Improving Ice Segmentation in Permafrost Cores using Computed Tomography

The segmentation of ice in X-ray Computed Tomography (CT) scans of permafrost samples has traditionally relied on the Hounsfield Unit (HU) thresholding approach while their accuracy is often limited by overlapping density ranges in complex and heterogeneous samples. Recent advances, including automated thresholding algorithms and machine learning techniques, offer improved precision by leveraging texture, contrast, and morphological features in CT images. This study investigates the evolution of ice segmentation methodologies by applying multiple approaches to a 164 cm long permafrost core drilled from a Yedoma upland in north-eastern Siberia. The core was analyzed using traditional HU thresholding, automated thresholding methods (e.g., Otsu and adaptive histogram-based segmentation), and machine learning models (e.g., random forests and convolutional neural networks). The results from CT scans and segmentation methods were validated and compared against laboratory measurements of ice content and density, ensuring a robust evaluation of each technique's accuracy and reliability.

The results provide critical insights into the strengths, weaknesses, and suitability of different segmentation methods for permafrost cores. These findings contribute to the development of standardized, high-precision methodologies for non-destructive characterization of ice-rich soils, supporting geotechnical and climate change studies in permafrost regions.