Investigating Buried Meltwater Lakes on an Antarctic Ice Shelf with Sentinel-1 SAR Imagery and Machine Learning Methods

Antarctic ice shelves, which encircle approximately 75% of the continent, play a pivotal role in moderating global mean sea level rise as their buttressing properties restrict the flow of inland ice. Each ice shelf is subject to distinct glaciological and climatic conditions that influence its susceptibility to partial break-up or total disintegration. One factor compromising the stability of ice shelves is the presence of both surface and sub-surface meltwater, which may accelerate firn-air depletion and induce flexural stresses, possibly leading to fractures within the ice shelf.

While the occurrence of surface meltwater has been studied extensively in recent years – documenting widespread meltwater systems across several ice shelves during the austral summer – our understanding of meltwater storage below the surface remains limited. In some regions, liquid water may persist within the ice-shelf surface throughout the year, insulated by overlying snow, firn, or ice layers. This subsurface meltwater, particularly in the form of buried lakes, represents a potential mechanism for hydrofracture – even outside the melt season. However, buried lakes are typically difficult to detect using optical imagery, complicating efforts to understand their dynamics and their impact on ice-shelf stability.

Here, we aim to evaluate the feasibility of applying machine learning methods, previously employed on the Greenland Ice Sheet, to detect meltwater lakes buried beneath the surface of Antarctic ice shelves. Using a convolutional neural network in a deep learning approach, we seek to classify ice-shelf surface and subsurface features in Sentinel-1 Synthetic Aperture Radar imagery (SAR), enabling the identification of buried lakes. Preliminary qualitative analysis of Sentinel-1 SAR data has revealed several possible buried meltwater lakes near the grounding line of the western Wilkins Ice Shelf near Merger Island. These lake findings provide an opportunity to assess the applicability of machine learning models developed for Greenlandic application in an Antarctic context. Additionally, it allows us to test the use of airborne radar data for validating buried lake identification in SAR imagery.