Skip to Main Content
Contrary to the traditional clustering methods (often based on parametric models), a recently popular non-parametric method, spectral clustering (SC), employs eigendecomposition of pairwise similarities, and has been shown successful. Despite the advantages of spectral clustering, due to its computational and spatial complexity, its use in remote sensing applications is possible only through approximate spectral clustering (ASC), i.e. SC of the data representatives obtained by quantization or sampling. In this study, we show that, compared to other quantization methods, neural network (self-organizing map or neural gas) based quantization produces better quantization for ASC, to achieve high clustering accuracies.