Fast Subspace Clustering Algorithm with Efficient Similarity-Constrained Sampling for Hyperspectral Images

Hyperspectral images (HSIs) are high-dimensional and complex images that provide rich spectral information of the scenes. Image processing and remote sensing communities are currently developing unsupervised learning methods for HSI classification due to the lack of labeled data. Subspace clustering (SC) methods based on spectral clustering have achieved high clustering performance in real data experiments. However, the computational complexity of such methods prevents their use on large HSI since they require building a similarity matrix that should account for all the pixels in the image. This work proposes an efficient SC-based method that reduces the temporal and spatial computational complexity by splitting the HSI clustering task using similarity-constrained sampling, which considers the spatial information to boost the clustering performance. Experimental results on two widely used HSI data sets show the proposed method's effectiveness, outperforming the baseline methods in more than 20% of overall accuracy.