Center-Pixel Guided Dual-Path Multihead Attention Fusion Transformer for Hyperspectral Image Classification

The performance of hyperspectral image classification (HSIC) relies on the deep integration of spatial and spectral information. Although mainstream Transformer architectures can capture global context, they often overlook the key inductive bias that “the central pixel determines the category,” and also incur high computational costs. Moreover, traditional linear fusion strategies (such as concatenation or element-wise addition) struggle to capture the complex nonlinear interactions between spatial and spectral dimensions. To address these issues, this paper proposes a central-pixel guided dual-path multi-head attention transformer (CPFormer). The model first introduces a spatial-spectral feature embedding (SSFE) module, which generates robust initial representations while preserving inherent spectral characteristics. Guided by the central pixel principle, a dual-path multi-head attention (CP-DMA) module is then proposed: the spatial path uses the central pixel as the Query and incorporates 2D Gaussian positional encoding to enhance local perception; the spectral path employs depthwise separable convolution and bidirectional spectral decay prior to model spectral (channel) continuity. Furthermore, CP-DMA adopts element-wise multiplication for spatial–spectral fusion, enabling adaptive and nonlinear feature interaction. Information plane analysis further confirms that, compared with linear fusion, the multiplication strategy can more effectively compress redundant information and drive feature representations closer to the “minimum sufficient statistic” in information theory. Extensive experiments on four benchmark datasets under limited training samples demonstrate that CPFormer outperforms state-of-the-art methods in classification performance. The source code will be available at: https://github.com/Jianghaob/CPFormer.