Multiphase Image Segmentation of Naturally Fractured Media: Benchmarking Deep Learning and Conventional Approaches

Characterizing subsurface reservoirs, more specifically naturally fractured subsurface reservoir rocks, is essential for the study of subsurface reservoir properties. Image segmentation is an important aspect of digital rock physics (DRP) workflows. Traditional image segmentation techniques are less accurate than deep learning-based segmentation algorithms. In this paper, we investigate the segmentation accuracy of a convolutional neural network U-net and compare it with traditional methods of Watershed and multi-Otsu thresholding for multiphase segmentation of grayscale images from a naturally fractured coal sample. The segmentation target involved multiphase classification of the matrix, fully-filled fractures with minerals, and open fractures. The results reveal that U-net outperformed the others with an Intersection over Union metric of 94.9%, a Dice metric of 97%, and a Recall metric of 97.5%. The results support the importance of multiphase, deep learning-based segmentation techniques to support DRP studies of naturally fractured rocks.