Superpixel-Based Dual-Neighborhood Contrastive Graph Autoencoder for Deep Subspace Clustering of Hyperspectral Image

Junhong Li, Renxiang Guan, Yuhang Han, Yaowen Hu, Zihao Li, Yanyan Wu, Ziwei Xu, Xianju Li

Abstract: Deep subspace clustering of hyperspectral image (HSI) holds paramount importance for the fine classification of ground elements like land cover and geological units. While graph-based deep subspace learning effectively captures feature representations for clustering, it does not fully harness the global spatial information essential for large-scale imagery requiring multiple graph nodes for representation. This study proposes a Superpixel-based Dual-neighborhood Contrastive Graph Autoencoder for Deep Subspace Clustering (SDCGSC), which consists of: (1) Superpixel-based dual-neighborhood graph autoencoder. Utilizing superpixels and their dual-neighborhood to construct two graphs facilitates the learning of local spatial information. (2) Contrastive graph learning. The model employs graph feature-level contrastive learning within each graph autoencoder branch and graph-level contrastive learning across branches to cultivate more robust features. (3) Self-expression reconstruction and clustering. Fused features from both branches, incorporating relative global information, are utilized for clustering. To validate the effectiveness of the proposed model, extensive experiments are conducted on multiple benchmark datasets. Results demonstrate that SDCGSC significantly outperformed existing state-of-the-art methods. In conclusion, superpixel and contrastive graph autoencoder-based deep subspace clustering is positive for HSI analysis.