Spatial-Spectral Middle Cross-Attention Fusion Network for Hyperspectral Image Superresolution

Xiujuan Lang,Tao Lu,Yanduo Zhang,Junjun Jiang,Zixiang Xiong
DOI: https://doi.org/10.14358/pers.24-00007r2
2024-10-31
Photogrammetric Engineering & Remote Sensing
Abstract:The spatial and spectral features of hyperspectral images exhibit complementarity, and neglecting them prevents the full exploitation of useful information for superresolution. This article proposes a spatial-spectral middle cross-attention fusion network to explore the spatial-spectral structure correlation. Initially, we learn spatial and spectral features through spatial and spectral branches instead of single ones to reduce information compression. Then, a novel middle-cross attention fusion block that includes middle features fusion strategy and cross-attention is proposed to fuse spatial-spectral features to enhance their mutual effects, which aims to explore the spatial-spectral structural correlations. Finally, we propose a spectral feature compensation mechanism to provide complementary information for adjacent band groups. The experimental results show that the proposed method outperforms state-of-the-art algorithms in object values and visual quality.
geosciences, multidisciplinary,geography, physical,remote sensing,imaging science & photographic technology
What problem does this paper attempt to address?