Segmentation of elemental EDS maps by means of multiple clustering combined with phase identification

Summary An imaging concept is proposed for the phase identification and segmentation of elemental map images from energy dispersive spectroscopy. The procedure starts with presegmentation using common clustering algorithms, continues with automated identification of the chemical compositions, followed by their screening by professional expertise. The ultimate phases are finally clustered by applying a minimum Euclidean distance classifier. The potential, performance and limitations of the approach are presented on energy dispersive spectroscopy maps acquired by a scanning electron microscope and conducted on samples produced from cement clinker, natural rock and hydrated cement mortar. Nevertheless, the technique is suitable for arbitrary types of materials and general devices for energy dispersive spectroscopy acquisition. It is an approach for extending common energy dispersive spectroscopy analysis by means of visual examination and ratio plots towards quantitative rating. Lay description From a sample, energy dispersive spectroscopy (EDS) systems yield element maps as multiple images at the same location, one map for each chemical element. Such sets of images call for chemical identification of the available phases coupled with the regional segmentation of their boundaries. If this purpose can be achieved as independently as possible from decisions of the evaluator, reproducible quantitative material assessments become feasible. The combination of automatic clustering algorithms with classifiers for ranking the likelihood of the occurrence of some chemical compounds provide the potential to approach this ambitious goal. We propose a suitable workflow and present its functionality on electron microscopy data from cement clinker, cement mortar and natural rock. The proposed technique is not limited to the presented materials, neither to electron microscopy only. Rather, it is applicable to arbitrary types of EDS data sets. The highly automated procedure is a fundamental move from qualitative inspection of EDS maps (always affected by human perception) towards quantitative assessment.