Enhancing classification in correlative microscopy using multiple classifier systems with dynamic selection

Correlative microscopy combines data from different microscopical techniques to gain unique insights about specimens. A key requirement to unlocking the full potential is an advanced classification method that can combine the various analytical signals into physically meaningful phases. The prevalence of highly imbalanced class distributions and high intra-class variability in such real applications makes this a difficult task, yet no study of classifier performance exists in the context of correlative microscopy. This paper investigates the application of both single classifiers as well as multiple classifier systems with dynamic selection. The test sample used for evaluation and comparison of the results is a volcanic rock where data from correlative Raman spectroscopy, Scanning Electron Microscopy (SEM) and Energy Dispersive X-ray Spectroscopy (EDS) are available and prepared for algorithmic evaluation. The results show that multiple classifier systems outperform single classifiers reaching an area under the curve of the receiver operating characteristic of 99% demonstrating the applicability of automated classification in correlative microscopy. Thus, this paper contributes by highlighting the potential of combining the research fields of correlative microscopy and machine learning.