Research on CT image segmentation and identification methods for micro/nanoscale pores and fractures in sandstone

Accurate characterization of micro/nanoscale pores and fractures in tight sandstone is fundamental for revealing the mechanisms of multiphase interactions within rocks, and it provides support for hydrocarbon exploitation, geothermal development, and geohazard assessment. Nevertheless, conventional image segmentation techniques suffer from obvious accuracy limitations when handling intricate pore-fracture structures. Moreover, few studies have systematically evaluated how segmentation algorithms affect subsequent three-dimensional pore reconstruction and seepage simulation results.

In this study, micro-focus X-ray computed tomography (μCT) was adopted to build a self-established CT dataset containing micro/nanoscale pores and fractures of tight sandstone. Two conventional segmentation approaches—grayscale thresholding and the watershed algorithm—were employed as baseline comparisons. Four convolutional-neural-network-based deep learning models including U-Net, SegNet, DeepLabv3-ResNet50, and DeepLabv3-ResNet101 were implemented for pore-fracture segmentation. A publicly available micrometer-scale sandstone dataset was further utilized to conduct cross-scale transfer validation. Model performance was comprehensively compared and evaluated in terms of two-dimensional segmentation accuracy, three-dimensional pore network model (PNM) reconstruction, and single-phase seepage simulation.

The experimental results showed that deep learning models outperformed traditional segmentation methods in general. Among all tested models, the U-Net achieved optimal comprehensive performance. Compared with the best results obtained using conventional segmentation methods, U-Net increased the intersection over union (IoU) and F1-score by 18.70% and 16.47%, respectively, with an overall accuracy of 99.03%. The corresponding standard deviations of these three indicators were 0.012, 0.010, and 0.004, demonstrating high stability and robustness. U-Net could effectively preserve the boundary integrity and fine-scale continuity of nanoscale pores and fractures. Its three-dimensional reconstructed porosity reached 2.408%, which was close to the original porosity of 2.785%, and the pore-throat parameters of the reconstructed pore network model (PNM) exhibited good consistency with the actual sample. It was noteworthy that data augmentation operations could destroy the inherent topological features of natural fractures and reduce model segmentation accuracy. The DeepLabv3 series models suffered from obvious over-segmentation artifacts, while traditional algorithms tended to produce missed detection of tiny pores or spurious pore connectivity. These defects further induced non-negligible deviations in permeability calculation and pressure-field distribution during seepage simulation.

Compared with traditional segmentation approaches, the pore-network topology derived from U-Net segmentation matches well with real pore-fracture structures. The U-Net model substantially improves the quantitative analysis accuracy of porosity, pore-throat geometry, and permeability properties. This study provides a feasible reference for automatic identification and digital modeling of micro/nanoscale pore-fracture systems within tight sandstone.