Answers are grounded in this paper and cite their sources.
Hyperspectral imaging offers a non-destructive route for wood species identification, but adjacent spectral bands are highly correlated and physically ordered. When Transformer-based models mix these channels without structural constraints, they can introduce redundant computation and reduce interpretability. Here, we propose SH-Swin, a Structured Spectral Head-Group Swin Transformer that uses spectral-band organisation as a modelling prior rather than treating all bands as homogeneous channels. The method combines continuity-constrained spectral band refinement and wavelength-aware subspace partitioning to reduce redundant cross-band interactions. It then uses Head-Group window attention to confine channel mixing within structured spectral subspaces, followed by Spectral Gating Exchange for adaptive communication across subspaces. Foreground masks generated by UNet++ are further incorporated through SegMask Attention and foreground-aware global average pooling (FG-GAP), strengthening target-region representation while suppressing background interference. On a hyperspectral wood dataset containing 10 tree species, SH-Swin increased overall classification accuracy from 0.892 to 0.960 and Macro-F1 from 0.891 to 0.960. Under the fixed preprocessing setting, the SH-Swin classifier required 6.794 GFLOPs and 21.598 M parameters, compared with 9.334 GFLOPs and 27.597 M parameters for the baseline Swin Transformer V1 classifier. Ablation studies showed that grouped spectral modelling reduces redundant classifier-side computation, whereas Spectral Gating Exchange and foreground-aware mechanisms improve cross-subspace interaction and robust feature modelling. These results support the use of SH-Swin as an interpretable classifier with favourable classifier-side computational efficiency under the fixed preprocessing setting, with potential use in automated wood identification and intelligent forest-product processing.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: