Hyperspectral Image Clustering Based on Weighted Spatial Denoising and Anchor Graph

Hyperspectral clustering is extensively utilized for the interpretation and information extraction of HyperSpectral Image (HSI). However, due to the large spectral variability and complex spatial distributions, HSI clustering presents considerable challenges. Traditional graph-based clustering algorithms of-ten encounter computational bottlenecks when processing large-scale problem. Furthermore, most existing methods inadequately address noise interference and fail to fully leverage spatial information. To address these issues, this paper proposes an HSI clustering algorithm based on weighted spatial denoising and anchor graph. To mitigate noise interference, the proposed method first partitions the original HSI into multiple superpixels and employs weighted averaging using the nearest neighbor pixels to locally smooth the pixels within each superpixel. To reduce computational complexity, kernel density estimation is then adopted to select the pixel with the highest density as an anchor. With the denoised pixels and anchors, each pixel is reconstructed again through a convex combination to capture both local feature and global structure. Finally, fast spectral clustering is performed on anchor graph to obtain clustering results. Extensive experimental results demonstrate that the proposed algorithm overcomes the computational bottlenecks while maintaining high clustering accuracy.