Interactive Machine Learning Improves Accuracy of Coal Porosity Segmentation in Focused Ion Beam-Scanning Electron Microscopy Images

Accurate characterization of the pore structure of coal is critical for predicting its behavior in applications, such as CO₂ storage and coal bed methane production. Focused ion beam scanning electron microscopy (FIB-SEM) is an ultra-high resolution imaging technique that is widely used to visualize the internal 3D structure of coal samples at the microscale and nanoscale. To reconstruct the pore structure, image segmentation algorithms are used to identify pores in the FIB-SEM images. However, traditional grayscale threshold segmentation can lead to significant errors due to image artifacts, including the pore-back effect, curtaining, and charging. In this study, we present a novel workflow for improving the accuracy of pore segmentation in coal that applies interactive machine learning software to FIB-SEM images. We analyzed a FIB-SEM data set for an anthracite coal sample with a vitrinite reflectance of 2.7%. We trained machine learning software (Ilastik) by manually labeling the pores in 1.25% of the 1200 images. We selected three features (intensity, edge, and texture) on different scales to train the classifier. Once trained, we then applied the classifier to the remaining images. Our results showed that the machine learning method segments the images far more accurately than grayscale segmentation: the comprehensive segmentation quality (F₁ score) of the machine learning approach (0.89-0.94) is approximately twice that of the grayscale threshold method (0.32-0.55). Crucially, relative to the machine learning method, grayscale threshold segmentation leads to estimates for porosity that are lower by up to 73%. Our results highlight the potential of using machine learning techniques to more reliably characterize the pore structure of coal and other geological materials.