首页|Optimizing recognition models for wood species identification using multi-spectral techniques
Optimizing recognition models for wood species identification using multi-spectral techniques
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
Walter De Gruyter
Abstract The identification of wood species is critical for effective forest management and conservation. However, existing methods frequently neglect the potential benefits of integrating multiple spectral properties. Current research often relies on uniform feature selection and modelling methods, despite the fact that each spectral technique has its own optimal recognition model. This study aims to address the gap in THz, NIR, and HSI data, systematically comparing these techniques to derive the best recognition models. This study employed the SNV transformation to preprocess the spectral data of four coniferous and one broad-leaved species. Subsequently, relevant frequency features were filtered using competitive adaptive weighting (CARS), uninformative variable elimination (UVE), and successive projections algorithm (SPA) to reduce dimensionality and enhance recognition efficacy. A comparative analysis of five algorithms – extreme learning machines (ELM), support vector machines (SVM), random forests (RF), long short-term memory networks (LSTM), and convolutional neural networks (CNN) – was conducted to identify the optimal recognition models. The results demonstrated impressive accuracy rates: 96.7 % for NIR (UVE + CARS + ELM), 96.7 % for HSI (CARS + LSTM), and 98.3 % for THz (SPA + RF). This analysis identifies the most effective models for each spectral type, advancing the field of spectral wood identification.
School of Technology, Beijing Forestry University||State Key Laboratory of Efficient Production of Forest Resources||Key Laboratory of National Forestry and Grassland Administration on Forestry Equipment and Automation
NETL
NSTL
Walter De Gruyter
