Reconstruction of three-dimensional fractures in coal based on multidirection slicing and U-Net segmentation

Accurate characterization of multi-size fractures in coal is crucial for estimating its transport properties. However, the extraction of narrow microfractures in 3D voxel-type CT images is difficult, which causes the loss of connectivity in the extracted fracture network and reduces the accuracy of the predicted transport properties. Improving the image quality and optimizing the segmentation process to deal with the inaccuracy of fracture extraction from coal CT images. In this study, the super-resolution image reconstruction and image noise filtering, were used to double the CT image size and enhance the recognition of narrow microfractures. The 3D image data were then sliced normal to the center vertical axis and around the center vertical axis to create two datasets. A convolution network, U-Net, was used to segment the two datasets, and a 3D binary fracture model was reconstructed by combining the segmented results. The permeability of the model was calculated using the lattice Boltzmann method. Compared with other widely used segmentation methods, multidirectional slicing and U-Net segmentation significantly improved the recognition of narrow microfractures. The segmented model had a higher effective porosity and permeability, indicating the good connectivity of the fracture network. These results suggest that the accurate extraction of narrow microfractures is critical to improve the estimation of fluid transport in a broad range of subsurface applications.