Lunga, D2014-03-042014-03-042013-12Lunga, D. 2013. A block structure Laplacian for hyperspectral image data clustering. In: Twenty-Fourth Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), Johannesburg, South Africa, 2-3 December 2013http://www.prasa.org/proceedings/2013/prasa2013-08.pdfhttp://hdl.handle.net/10204/7270Twenty-Fourth Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), Johannesburg, South Africa, 2-3 December 2013Over the past decade, the problem of hyperspectral data clustering has generated a growing interest from various fields including the machine learning community. This paper presents an analysis of the traditional spectral clustering approach and points to new directions that boost unsupervised pattern classification. In particular, the paper offers design insights on the generation of a well structured graph Laplacian based on an affinity function that induces context-dependence to create compact neighborhoods. A novel bilateral-kernel (affinity) function exploits the spatial information to generate a diagonal-block structured Laplacian. Experimental validations through the analysis of eigenvalues and eigenvectors demonstrate the benefits of seeking block structured affinities in hyperspectral image clustering and visualization.enHyperspectral image data clusteringHyperspectral Laplacian eigenspectrum analysisNormalized graph LaplacianConference PresentationLunga, D. (2013). A block structure Laplacian for hyperspectral image data clustering. PRASA 2013 Proceedings. http://hdl.handle.net/10204/7270Lunga, D. "A block structure Laplacian for hyperspectral image data clustering." (2013): http://hdl.handle.net/10204/7270Lunga D, A block structure Laplacian for hyperspectral image data clustering; PRASA 2013 Proceedings; 2013. http://hdl.handle.net/10204/7270 .TY - Conference Presentation AU - Lunga, D AB - Over the past decade, the problem of hyperspectral data clustering has generated a growing interest from various fields including the machine learning community. This paper presents an analysis of the traditional spectral clustering approach and points to new directions that boost unsupervised pattern classification. In particular, the paper offers design insights on the generation of a well structured graph Laplacian based on an affinity function that induces context-dependence to create compact neighborhoods. A novel bilateral-kernel (affinity) function exploits the spatial information to generate a diagonal-block structured Laplacian. Experimental validations through the analysis of eigenvalues and eigenvectors demonstrate the benefits of seeking block structured affinities in hyperspectral image clustering and visualization. DA - 2013-12 DB - ResearchSpace DP - CSIR KW - Hyperspectral image data clustering KW - Hyperspectral Laplacian eigenspectrum analysis KW - Normalized graph Laplacian LK - https://researchspace.csir.co.za PY - 2013 T1 - A block structure Laplacian for hyperspectral image data clustering TI - A block structure Laplacian for hyperspectral image data clustering UR - http://hdl.handle.net/10204/7270 ER -