A graph-based clustering method with special focus on hyperspectral imaging

Abstract

A common trait of the more established clustering algorithms such as K-Means and HCA is their tendency to focus mainly on the bulk features of the data which causes minor features to be attributed to larger clusters. For hyperspectral imaging this has the consequence that substances which are covered by only a few pixels tend to be overlooked and thus cannot be separated. If small lateral features such as particles are the research objective this might be the reason why cluster analysis fails. Therefore we propose a novel graph-based clustering algorithm dubbed GBCC which is sensitive to small variations in data density and scales its clusters according to the underlying structures. The analysis of the proposed method covers a comparison to K-Means, DBSCAN and KNSC using a 2D artificial dataset. Further the method is evaluated on a multisensor image of atmospheric particulate matter composed of Raman and EDX data as well as an FTIR image of microplastics.

Keywords: Density estimation; Digraph; Exploratory analysis; Graph-based clustering; Hyperspectral imaging; Nearest neighbors.