Machine Learning-Based Threshold Segmentation Method for Deriving Methane Hydrate Saturation in Synchrotron Radiation CT Images

Natural gas hydrates in marine sediments and permafrost are a promising transition fuel, but effective recovery is hindered by reservoir instability and low production efficiency, largely affected by geomaterial microstructure and hydrate distribution. Synchrotron radiation X-ray computed tomography (SR-XCT) has been used to characterize these parameters. This technique enables high-resolution imaging of hydrate-bearing sediments, but the similar X-ray attenuation coefficients of hydrates and water hinder their distinction. To address this, we proposed a machine learning approach combining convolutional neural networks (CNNs) with K-Means clustering for threshold segmentation. Leighton Buzzard sand samples (porosity: 35%) were imaged during methane hydrate formation using SR-XCT at the TOMCAT beamline, Swiss Light Source. A CNN-based edge detection framework employing the Sobel operator first identified grayscale boundaries of solid particles and methane gas. The K-Means algorithm (K = 5, validated by the elbow method) segmented edge and nonedge regions, isolating grayscale values specific to water and hydrates. After removing particle and bubble regions, min-max normalization generated probability density diagrams of the remaining phases. Temporal analysis revealed a leftward peak shift in normalized distributions, indicating hydrate formation within the grayscale range of 0–0.18 on the maximum-minimum normalized interval. Validation against 366 manually annotated data sets confirmed high classification accuracy. This method significantly improves segmentation performance, providing a more reliable quantification of hydrate saturation and pore-scale evolution during formation and dissociation, thereby supporting the assessment of reservoir behavior and production efficiency.