Coal Cleat/Fracture Segmentation Using Convolutional Neural Networks

Cleats/fractures play a vital role in gas extraction from coal bed methane reservoirs. As such, fractures are identified in microcomputed tomography images by segmentation in detailed fracture characterization studies. Conventional segmentation methods include thresholding-based, region growing, and hybrid methods. In digital images, these methods are unable to differentiate fractures from other coal constituents with a similar intensity on the gray scale. However, a supervised segmentation is not a wise approach, since it requires a considerable amount of time. Recently, machine learning-more specifically, the convolutional neural networks (CNN)-has demonstrated excellent performance in machine vision applications. This study uses CNN-based methods to segment cleats/fractures of two types of coal samples: homogenous (with typical cleat system) and heterogeneous (with complex background and irregular fracture system). Two CNNs (2D and 3D) with similar architectures are applied to classify the central pixel (or voxel) of the input image as either fracture or nonfracture classes. In the case of the homogenous coal, the existing fractures in one 2D image are manually segmented and used for the training, validation, and test dataset. Results indicate that the 3D CNN is more robust and accurate than the 2D CNN, as the former extracts the features in 3D space. For the complex and heterogeneous sample, five 2D images are segmented for a more representative training dataset. The extracted fracture system using the 3D CNN displays an accuracy of 96.7%. A comparison with the conventional methods in both cases showed that the CNN-based approach not only detects the fracture system more efficiently but also generates fractures with a correct aperture size. These results reveal that CNN-based methods successfully discern fractures from other constituents of coal with similar grayscale intensity.