Inlay

Dr Joe Phillips (Lancaster University) will present this science as part of Session CR6.5 on Friday, 08 May, 14:45–14:55 (CEST) in Room L2. Radar altimetry satellites can measure the elevation of ice sheets by firing radio waves at the surface and timing how long the echo takes to return. However, with only a single antenna, these systems cannot tell exactly where on the surface each echo originated from. Current approaches work around this by making simplifying assumptions that reduce each echo to a single elevation estimate, discarding most of the information the waveform contains. This work takes a fundamentally different approach. Rather than throwing away that ambiguity, a probabilistic deep learning framework was trained to extract the full range of plausible surface elevations encoded within each echo. An ensemble of 16 deep learning models was trained on 600,000 radar echoes collected by CryoSat-2 over Antarctica between 2012 and 2021, using the Reference Elevation Model of Antarctica (REMA) as ground truth. The framework was tested over Pine Island Glacier – a region kept entirely separate from training – where it successfully reproduced well-established patterns of ice thinning of 2–3 metres per year. Encouragingly, results closely matched those from CryoSat-2’s interferometric products, which rely on additional information from a second antenna that many satellites do not carry. This matters because elevation change underpins almost everything we calculate about ice sheets: how much ice is being lost, how much seas are rising, and how reliable our future projections are. Extracting more information from each satellite echo – including from historical missions and future satellites that lack a second antenna – could meaningfully improve all of these estimates. Find out more by reading the abstract and attending his presentation online or in-person at EGU26. Feature image credit: ESA Header image credit: Professor Alison Banwell