusion can realize the efficient and accurate traceability of the origin of food products and support new alternatives for the realization of brand construction, origin traceability, and quality control of geographical indication food products.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging cross-modal learning with CNN-LSTM: hyperspectral technology for geographical origin tracing of lotus seeds

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Jing Zhao,
  • Yue Yu,
  • Di Wu,
  • Ting Yin,
  • Yifan Tu,
  • Siheng Lu,
  • Qian Qin,
  • Zhanming Li,
  • Qiming Wu

摘要

Multimodal data complementarity can make use of different data feature advantages, solving single-modal problems of one-sided information and weak anti-interference, and improving quantitative analysis accuracy. This study proposed lotus seed origin traceability using hyperspectral imaging (HSI) and Bayesian-optimized convolutional neural network-long short-term memory network (CNN-LSTM) cross-modal fusion. Results showed its 97.22% classification accuracy (highest, better than HSI gray level co-occurrence matrix (GLCM) feature splicing and single modality). The gradient-weighted class activation mapping (GradCAM) visualized the model’s attention area, revealing links between discrimination basis and origin environmental conditions, and providing intuitive evidence for the morphological interpretation of the origin of lotus seeds. As above, the combination of HSI and Bayesian-optimized CNN-LSTM cross-modal fusion can realize the efficient and accurate traceability of the origin of food products and support new alternatives for the realization of brand construction, origin traceability, and quality control of geographical indication food products.