Scholarlink

<p>The rapid and accurate discrimination of cultivars and genotypes is of considerable importance for breeding programs, seed certification, genetic purity assessment, and digital phenotyping applications. In this study, an interpretable image-based phenotyping framework integrating morphological, colorimetric, and textural traits was developed for soybean genotype classification. In addition, chemical composition traits were evaluated as complementary genotype-level descriptors to support the biological interpretation of phenotypic differences among soybean genotypes. The study utilized a total of 2520 individual seed images from five different soybean genotypes: Ilksoy, Kristal, Lider, Mersoy, and Traksoy. The images were captured using a flatbed scanner with a resolution of 1200 dpi under standardized imaging conditions. Following the segmentation process, morphological features, Red–Green–Blue (RGB)-based color parameters, and Gray Level Co-occurrence Matrix (GLCM)-based texture features were extracted from the seeds. Among the machine learning models, the SMO-SVM model demonstrated the most balanced performance, with a precision of 71.9%, a recall of 71.4%, an F1-score of 71.4%, and an MCC value of 0.644. The ROC-AUC values ranged from 91.8 to 92.1% across all machine learning models. Deep learning models, however, outperformed traditional machine learning algorithms. While the VGG16 model produced the best results with approximately 93% accuracy and an MCC value of 0.915, the ResNet101 model achieved an accuracy rate of approximately 92%. Feature importance analyses showed that area, perimeter, aspect ratio, and certain GLCM texture parameters were the most effective variables for genotype discrimination. Grad-CAM analyses revealed that deep learning models focused on biologically meaningful seed regions. The obtained results indicate that image-based seed phenotyping using morphological, colorimetric, and textural traits provides an effective approach for soybean genotype classification, while chemical composition data offer complementary biological information for genotype characterization and interpretation.</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

Image-based multimodal seed phenotyping using morphological, colorimetric, and textural traits for soybean genotype classification

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Sevim Akçura

摘要

The rapid and accurate discrimination of cultivars and genotypes is of considerable importance for breeding programs, seed certification, genetic purity assessment, and digital phenotyping applications. In this study, an interpretable image-based phenotyping framework integrating morphological, colorimetric, and textural traits was developed for soybean genotype classification. In addition, chemical composition traits were evaluated as complementary genotype-level descriptors to support the biological interpretation of phenotypic differences among soybean genotypes. The study utilized a total of 2520 individual seed images from five different soybean genotypes: Ilksoy, Kristal, Lider, Mersoy, and Traksoy. The images were captured using a flatbed scanner with a resolution of 1200 dpi under standardized imaging conditions. Following the segmentation process, morphological features, Red–Green–Blue (RGB)-based color parameters, and Gray Level Co-occurrence Matrix (GLCM)-based texture features were extracted from the seeds. Among the machine learning models, the SMO-SVM model demonstrated the most balanced performance, with a precision of 71.9%, a recall of 71.4%, an F1-score of 71.4%, and an MCC value of 0.644. The ROC-AUC values ranged from 91.8 to 92.1% across all machine learning models. Deep learning models, however, outperformed traditional machine learning algorithms. While the VGG16 model produced the best results with approximately 93% accuracy and an MCC value of 0.915, the ResNet101 model achieved an accuracy rate of approximately 92%. Feature importance analyses showed that area, perimeter, aspect ratio, and certain GLCM texture parameters were the most effective variables for genotype discrimination. Grad-CAM analyses revealed that deep learning models focused on biologically meaningful seed regions. The obtained results indicate that image-based seed phenotyping using morphological, colorimetric, and textural traits provides an effective approach for soybean genotype classification, while chemical composition data offer complementary biological information for genotype characterization and interpretation.