ASCRNet: An Abundance-Guided Spectral-Spatial Collaborative Reconstruction Network for Single Hyperspectral Image Super-Resolution

Abstract

Single hyperspectral image super-resolution (HSI-SR) aims to enhance the spatial resolution of hyperspectral images without modifying imaging hardware. However, existing methods still face challenges in efficiently utilizing high-dimensional spectral information, exploiting intrinsic spectral mixing characteristics, and simultaneously recovering local details and long-range spectral-spatial dependencies. To address these issues, we propose the Abundance-Guided Spectral-Spatial Collaborative Reconstruction Network (ASCRNet) for single HSI-SR. Specifically, ASCRNet first transforms the input low-resolution hyperspectral image into a compact abundance-guided representation space through an Abundance Generation Module (AGM). This strategy not only reduces reconstruction complexity in the original high-dimensional spectral domain but also introduces spectral mixing priors to facilitate more effective spectral representation learning. Subsequently, a Residual Abundance Reconstruction Network (RARN) performs residual refinement in the abundance-guided space to recover missing high-frequency spectral-spatial details while capturing long-range contextual dependencies. Finally, a Shared Spectral Reconstruction Layer (SSRL) projects the refined representations back to the hyperspectral domain, enabling consistent spectral reconstruction and generating the final super-resolved image. Experiments on the Chikusei and MDAS datasets demonstrate that ASCRNet achieves competitive reconstruction accuracy and spectral fidelity compared with recent representative methods, particularly under challenging large-scale upsampling settings.