Graph convolutional and random Fourier feature mapping for hyperspectral image clustering

Using subspace clustering to cluster large-scale hyperspectral images (HSI) is an important task. However, most subspace clustering methods typically involve a large amount of computation and are based on linear assumptions. They struggle to capture subspaces and local features, making them unsuitable for processing large-scale HSI. To address these challenges, we propose a novel approach called graph convolutional and random Fourier feature mapping (GCRFFM) for hyperspectral image clustering. First, GCRFFM uses a fast spectral embedding technique to obtain a low-dimensional representation of HSI data, and introduces random Fourier feature mapping (RFFM) to approximate kernel functions, thereby reducing the computational complexity of processing HSI data. Second, we further utilize RFFM during the subspace clustering process by mapping nonlinear data into a high-dimensional linear space to enhance the efficiency of GCRFFM in handling complex nonlinear HSI data. Finally, by using the graph convolutional self-representation model and combining it with RFFM, to fully utilize the spatial and nonlinear features. Experiments were conducted on five common HSI datasets, and the results demonstrated the efficiency and effectiveness of GCRFFM.