Information fusion technology for terahertz spectra and hyperspectral imaging in wood species identification

This paper proposes the use of information fusion technology to identify different wood species by combining spectral and spatial information from hyperspectral images and terahertz (THz) spectra. The study utilized five species of coniferous wood as experimental samples. The hyperspectral and terahertz raw images and spectra acquired by the spectroscopic instruments were preprocessed using standard normal variational transform (SNV). Three methods, namely, competitive adaptive reweighting (CARS), uninformative variable elimination (UVE), and random frog hopping (RF), were employed to select relevant frequency features in both hyperspectral image spectral information and THz spectra. For hyperspectral image spatial information, three algorithms, grayscale co-occurrence matrix (GLCM), local binary pattern (LBP), and Gaussian Markov random field (GMRF) were used to extract texture features. Subsequently, these three sets of extracted features were recognized separately using an extreme learning machine (ELM) model. The results showed that the accuracies achieved by the three features alone in wood identification were 71.8% for the spectral information, 85% for the hyperspectral image spatial information, and 91.7% for THz spectra. However, there was still room for improvement in terms of accuracy. Consequently, the study fused the hyperspectral image spectral and spatial information with THz spectral information, and the ELM model was employed to recognize the fused data. The results indicated that this fusion method led to a substantial enhancement in wood identification accuracy, achieving an impressive 96.7%. This accuracy markedly surpassed the highest recognition accuracy achieved by a single information feature.