Unsupervised wood species identification based on multiobjective optimal clustering and feature fusion

This paper proposes an unsupervised wood species identification approach utilizing multiobjective optimization clustering and feature fusion. To address the inherent limitations of single-band spectra in capturing comprehensively wood characteristics, this approach integrates preprocessed low-dimensional terahertz (THz) and hyperspectral data. Additionally, to address the challenge of selecting the optimal k-value in clustering, an unsupervised wood clustering algorithm was developed, employing multiobjective optimization and evolutionary algorithms. This proposed algorithm incorporated a prototype coding method for initialization, density peak clustering for pattern identification, and an improved firefly optimization algorithm to introduce cross-variation and maintain population diversity. To further refine the clustering process, a selection operator based on grid division and fast non-dominated sorting was designed, optimizing the clustering performance. The model was evaluated on a dataset containing hyperspectral and THz spectra from 400 samples, representing ten wood species—five coniferous and five broadleaf species. Experimental results indicated that fusing the spectral data resulted in a 3.5% increase in clustering purity compared to individual datasets. Moreover, the proposed algorithm outperformed established clustering methods such as DBSCAN, OPTICS, and density peak clustering, achieving a maximum clustering purity of 91.25% across both internal and external clustering metrics. These findings demonstrate the effectiveness of the multi-spectral fusion approach and the proposed algorithm in enhancing wood species identification accuracy, offering a promising avenue for improving non-destructive evaluation methods in forestry and material sciences.