Scholarlink

<p>Fossil diatoms preserved in sediments serve as valuable bioindicators for reconstructing past climate and environmental conditions, owing to their sensitivity to ecological changes and exceptional preservation. Nevertheless, conventional manual classification methods are resource-intensive, time-consuming and subjective, particularly when conducted by individuals without specialized expertise. Such challenges constrain the efficiency and broader applicability of diatom-based paleoclimate reconstructions, reinforcing the necessity for automated classification approaches. To address this challenge, we propose an automated diatom classification approach using deep learning. Given the limited availability of diatom data, we employ (1) data augmentation to enhance training data diversity, (2) transfer learning to adapt and fine-tune pre-trained models to diatom-specific features with limited data, and (3) ensemble learning to improve classification reliability. We adopt ResNet, DenseNet, and Wide ResNet, fine-tuned with augmented diatom data and combined using various ensemble strategies. Our experiment results demonstrate improved classification accuracy, achieving approximately 91% accuracy for the classification of 50 species and approximately 86% for the classification of 131 species. This demonstrates the potential to enhance the speed and reliability of diatom analysis, enabling more precise and scalable paleoclimate reconstructions.</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

Robust automated identification of diatom microfossils under severe data scarcity via deep ensemble learning

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Yun Chang Choi,
  • Tae Hyung Kim,
  • Jungjae Park,
  • Jiwoo Han

摘要

Fossil diatoms preserved in sediments serve as valuable bioindicators for reconstructing past climate and environmental conditions, owing to their sensitivity to ecological changes and exceptional preservation. Nevertheless, conventional manual classification methods are resource-intensive, time-consuming and subjective, particularly when conducted by individuals without specialized expertise. Such challenges constrain the efficiency and broader applicability of diatom-based paleoclimate reconstructions, reinforcing the necessity for automated classification approaches. To address this challenge, we propose an automated diatom classification approach using deep learning. Given the limited availability of diatom data, we employ (1) data augmentation to enhance training data diversity, (2) transfer learning to adapt and fine-tune pre-trained models to diatom-specific features with limited data, and (3) ensemble learning to improve classification reliability. We adopt ResNet, DenseNet, and Wide ResNet, fine-tuned with augmented diatom data and combined using various ensemble strategies. Our experiment results demonstrate improved classification accuracy, achieving approximately 91% accuracy for the classification of 50 species and approximately 86% for the classification of 131 species. This demonstrates the potential to enhance the speed and reliability of diatom analysis, enabling more precise and scalable paleoclimate reconstructions.