# Kim Bente: Automated Extraction of Basal Features from Radargrams with Contrastive Learning

**URL:** <https://forum.access-hive.org.au/t/kim-bente-automated-extraction-of-basal-features-from-radargrams-with-contrastive-learning/7196>\
**Category:** COSIMA/Cryosphere Workshop 2026 Abstracts\
**Created:** [5 November 2026 21:55 UTC](https://forum.access-hive.org.au/t/kim-bente-automated-extraction-of-basal-features-from-radargrams-with-contrastive-learning/7196 "2026-11-05T21:55:49Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![lachlanswhyborn](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.access-hive.org.au/lachlanswhyborn/32/1837_2.png) [@lachlanswhyborn](https://forum.access-hive.org.au/u/lachlanswhyborn)\
**Post date:** [5 November 2026 21:55 UTC](https://forum.access-hive.org.au/t/kim-bente-automated-extraction-of-basal-features-from-radargrams-with-contrastive-learning/7196/1 "2026-11-05T21:55:49Z")

</div>

Kim Bente: Automated Extraction of Basal Features from Radargrams with Contrastive Learning

The conditions at the base of the ice sheet, where ice and bed meet, can strongly influence how and how fast the ice above flows. For example, when ice at the base is thawed, meltwater can accelerate ice flow, while hard, rough, and dry bedrock can decelerate it. Although characterising such thermal, material, and geometric conditions in the Ice-Bed Interface Zone (IBIZ) is vitally important, notably for improving how basal processes are represented in ice sheet models, direct observations remain limited. However, radargrams, the most widely collected geophysical data of the Antarctic ice sheet’s subsurface, are often processed only to extract ice thickness, leaving much of the information they contain about basal conditions untapped.

To make better use of these rich data archives, we introduce a machine learning method that automatically characterises basal conditions directly from radargrams. Using a technique called contrastive learning, the model learns from ‘IBIZ images’ cropped from radargrams and paired with co-located ice velocities from satellites to extract radioglaciological features that distinguish fast-flowing from slow-flowing regions. Applying the method to airborne radar data from the Totten-Vanderford system in East Antarctica, we present first insights on how such learned numerical representations, i.e., embeddings, encode spatial variability of basal conditions. Furthermore, we show how these embeddings can be combined with insights from sparser seismic or other geophysical data to accelerate data-driven scientific discovery.