PRNet: Pseudotext Reconstruction Network for Domain-Generalized HSI Classification
Tan Guo, Youjinyang Li, Fulin Luo, Chuan Fu · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
Cross-domain hyperspectral image classification remains a significant challenge due to the severe domain shift and the limited ability of existing language-aware models. Specifically, current methods suffer from (1) coarse semantic descriptions that fail to bridge the vision-language gap; (2) inefficient visual encoding of coupled spatial-spectral features; (3) the absence of textual guidance during inference, which leads to a modality imbalance. To overcome these limitations, we propose the Pseudo-text Reconstruction Network (PRNet). Firstly, to enhance domain-invariant semantic expression, a dual-text cross-modal alignment strategy is developed, which aligns visual features with both task-oriented and attribute-level textual descriptions to achieve multi-granularity semantic grounding. Secondly, we design a dual-branch image encoder, termed VSS-CNN Network(VCNNet), which synergistically integrates convolutional neural networks (CNNs) and Visual State Space Blocks (VSSBlocks) to effectively model local spatial structures and long-range spectral dependencies. Thirdly, a visual-to-language feature translation (V2L) module is introduced to reconstruct semantic-consistent pseudo-text from visual embeddings. By distilling these pseudo-text features against real textual prototypes, PRNet ensures deep semantic consistency and enables effective vision-language interaction during both training and inference stages. Experimental results on three challenging HSI benchmarks demonstrate that PRNet effectively captures high-level, transferable semantics and significantly outperforms state-of-the-art domain generalization methods.