A hybrid similarity measure for approximate spectral clustering of remote sensing images

Clustering has been a widely-used method for land cover identification using remote sensing images, thanks to its requirement of limited or no priori information. Among many methods, approximate spectral clustering, which depends on eigendecomposition of a similarity measure, has been popular due to its success and ability to extract arbitrarily-shaped clusters. The similarity measure, which is defined either based on distances or recently on density information, often underutilizes available information for accurate representation of dissimilarity. To address this challenge, a hybrid criterion merging density and distance information is proposed for approximate spectral clustering. Experimental results on remote-sensing images show that the hybrid similarity achieves accuracies greater than the accuracies obtained by the similarity solely based on distance or density.