Dual-Branch Rectangle Transformer for Hierarchical Hyperspectral Super Resolution via Spectral Reversion Contrastive Learning

Fusion-based hyperspectral super-resolution methods use complementary data to address single-source limitations. However, existing approaches prioritize spatial resolution and color fidelity, neglecting spectral preservation—critical for data interpretability. This oversight, combined with the limited adaptability of conventional spectral constraint methods, causes significant spectral distortions in high-resolution outputs. Multiscale feature integration issues further lead to inadequate detail restoration. To tackle these, we propose DRT-Net. Our dual-branch rectangular transformer uses cross-attention to fuse spatial–spectral features, enhancing intermodality interactions. The scale-adaptive aggregator dynamically selects multiscale features for accurate spectral modeling and detail recovery. Additionally, a contrastive learning mechanism enforces spectral feature similarity to ground truth, offering a new approach for modeling complex hyperspectral distributions. Extensive experiments conducted on six mainstream hyperspectral datasets a real-world dataset demonstrated that the proposed DRT-Net outperformed state-of-the-art methods. Furthermore, our network exhibited robust performance and versatility in various scenarios.