文献类型：期刊文献

英文题名：A Multi-Scale Convolutional Neural Network Combined with a Portable Near-Infrared Spectrometer for the Rapid, Non-Destructive Identification of Wood Species

作者：Pan, Xi[1] Yu, Zhiming[2] Yang, Zhong[1]

通信作者：Yang, Z[1]

机构：[1]Chinese Acad Forestry, Res Inst Wood Ind, Beijing 100091, Peoples R China;[2]Beijing Forestry Univ, Coll Mat Sci & Technol, Beijing 100083, Peoples R China

年份：2024

卷号：15

期号：3

外文期刊名：FORESTS

收录：;EI(收录号：20241315826073);Scopus(收录号：2-s2.0-85188800191);WOS:【SCI-EXPANDED(收录号:WOS:001191837400001)】；

基金：The authors gratefully acknowledge the Xylarium of Southwest Forestry University for supporting the wood specimens used in this study. Additionally, the authors thank Beijing Great Technology Co., Ltd. (Beijing, China) for supporting the portable NIR spectrometer.

语种：英文

外文关键词：wood species identification; near-infrared spectroscopy; convolutional neural network; machine learning; feature extraction

摘要：The swift and non-destructive classification of wood species holds crucial significance for the utilization and trade of wood resources. Portable near-infrared (NIR) spectrometers have the potential for rapid and non-destructive wood species identification, and while several studies have explored related methodologies, further research on their practical application is needed. To address this research gap, this study proposes a multi-scale convolutional neural network (CNN) combined with a portable NIR spectrometer (wavelengths range: 908 to 1676 nm) for wood species identification. To enhance the capability of directly extracting robust features from NIR spectral data collected by a portable spectrometer, the Gramian angular field (GAF) method is introduced to transform 1-dimensional (1D) NIR spectral data into 2-dimensional (2D) data matrices. Furthermore, a multi-scale CNN model is utilized for direct feature extraction. The representation by 2D matrices, instead of 1D NIR spectral data, aligns with 2D convolutional operations and enables a more robust extraction of discriminative features. In the experimental phase, eight wood species were identified using the proposed method, alongside commonly used multivariate data analysis and machine learning (ML) methods. The StratifiedGroupKFold dataset partitioning approach and five-fold cross-validation were used. Additionally, nine spectral preprocessing methods were compared, and principal component analysis (PCA) was used for feature extraction in the ML method. Evaluation metrics, such as accuracy, precision, and recall, were adopted to assess the performance of the methods. The proposed multi-scale CNN model, in combination with 2D GAF matrices of the 1D spectral data, yielded the most accurate results with a mean accuracy of 97.34% in the five-fold validation. These findings present a new approach for the construction of a rapid, non-destructive, and automatic wood species identification method using a portable NIR spectrometer.