SHCFN: Superpixel-based hybrid Cooperative Fusion Network for Hyperspectral Image Classification

Graph-based methods demonstrate superior ability in capturing non-Euclidean relationships among pixels, thereby making them particularly appropriate for applications in hyperspectral image classification (HSIC). However, these methods encounter excessive computational complexity when processing a large volume of pixels. Additionally, existing graph-based methods tend to emphasize local correlation modeling, which may limit their ability to fully capture global dependencies. To address this challenge, this paper proposes a superpixel-based hybrid cooperative fusion network (SHCFN) for HSIC. First, a superpixel-based multi-hop graph attention module (SMGAT) is designed to effectively capture the multi-level local spatial correlations of HSI while reducing graph modeling complexity. Second, a multi-scale agent-attention module (MSAAT) is introduced to enhance the spatial representation modeling capability of local multi-scale features and global dependencies. Third, a graph channel attention module (GCAT) is proposed to mine inter-channel dependencies while suppressing spectral redundancy, so as to enhance the discriminability of spectral features. Finally, an adaptive gated fusion module (AGFM) is designed to ena ble effective feature fusion between the MSAAT and GCAT modules by filtering redundant information. Experimental results on five publicly available hyperspectral datasets demonstrate that our SHCFN achieves competitive classification performance compared to some existing HSIC approaches.