Multiscale adaptive graph attention band selection for hyperspectral rosewood species classification

The identification of precious timber represented by national standard Rosewood remains challenged by the time-intensive and destructive nature of conventional approaches. Hyperspectral imaging offers a non-destructive alternative, yet its application is hindered by high dimensionality and spectral redundancy. Existing methods often struggle to simultaneously capture global spectral dependencies and localized chemical features. This study presents an integrated framework combining a multiscale adaptive graph attention-based band selection method with stability consistency (MAGABSS-SC) and a dual-domain resonant state space network (D2R-S2Net). MAGABSS-SC employs supervised multiscale attention and PageRank propagation to identify topologically robust bands, which are directly fed into D2R-S2Net to integrate frequency-domain filtering with temporal state space sequence modeling.

Evaluated on a hyperspectral dataset of eighteen Rosewood species comprising 1800 spectral samples, D2R-S2Net achieved an accuracy of 93.33% using full-band input, outperforming support vector machines, random forests, and one-dimensional vision transformers. When combined with MAGABSS-SC, the classification accuracy reached 95.93% while retaining 80% of the spectral bands. The selected wavelengths exhibited contiguous cluster distributions aligned with general absorption intervals reported in wood spectroscopy literature.

The proposed framework provides an accurate and non-destructive approach for the classification of endangered rosewood species. By mitigating spectral redundancy and noise interference, it demonstrates the feasibility of combining graph-based band selection with dual-domain deep learning for timber identification and forest resource regulation.