InterPore2026

Improving Ground Ice Segmentation in Permafrost Cores Using X-ray CT

22 May 2026, 10:20

1h 30m

Poster Presentation (MS01) Porous Media for a Green World: Energy & Climate Poster

Speaker

Mahya Roustaei (Research associate)

Description

Ground ice strongly controls how permafrost responds to warming, influencing thaw settlement, thermokarst development and drainage changes. For predicting thaw settlement and designing resilient infrastructure to expected climate conditions, ice content estimates must be accurate and comparable across cores and sites. X-ray Computed Tomography (CT) is a practical non-destructive tool for measuring ice distribution, but the standard practice of segmenting ice using fixed Hounsfield Unit (HU) thresholds often fails in heterogeneous permafrost because sediment, organic matter, and ice can overlap in apparent density and mixed voxels are common. These effects can bias inferred ice volumes and, in turn, assessments of thaw vulnerability.

We evaluate how segmentation choices affect ice quantification using a 164 cm long permafrost core from a Yedoma upland in north-eastern Siberia spanning variable cryostructures and sediment compositions. We compare (i) conventional HU thresholding, (ii) automated thresholding methods (including Otsu and adaptive histogram-based approaches), and (iii) machine-learning models (random forests and convolutional neural networks) that incorporate texture and morphological context in addition to intensity. CT-derived ice content and bulk density estimates are validated against independent laboratory measurements to quantify bias and uncertainty across core intervals rather than relying on visual agreement alone.

Results show that no single method is robust for all materials. Threshold-based workflows can perform adequately in simpler intervals but become unstable where partial-volume effects and phase overlap are strong. Automated and learning-based approaches reduce some of these errors, but their performance depends on parameter choices, training data, and transferability between contrasting textures. We summarize strengths and limitations across cryostructures and provide guidance for selecting segmentation workflows when the end use is climate- and hazard-relevant ice quantification. The study supports standardized, non-destructive CT-derived datasets needed for comparing permafrost cores and improving projections of thaw impacts in rapidly changing Arctic regions.