Scholarlink

<p>To achieve nondestructive classification of green pepper quality, a classification framework integrating Gramian Angular Summation Field (GASF) encoding, Borderline-SMOTE, three-channel input, and ResNet-SVM was developed using hyperspectral data in the range of 372.66–1039.65&#xa0;nm. One-dimensional spectra were converted into GASF images to enhance the representation of inter-band correlation information. Borderline-SMOTE was employed to alleviate class imbalance and improve the recognition of boundary samples. A three-channel input strategy integrating raw, SG-, and SNV-processed spectra was employed to enhance feature representation. The results showed that ResNet-SVM outperformed the standalone ResNet, demonstrating the synergistic advantage of deep feature extraction and SVM-based discrimination. Compared with single-channel input, the three-channel strategy further improved classification performance. Among all tested configurations, ResNet-SVM combined with Borderline-SMOTE and SG-SNV achieved the best overall performance, attaining a test accuracy of 95.0%. SHAP and Grad-CAM analyses further identified the key spectral bands associated with model decisions. This study provides an accurate, robust, and interpretable framework for the nondestructive classification of visible quality conditions in green peppers.</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

A Multi-Channel GASF-Based hyperspectral classification framework integrating feature enhancement and interpretability for green chili pepper quality assessment

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Xiong Li,
  • Yawen Guo,
  • Xinlin Xiong,
  • Shaohe Jiao,
  • Yande Liu,
  • Ouyang Aiguo

摘要

To achieve nondestructive classification of green pepper quality, a classification framework integrating Gramian Angular Summation Field (GASF) encoding, Borderline-SMOTE, three-channel input, and ResNet-SVM was developed using hyperspectral data in the range of 372.66–1039.65 nm. One-dimensional spectra were converted into GASF images to enhance the representation of inter-band correlation information. Borderline-SMOTE was employed to alleviate class imbalance and improve the recognition of boundary samples. A three-channel input strategy integrating raw, SG-, and SNV-processed spectra was employed to enhance feature representation. The results showed that ResNet-SVM outperformed the standalone ResNet, demonstrating the synergistic advantage of deep feature extraction and SVM-based discrimination. Compared with single-channel input, the three-channel strategy further improved classification performance. Among all tested configurations, ResNet-SVM combined with Borderline-SMOTE and SG-SNV achieved the best overall performance, attaining a test accuracy of 95.0%. SHAP and Grad-CAM analyses further identified the key spectral bands associated with model decisions. This study provides an accurate, robust, and interpretable framework for the nondestructive classification of visible quality conditions in green peppers.