Graph Spatial-Spectral Deep Subspace Clustering for Hyperspectral Image Classification

Abstract

Deep subspace clustering (DSC) has become a research hotspot and achieved considerable success in unsupervised hyperspectral image (HSI) classification domain. However, previous researches seldom consider global spatial-spectral information, which leads to an unsatisfactory performance when processing HSI with background (can be seen as a type of noise). In this work, we propose a Graph Spatial-Spectral Deep Subspace Clustering (GS2DSC) method which can make full use of the spatial-spectral information of HSI data. Specifically, a graph attention auto-encoder is utilized to extract the graph-structed features of bands and the spectral information. Then, we employ a fusion strategy to fuse the band-level global spatial-spectral features and the patch-level local spatial-spectral features into joint features to enhance the spatial-spectral information. Our proposed method alleviates the bad influence of the background information and produce an informative affinity matrix. The experimental results on two classic HSI datasets show that the overall accuracy of our approach is improved by 5.31% and 3.05% respectively compared with other DSC-based methods, which demonstrates that the structure can enhance the spatial-spectral information and perform better classification results.