文献类型：期刊文献

英文题名：Deep learning-based species identification of gymnosperm xylem: The practice in digital forestry

作者：Zheng, Chang[1,2] He, Tuo[1,2,3] Jiao, Lichao[1,2] Lu, Yang[1,2] Liu, Shoujia[1,2] Li, Tianxiao[1,2] Miao, Yuanyuan[2,4] Feng, Dejun[2,5] Lin, Jian[2,6] Feng, Xian[2,7] Huang, Zhonghua[2,8] Yu, Min[2,9] Qi, Jinqiu[2,10] Zhou, Haibin[1] Zhang, Baohua[11] Yin, Yafang[1,2]

第一作者：Zheng, Chang

通信作者：Jiao, LC[1];Yin, YF[1];Zhang, BH[2]

机构：[1]Chinese Acad Forestry, Res Inst Wood Ind, Beijing 100091, Peoples R China;[2]Natl Forestry & Grassland Adm, Wood Specimen Resource Ctr, Beijing 100091, Peoples R China;[3]Natl Forestry & Grassland Adm, Wildlife Conservat Monitoring Ctr, Beijing 100714, Peoples R China;[4]Northeast Forestry Univ, Coll Mat Sci & Engn, Harbin 150040, Peoples R China;[5]Northwest A&F Univ, Coll Forestry, Yangling 712100, Peoples R China;[6]Beijing Forestry Univ, Coll Mat Sci & Technol, Beijing 100083, Peoples R China;[7]Yunnan Acad Forestry & Grassland, Kunming 650233, Peoples R China;[8]Sichuan Acad Forestry Sci, Chengdu 610081, Peoples R China;[9]Anhui Agr Univ, Sch Forestry & Landscape Architecture, Hefei 230036, Peoples R China;[10]Sichuan Agr Univ, Coll Forestry, Chengdu 611130, Peoples R China;[11]Nanjing Agr Univ, Coll Artificial Intelligence, Nanjing 211800, Peoples R China

年份：2025

卷号：237

外文期刊名：COMPUTERS AND ELECTRONICS IN AGRICULTURE

收录：;EI(收录号：20252318539727);Scopus(收录号：2-s2.0-105006999040);WOS:【SCI-EXPANDED(收录号:WOS:001503995200001)】；

基金：Funding This study was supported financially by the National Key Research and Development Program of China (Grant No. 2023YFF0906301) and Science & Technology Fundamental Resources Investigation Program (Grant No. 2023FY101400) .

语种：英文

外文关键词：Agroforestry health monitoring; Computer vision; Feature visualization; Gymnosperm tree; Wood anatomy; Wood identification

摘要：Computer vision methods have been proven to be feasible in the field of digital forestry, including tree species identification on angiosperm xylem (hardwood) and can overcome the limitations of the traditional wood anatomical method in recognition accuracy. However, few studies exist on gymnosperm xylem (softwood). Owing to imperfections in the wood voucher specimen image database and the lack of a scientific approach to dataset division, its generalizability in practical applications is hindered. In this study, the largest Pinaceae Wood Cross-section Macroscopic (PWCM) Dataset was constructed, including 38,953 images of 22 common species within Pinaceae family distributed in China. Two approaches for dividing the datasets were compared, and four algorithms-ResNet50, SeResNet50, RepVGG-B2, and RepLKNet-31B-were established. Through algorithmic visualization and statistical analysis of misidentified images, an image selection principle applicable to the identification of Pinaceae wood was proposed. The results indicated that the computer vision method was suitable for verifying unknown images in known samples. The best algorithm, RepLKNet-31B, achieved 98.55% and 80.11% accuracy at the genus and species levels, respectively. The results of the algorithm visualization demonstrated that the main features affecting identification were changes in tracheid morphology during the transition between earlywood and latewood, as well as the growth rings and their adjacent tracheids. The error rate statistics of the growth ring widths showed that the best recognition results were achieved when the image contained one to three complete growth rings. Four image categories related to wood anatomical traits-the transition region between juvenile and mature woods, region around the tree pith, sapwood region near the bark, and region with sharp width variations between adjacent growth rings-significantly interfered with the accuracy of the algorithm recognition. This contributes a potential reference for softwood image acquisition standards. This study will provide technical support for agroforestry ecology safety supervision, forest biodiversity conservation, and forensic identification in related fields.