Mudstones and shales are of great geoscientific interest given their potential as primary energy carriers
(“shale oil/gas”), as well as their importance to numerous other energy-related fields such as hydrogen
underground storage, drilling for geothermal energy, or nuclear waste disposal. Electron beam
techniques and particularly broad ion beam - scanning electron microscopy (BIB-SEM) have become
more and more important in the investigation of such kinds of fine-grained sedimentary rocks, as they
can provide porosity quantification and information about the pore space morphology in 2D (Desbois
et al., 2009; Klaver et al., 2012; Misch et al., 2018a, 2021). BIB-SEM maps are large digital image files
that have to be processed in order to obtain reliable pore characteristics, such as total porosity and pore
size distributions (PSDs). Image processing tools like Photoshop, imageJ, and ArcGIS are widely used
for image segmentation based on the thresholding principle (Abràmoff et al., 2004; Desbois et al., 2009;
Klaver et al., 2012, 2015; Li et al., 2019; Misch et al., 2018b). However, different image processing
approaches may lead to discrepancies in the obtained pore information and generate highly variable
results. The effects of different image segmentation approaches on the obtained PSDs and pore
geometry factors have not been quantified nor well documented yet.
Variability in image processing results is not only caused by the fundamentally different methods, but
also inherently linked to i) the physical resolution limit of the respective imaging technique, as well as
ii) the inconsistent subjective interpretation of features by the human eye. As shown in Figure 1,
boundaries of pore cross-sections in BIB-SEM images can be ambiguous for features approaching the
maximum resolution/minimum pixel size of an image. The grey value pattern of pore boundaries is
typically gradual rather than sharp and clear, as shown in the grey value profile over a porous domain
displayed in Figure 1c. In such cases, a clear boundary determination may be difficult, leading to
possible variations in pore quantification data (“boundary effect”).
This study addresses the abovementioned variability in image processing by conventional thresholding
vs. advanced machine learning methods. The test data set consists of BIB-SEM large-area maps
acquired from 12 Middle Miocene mudstone samples within a single stratigraphic interval in the Vienna
Basin, Austria. Pore space segmentation of SEM maps was processed using the conventional
deterministic thresholding method in the software imageJ (Schindelin et al., 2012, 2015) and the
machine learning-based pixel classification in the open-source tool kit ilastik (Berg et al., 2019; Sommer
et al., 2011). Total porosity, PSDs and pore morphology were resolved by both methods. The study
aims were i) to test the sensitivity of segmentation results to thresholding parameter variations in
ImageJ, and ii) to compare the conventional thresholding method with the considerably faster machine
learning-based identification by ilastik.