Answers are grounded in this paper and cite their sources.
The effectiveness of modeling joint spatial-spectral features is crucial for the performance of hyperspectral image (HSI) classification. However, existing convolutional neural network (CNN) and Transformer-based methods fail to effectively address rotation invariance in remote sensing scenarios and spatial–spectral feature ambiguity in multiscale feature extraction. To address these issues, we develop MRC-Former, a multiscale rotation-invariant and center-preserving transformer. MRC-Former contains two key modules. One is a symmetric rotation-invariant encoding-based attention (SRIEA) mechanism that models long-range spatial interactions by integrating symmetric encoding expansion with discrete rotations, producing rotation-invariant features. The other is a center-preserving symmetric pooling (CPSP) strategy for spatial downsampling, which retains the central pixel to reduce spectral distortion and spatial inconsistency. Based on these designs, MRC-Former builds hierarchical multiscale spectral–spatial representations, maintaining rotation consistency and preserving central spectral information across layers. Experiments on three benchmark datasets show that it achieves 99.17% OA on Indian Pines, 98.75% on University of Pavia, and 98.89% on Houston 2013. MRC-Former is lightweight, with only 0.08 M parameters compared with 0.15 M for SSFTT.
No takes yet. Share an insight, caveat, or question.
Weng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: