Item

Abstract

Ice cores are among the most important natural archives that can provide valuable information from the past environment of our planet. The signals related to the history of the earth are preserved in structure, bubbles, water isotopes, and impurities, which can be retraced by studying polar ice cores. Although recent technological advancements have made it possible to perform non-destructive tests such as micro-CT scans to study structure and bubbles, performing a high-quality automated segmentation of ice cores from different regions and depths is still a challenge. The CT images of various depths have different pixel intensities, and they might appear with a range of noise, artifacts, and beam hardening issues. The traditional segmentation methods, such as thresholding and edge finding is tedious to be applied on all sort of different ice core CT images, thus, we took advantage of deep learning algorithms to facilitate this task. Besides the image noise diversities, scanning with high resolution is very time-consuming and makes it impossible to scan a full column. Thus implementing A.I. is suggested to have high-quality segmentation for all sorts of ice cores. At first, the ground truth was provided by weak unsupervised learners such as the Gaussian mixture model and manual checking from high-resolution images, then a supervised model (U-net) was developed having low-resolution as input and downsampled high-resolution as ground truth. For this purpose, three ice core specimens from different locations (Greenland and Antarctica) and various depths (snow, firn, bubbly ice) were scanned with 60 and 30-micrometer resolutions.​