Monitoring supraglacial hydrology on the Greenland Ice Sheet using remote sensing and machine learning

Glen, Emily and Leeson, Amber and McMillan, Mal and Maddalena, Jennifer (2025) Monitoring supraglacial hydrology on the Greenland Ice Sheet using remote sensing and machine learning. PhD thesis, Lancaster University.

Text (2025GlenPhD)
2025GlenPhD.pdf - Published Version
Available under License Creative Commons Attribution.
Download (72MB)

Abstract

The Greenland Ice Sheet (GrIS) is experiencing accelerating surface melt rates, with supraglacial meltwater influencing ice dynamics, mass balance, and sea level rise. Yet key gaps remain in understanding the spatial and temporal evolution of supraglacial melt and hydrological features, especially for understudied features like slush. While traditional remote sensing methods are widely used, they often struggle to capture the complexity of these systems. Machine learning (ML) offers scalable, automated solutions suitable for large-scale or near-real-time monitoring, yet current applications remain limited by challenges in accuracy, generalisation, and training data availability. This thesis develops and applies satellite remote sensing and ML methodologies to improve the detection and mapping of supraglacial hydrology across the GrIS. In doing so, this thesis also produces new datasets that advance understanding of supraglacial meltwater distribution at ice-sheet scale, including detailed assessments of regional, interannual, and seasonal variability, with a focus on contrasting low and high melt years. First, a comparative study of meltwater features in southwest Greenland during a low-melt year (2018) and a high-melt year (2019), using optical satellite imagery and threshold-based classification algorithms individually developed to delineate different components of the supraglacial hydrological system (i.e., supraglacial lakes, channels, and slush), reveals a substantial increase in meltwater extent, connectivity, and elevation in 2019. Slush emerges as a dominant yet previously under-recognized component of the supraglacial hydrological system. Second, a near-decadal, ice-sheet-wide analysis of slush from 2016 to 2024 using Sentinel-2 optical imagery and ML classification demonstrates that slush – underrepresented in existing meltwater inventories – is widespread, but highly variable, across regions and years. Third, to evaluate and optimise large-scale, cloud-based meltwater mapping, a systematic assessment of seven ML classifiers within Google Earth Engine identifies Random Forest as the most transferable across space and time, while Gradient-Boosted Decision Trees achieves the highest overall accuracy but are more sensitive to mislabelled training data. Together, the findings in this thesis advance supraglacial hydrology monitoring by (1) establishing slush as an important component of the meltwater system, (2) revealing pronounced interannual variability and climate sensitivity in meltwater dynamics, (3) promoting a more holistic view of supraglacial hydrology as a continuum of interconnected features, and (4) evaluating the utility and scalability of various cloud-based ML approaches for large-scale meltwater mapping. These contributions enhance our understanding of meltwater distribution and dynamics on the GrIS and improve our capacity to monitor all components of supraglacial hydrology, including slush, over large spatial and temporal scales. These findings also lay the groundwork for optimizing ML-based classification approaches, with future implications for long-term monitoring and for assessing the impact of supraglacial hydrology on ice sheet stability in a warming climate – not only on the GrIS, but also on the Antarctic Ice Sheet and mountain glaciers worldwide.