S 2 TA-Fuse: Semantic-Superpixel Tokenized Attention for Spatial–Spectral Fusion

Superpixel paradigms have long been regarded as a well-established approach of mitigating spatial redundancy in hyperspectral images, yet their reliance on non-differentiable and irreversible segmentation renders them unsuitable for end-to-end spatial–spectral fusion (SSF). To tackle this limitation, this study introduces S²TA-Fuse, a transformer-based solver namedSemantic-Superpixel Tokenized Attention for Fusion, which preserves the efficiency of superpixels while removing the need for explicit segmentation. The central design lies in a semantic attention mechanism that adaptively organizes pixels into deformable and content-aware semantic groups. Pixels sharing similar latent states are softly aggregated and encoded as compact tokens, upon which attention is computed to capture intricate long-range dependencies. This formulation endows the model with an inherent ability to accommodate scale variations while maintaining linear computational complexity with respect to the number of pixels. On top of the semantic backbone, two complementary components are devised. The Local Spectral Pyramid enhances the representation of multi-scale spectral cues in the spatial domain, whereas FreqNet supplements global information by modeling frequency-dependent variations through amplitude and phase decomposition. Comprehensive experiments on widely used benchmarks for spatial–spectral fusion demonstrate that S²TA-Fuse consistently surpasses the state of the art both in quantitative accuracy and visual fidelity.