PST-DUNet: A Progressive Spectral Topology-Aware Deep Unfolding Network for Hyperspectral Image Super-Resolution
Abstract
Hyperspectral images (HSIs) contain rich spectral fingerprints. However, restricted by hardware bottlenecks, the acquired images often fail to satisfy the requirements for fine-grained observation. Although HSI super-resolution has achieved significant progress, many existing methods still primarily emphasize spatial detail recovery and exploit cross-band information through implicit feature interactions or output-level spectral constraints. Such a spatial-first reconstruction paradigm may provide insufficient cross-band coordination during progressive detail generation, leading to inconsistent interband response patterns and spectral distortion. To address this issue, we propose a progressive spectral topology-aware deep unfolding network (PST-DUNet). In particular, we adopt a progressive low-resolution (LR)–mid resolution (MR)–high-resolution (HR) reconstruction strategy to reduce the difficulty of direct large-scale reconstruction. In the MR stage, a parameterized observation-consistency correction is introduced to constrain the intermediate estimate with the LR observation. Meanwhile, we design an instance-adaptive dynamic spectral topology inference (DSTI) module to infer reconstruction-state-conditioned interband dependencies. Furthermore, a spectral topology rectifier (STR) is introduced to transform the topology-aggregated spectral representation into channelwise gating coefficients, thereby modulating spatial feature updates and alleviating spectral distortion. Extensive experiments on the evaluated public HSI datasets like Pavia Center, Chikusei, and CAVE demonstrate that PST-DUNet achieves superior reconstruction performance under the nonblind degradation setting, providing a favorable balance between spatial detail recovery and spectral consistency preservation.