Scholarlink

<p>Pollen classification based on surface morphology remains a challenging problem due to high inter-class similarity and subtle microstructural variations. In this study, a physics-guided deep learning framework is proposed for classifying pollen species using Scanning Electron Microscopy (SEM) images. The dataset consists of 960 SEM images from six plant species, with a balanced class distribution. SEM images encode not only visual texture but also physically meaningful information related to surface roughness, ridge orientation, and pore distribution. Conventional convolutional neural networks rely primarily on pixel-intensity patterns and often fail to capture morphology-driven differences. To address this limitation, a hybrid representation learning framework is developed by integrating physics-based texture descriptors with deep convolutional features. A 68-dimensional morphology-aware descriptor set is constructed, including roughness statistics, gray-level co-occurrence features, Gabor-based orientation responses, ridge density, and pore characteristics. These descriptors are fused with features extracted from a ResNet-based convolutional backbone using a structured feature fusion strategy. The model is trained using AdamW optimization with cosine annealing, and evaluated under a stratified train–validation–test protocol. The proposed framework achieves 99.31% classification accuracy, 0.9931 F1-score, and 1.00 AUC, with only one misclassification on the test set. While baseline CNN models reach similar accuracy, the proposed method improves latent space separability by ~ 15% in Silhouette score and demonstrates significantly higher robustness under noise perturbations. The integration of physics-guided descriptors enhances representation stability and interpretability without increasing computational complexity. This study introduces a microscopy-aware, physics-informed learning paradigm for SEM image analysis. The proposed approach bridges the gap between data-driven deep learning and physically grounded modeling, providing a robust and interpretable framework for microstructural classification tasks.</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

Physics-guided deep representation learning for SEM-based pollen classification via morphology-aware hybrid feature integration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Fatih Ekinci,
  • Aycan Tosunoğlu,
  • Özge Demir,
  • Eda Kumru,
  • Mustafa Sevindik,
  • Ayhan Aydın,
  • Güney Uğurlu,
  • Mehmet Serdar Güzel,
  • Ilgaz Akata

摘要

Pollen classification based on surface morphology remains a challenging problem due to high inter-class similarity and subtle microstructural variations. In this study, a physics-guided deep learning framework is proposed for classifying pollen species using Scanning Electron Microscopy (SEM) images. The dataset consists of 960 SEM images from six plant species, with a balanced class distribution. SEM images encode not only visual texture but also physically meaningful information related to surface roughness, ridge orientation, and pore distribution. Conventional convolutional neural networks rely primarily on pixel-intensity patterns and often fail to capture morphology-driven differences. To address this limitation, a hybrid representation learning framework is developed by integrating physics-based texture descriptors with deep convolutional features. A 68-dimensional morphology-aware descriptor set is constructed, including roughness statistics, gray-level co-occurrence features, Gabor-based orientation responses, ridge density, and pore characteristics. These descriptors are fused with features extracted from a ResNet-based convolutional backbone using a structured feature fusion strategy. The model is trained using AdamW optimization with cosine annealing, and evaluated under a stratified train–validation–test protocol. The proposed framework achieves 99.31% classification accuracy, 0.9931 F1-score, and 1.00 AUC, with only one misclassification on the test set. While baseline CNN models reach similar accuracy, the proposed method improves latent space separability by ~ 15% in Silhouette score and demonstrates significantly higher robustness under noise perturbations. The integration of physics-guided descriptors enhances representation stability and interpretability without increasing computational complexity. This study introduces a microscopy-aware, physics-informed learning paradigm for SEM image analysis. The proposed approach bridges the gap between data-driven deep learning and physically grounded modeling, providing a robust and interpretable framework for microstructural classification tasks.