Scholarlink

<p>Reliable and rapid identification of tropical timber is essential for protecting threatened species, deterring illegal logging, and strengthening supply-chain control. This study evaluated four supervised classification models for fifteen commercially and conservation-relevant tropical timber species from Costa Rica. Absorbance spectra from 400 to 2500&#xa0;nm were measured on wood surfaces and preprocessed using standard normal variate transformation. The resulting data were analyzed using partial least squares discriminant analysis (PLS-DA), <i>k</i>-nearest neighbors (<i>k</i>-NN), piecewise linear support vector machines (PL-SVM), and linear discriminant analysis (LDA) under a common analytical framework. Model performance was evaluated using accuracy, precision, recall, F1-score, and Cohen’s kappa, together with confusion matrices and species-specific discriminant band analysis. LDA showed the best overall performance, with all evaluation metrics above 0.90. Misclassifications were infrequent and occurred mainly among congeneric species, whereas several species were classified without error. Most discriminant bands were in the SWIR region, especially around 1810–2020&#xa0;nm and 2100–2350&#xa0;nm, with additional contributions from 1420 to 1530&#xa0;nm and selected NIR regions around 1100–1330&#xa0;nm. Stability and composite score values indicated that these bands were repeatedly selected across bootstrap resampling. These findings indicate that linear decision boundaries are sufficient for reliable species-level discrimination in this spectral space and identify stable wavelength ranges with potential for the development of reduced-band sensors. Integrating spectral and machine-learning approaches into checkpoint screening, forensic verification, and timber-governance systems may support efforts to strengthen legal and sustainable timber trade.</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

Discriminating tropical timber species using absorbance spectra and machine learning

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Juan C. Valverde,
  • Dagoberto Arias-Aguilar,
  • Nelson Zamora-Villalobos,
  • Marvin Castillo-Ugalde,
  • Geovanny Figueroa-Mata,
  • Arnoldo Ramirez-Rodríguez

摘要

Reliable and rapid identification of tropical timber is essential for protecting threatened species, deterring illegal logging, and strengthening supply-chain control. This study evaluated four supervised classification models for fifteen commercially and conservation-relevant tropical timber species from Costa Rica. Absorbance spectra from 400 to 2500 nm were measured on wood surfaces and preprocessed using standard normal variate transformation. The resulting data were analyzed using partial least squares discriminant analysis (PLS-DA), k-nearest neighbors (k-NN), piecewise linear support vector machines (PL-SVM), and linear discriminant analysis (LDA) under a common analytical framework. Model performance was evaluated using accuracy, precision, recall, F1-score, and Cohen’s kappa, together with confusion matrices and species-specific discriminant band analysis. LDA showed the best overall performance, with all evaluation metrics above 0.90. Misclassifications were infrequent and occurred mainly among congeneric species, whereas several species were classified without error. Most discriminant bands were in the SWIR region, especially around 1810–2020 nm and 2100–2350 nm, with additional contributions from 1420 to 1530 nm and selected NIR regions around 1100–1330 nm. Stability and composite score values indicated that these bands were repeatedly selected across bootstrap resampling. These findings indicate that linear decision boundaries are sufficient for reliable species-level discrimination in this spectral space and identify stable wavelength ranges with potential for the development of reduced-band sensors. Integrating spectral and machine-learning approaches into checkpoint screening, forensic verification, and timber-governance systems may support efforts to strengthen legal and sustainable timber trade.