A multiple spectral important feature fusion method for wood species identification

This study proposes a novel method for wood species identification, that employs importance-based feature selection integrated with a multiple spectral fusion technique. Specifically, the fusion integrates near-infrared spectroscopy (NIR), hyperspectral imaging spectral information, and terahertz (THz) spectroscopy. The experimental samples comprised four conifers and one broadleaf wood. Preprocessing of the spectral data was conducted using a combination of Savitzky-Golay smoothing (SG), Standard Normal Variate (SNV) correction, and normalization techniques. A hybrid feature selection method, combining random forest (RF) and gradient boosting decision tree (GBDT) algorithms, was then employed to extract the most important spectral features. To enhance clustering stability and mitigate the risk of overfitting, data augmentation was performed using a variational auto-encoder (VAE) augmented with self-attention (SA) mechanisms. Subsequently, the fused multiple spectral data, containing the most significant features from both individual and combined spectra, were subjected to K-means clustering. The clustering performance was assessed using metrics such as accuracy (ACC), normalized mutual information (NMI), and adjusted rand index (ARI). The results revealed that the fusion of NIR features with the top 50 features with the highest importance of the top 60 THz features yielded the most optimal results. The clustering evaluation metrics demonstrated an ACC of 0.945, an NMI of 0.957, and an ARI of 0.959. The hybrid feature selection approach facilitates a deeper understanding of the critical features influencing the performance of wood species identification models, thereby enabling more effective feature selection during the development of machine learning models.