Advancing Satellite Radar Altimetry for Ice Sheet Monitoring using High Resolution Digital Elevation Models and Deep Learning

Phillips, Joe and McMillan, Mal (2026) Advancing Satellite Radar Altimetry for Ice Sheet Monitoring using High Resolution Digital Elevation Models and Deep Learning. PhD thesis, Lancaster University.

Text (2026PhillipsPhD)
2026phillipsphd.pdf - Published Version
Available under License Creative Commons Attribution.
Download (57MB)

Abstract

Satellite radar altimeters have provided a near-continuous record of polar ice sheets since 1991, delivering new understanding of ice sheet contributions to sea level rise. The complex topography of the ice margins, however, presents several deep-seated challenges; namely, difficulty in tracking the ice surface, and fundamental ambiguity in determining the origin of surface reflections. This thesis develops methods to better understand and overcome these deficiencies, through the use of high-resolution Digital Elevation Models and novel deep learning methodologies. In doing so, it aims to characterise key limitations in conventional processing methods, and demonstrate how new data-driven approaches can better exploit the information encoded within radar waveforms. Three interconnected objectives were pursued: assessment of current operational systems, development of novel methods for topographic characterisation, and creation of probabilistic frameworks that embrace waveform ambiguity as an information source rather than a processing obstacle. First, comprehensive evaluation of Sentinel-3 SAR altimetry over Antarctica using high-resolution topographic datasets revealed systematic breakdown of core assumptions in complex terrain. Novel Singular Value Decomposition methods created continent-wide slope and roughness datasets, establishing quantifiable relationships between topography and instrument performance. Second, a probabilistic deep learning framework was developed for CryoSat-2, using ensemble models trained with quantile regression to predict full elevation distributions across the altimeter swath from power waveforms alone. The framework demonstrates robust performance whilst explicitly quantifying both fundamental physical ambiguity and model confidence. Third, the practical application of this new probabilistic approach was investigated, by using the framework to generate swath predictions of ice sheet elevation change. The framework successfully reproduced known elevation change trends, and was benchmarked against a suite of other datasets. These advances demonstrate that new machine learning methods and high-resolution datasets can address longstanding limitations in SAR altimetry, impacting past, current, and future missions and, ultimately, our understanding of ice sheet change.