Scholarlink

<p>Species identification is crucial in biodiversity conservation including combating the illegal trade in timbers. Traditional methods usually cannot identify timbers to the species-level and the sharp decline in the number of taxonomists has exacerbated this challenge. Several attempts have been made to utilize computer vision for wood identification, but some fundamental problems remain regarding dataset split (training, validation and test dataset), model performance, and how deep learning models interpret complex wood anatomical features. Cross-sectional images of seven endangered <i>Pterocarpus</i> species were obtained from the scientific wood collection (Wood Collection of Chinese Academy of Forestry), and four convolutional neural network models (ResNet-50, ResNet-152, WideResNet-50, and SEResNet-50) were trained and tested at specimen-level after image data augmentation, i.e. Crop (C), Rotating before Center Cropping (RC). Layer class activation mapping (Layer-CAM) was used to investigate diagnostic characters to identify each species. The results indicated that the accuracy of the four models was higher when the images were preprocessed using the RC strategy than C strategy. We found that WideResNet-50 identified <i>Pterocarpus</i> samples to 87.56% accuracy, outperforming the other three models. The heat maps showed that the models identified the same features recognized by the human eyes. All four deep learning models focused on the axial parenchyma groupings and vessel groupings of the xylem, although the features detected varied slightly for the different models. These results demonstrate that computer vision-based species identification is a practical means to identify wood samples and can be used to help prevent the illegal trade of timbers and conserve species diversity without relying on taxonomic knowledge and expertise.</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

Automated species discrimination and feature visualization of closely related Pterocarpus wood species using deep learning models: comparison of four convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Shoujia Liu,
  • Chang Zheng,
  • Tuo He,
  • Weihui Zhan,
  • Peter Gasson,
  • Yang Lu,
  • Yafang Yin

摘要

Species identification is crucial in biodiversity conservation including combating the illegal trade in timbers. Traditional methods usually cannot identify timbers to the species-level and the sharp decline in the number of taxonomists has exacerbated this challenge. Several attempts have been made to utilize computer vision for wood identification, but some fundamental problems remain regarding dataset split (training, validation and test dataset), model performance, and how deep learning models interpret complex wood anatomical features. Cross-sectional images of seven endangered Pterocarpus species were obtained from the scientific wood collection (Wood Collection of Chinese Academy of Forestry), and four convolutional neural network models (ResNet-50, ResNet-152, WideResNet-50, and SEResNet-50) were trained and tested at specimen-level after image data augmentation, i.e. Crop (C), Rotating before Center Cropping (RC). Layer class activation mapping (Layer-CAM) was used to investigate diagnostic characters to identify each species. The results indicated that the accuracy of the four models was higher when the images were preprocessed using the RC strategy than C strategy. We found that WideResNet-50 identified Pterocarpus samples to 87.56% accuracy, outperforming the other three models. The heat maps showed that the models identified the same features recognized by the human eyes. All four deep learning models focused on the axial parenchyma groupings and vessel groupings of the xylem, although the features detected varied slightly for the different models. These results demonstrate that computer vision-based species identification is a practical means to identify wood samples and can be used to help prevent the illegal trade of timbers and conserve species diversity without relying on taxonomic knowledge and expertise.