Self-Supervised Learning Using Vision Transformer Architecture for Rock Image Segmentation

Abstract The segmentation of microstructural features in Scanning Electron Microscopy (SEM) images of shale samples is critical for petrophysical analyses, including mineralogy quantification, pore network analysis, and pore system identification. However, processing these images efficiently and accurately typically requires supervised deep learning-based methods, such as semantic segmentation algorithms. Semantic segmentation classifies each pixel in an image into a predefined category (e.g., organic material, inorganic material, pore), regardless of the number of times that category appears in the image. This type of segmentation enables the detailed quantification of microstructural features in shales, but supervised methods require large, manually annotated datasets. The creation of these datasets, particularly for SEM images, is resource-intensive, both in terms of time and cost. Self-supervised semantic segmentation, on the other hand, learns useful features from the image itself without needing manual annotation, significantly reducing the time and cost involved in dataset preparation. The specific algorithm in this work is based on a Vision Transformer architecture. Our dataset comprises of FIB-SEM images from 22 shale plays across North and South America. We first begin with careful image augmentation, which is especially critical for self-supervised semantic segmentation algorithms, to avoid generating trivial solutions while also enabling the algorithm to focus on global details as well as local variations in an image. In this work, our data augmentation tasks include generating random crops of each image and enhancing the contrast. This is followed by training a self-supervised segmentation algorithm on a subset of images from our augmented dataset. Finally, we use the trained model to segment the image into 12 sub-classes which are subsequently grouped into the primary classes of organic, and inorganic material and pores. This approach enables the model to capture subtle differences in grayscale shades within each primary class, resulting in more refined image segmentation. We evaluate the self-supervised model across several complex scenarios to test its accuracy and robustness and show that the model reliably segments organic, inorganic, and pore regions in these images, allowing for large-scale analyses of shale images and eliminating the dependence on large, expensively annotated datasets. However, while our approach is promising, we also document instances with poor segmentation performance, which occurs in about 5% of the images we test. Nevertheless, considering the rapidity and fidelity of this approach, especially for instances where labeled data is scarce and expensive to acquire, self-supervised segmentation has the potential to streamline the analyses of microstructural features in shales, making it a valuable tool for subsequent petrophysical and geological applications.