Cross-Scale Spectral–Spatial Interaction With Nonnegative Positional Encoding for Hyperspectral Image Classification

Transformers are emerging as a powerful alternative to convolutional neural networks for hyperspectral image (HSI) classification. However, most existing approaches either neglect the hierarchical spectral-spatial interactions across scales or suffer from unstable performance due to inadequate spectral encoding. We introduce the Cross-Scale Spectral-Spatial Transformer (CSSTFormer), which integrates two key innovations: (i) Cross-Scale Spectral-Spatial Attention (CSSA), a set of bidirectional pathways linking fine-resolution, medium-resolution, and coarse-resolution feature maps to exchange hierarchical context; and (ii) a learnable non-negative spectral positional encoding (NN-SPE) that reweights principal-component features via a magnitude-only (non-negative) component-wise bias, providing an interpretable spectral gating and acting as a lightweight regularizer that is empirically associated with improved robustness in our ablations. Extensive experiments on four public HSI benchmarks demonstrate that our method consistently outperforms state-of-the-art CNN and transformer baselines in overall accuracy (OA), average accuracy (AA), and Kappa coefficient. Comprehensive ablation studies reveal that cross-scale interaction serves as the primary driver of performance gains, while the positional encoding and non-negative constraint further contribute to improved robustness and generalization.