Pore segmentation in electron micrographs: A probabilistic approach by ensemble machine learning

In this study, an automatic method for segmenting pores in scanning electron microscopy images was developed. An ensemble of machine learning classifiers was combined with a fully connected conditional random field to obtain a spatial pore probability field. This field was then thresholded to produce coherent binary pore masks, and a confidence per pore C_l was defined to quantify the reliability of the segmentation. The approach was demonstrated on a broad-ion-beam polished sample of the shaley facies of the Opalinus Clay. Accurate segmentation enabled the derivation of pore size distributions (PSD), pore morphologies, orientations, and spatial statistics. By using the median of C_l per size range, a data-driven lower truncation limit for PSD fitting was established. The resulting microstructural metrics supported the interpretation of rock properties such as permeability. These results highlighted the method’s relevance for materials such as Opalinus Clay, which is investigated as a potential candidate for a host rock for nuclear waste storage.