A comprehensive assessment of image processing variability in pore structural investigations: Conventional thresholding vs. machine learning approaches

Abstract

Broad ion beam - scanning electron microscopy (BIB-SEM) is an important technology for quantitative description of pore space and detailed investigation of pore structure in mudstones and shales. To obtain quantitative data from BIB-SEM images, image processing is required. However, the effect of different image processing approaches on the segmentation results has not been well documented yet. Particularly the interpretation of pore boundaries approaching the maximum image resolution can represent a major challenge. This study addresses the variability in image processing by comparing conventional thresholding with advanced machine learning methods and analysing the sensitivity of results to different segmentation. The test dataset consists of large area BIB-SEM maps of 12 Middle Miocene mudstone samples from the Vienna Basin, Austria. Segmentation of pore space and organic matter was performed using the thresholding method in ImageJ and the machine learning-based pixel classification in ilastik. The results show that the selected greyscale thresholding algorithm in ImageJ significantly affected the resolved porosity (1.7% to 2.8% variation) relative to the average total porosity of 4.5%. Compared to ImageJ, ilastik provides comparable results with better efficiency in separating pores from organic matter. However, the obtained pore information is inevitably influenced by variations in boundary delineation and feature identification due to inherent differences in the image processing methods and varying representativeness of image maps themselves. Overall, the study findings highlight the importance of carefully selecting image processing methods and considering their potential impact on the accuracy and reliability of pore characterization in mudstones and shales.