Extracting Swath Elevation Information from Non-Interferometric Radar Altimetry using Probabilistic Deep Learning

Satellite radar altimetry provides three decades of continuous ice sheet observations, yet non-interferometric systems face a fundamental constraint: each waveform records only the return time of reflections, not their origin within the beam footprint. As a result, current approaches - which have remained largely unchanged for 30 years - rely on simplifying assumptions and fail to extract the full information content from the recorded echoes. Interferometric processing resolves this by measuring arrival angle with two antennas, enabling swath processing with multiple across-track elevation points. However, historical missions (ERS-1/2, Envisat), current operations (Sentinel-3), and planned systems (CRISTAL Ka-band) lack this capability.

Conventional processing instead eliminates ambiguity through dimensionality reduction: retracking identifies a single range from the leading edge, while slope correction locates the point of closest approach (POCA) to attribute this to. This approach discards most of the waveform information, extracting only singular elevation estimates.

Here, we present a fundamentally alternative approach to processing radar altimetry echoes over ice sheets using probabilistic deep learning to extract the distribution of plausible surface elevations encoded within each waveform. Rather than reducing ambiguity through simplifying assumptions, we treat the full waveform as exploitable information, directly modelling the distribution of surface elevations consistent with each measurement.

Specifically, we train an ensemble of 16 ResNet-RS models on 600,000 CryoSat-2 SARIn power waveforms (2012–2022) over Antarctica using the Reference Elevation Model of Antarctica (REMA) as ground truth. The framework predicts 5th, 50th, and 95th elevation quantiles across 150 points spanning the 15 km swath, providing 150 elevation estimates where conventional methods yield one. Importantly, our new probabilistic approach allows us to directly quantify both aleatoric (surface ambiguity) and epistemic (model confidence) uncertainty.

Once trained, we validated the models against unseen ice-sheet-wide REMA data (2023), demonstrating robust prediction of swath elevation distributions. Prediction interval coverage (5th–95th) averaged ~80% (10 points below the 90% nominal target) ice-sheet-wide, with consistent performance achieved with respect to ICESat-2 over Pine Island Glacier (2019–2025) - a region entirely withheld from training. The upper 95th percentile exhibited strong stability across all conditions, reflecting its physical anchoring in leading-edge returns, while lower quantiles showed systematic under-coverage increasing with topographic complexity.

Practical utility of this new approach was demonstrated through elevation change monitoring over Pine Island Glacier, where our deep learning framework reproduced established spatial thinning patterns of 2-3 m yr^-1 with temporal elevation change differences of -0.03 ± 0.11 m yr^-1 relative to ICESat-2. Despite using only power waveforms, performance matched CryoSat-2's interferometric POCA and EOLIS products - detecting sub-metre annual changes from probability distributions spanning tens of metres.

This represents the first swath processing from non-interferometric altimetry using power waveforms alone. By reframing waveform ambiguity as quantifiable distributional information, rather than a processing limitation, this deep learning approach demonstrates how machine learning can fundamentally rethink conventional altimetry processing, establishing new capabilities for ice sheet monitoring across past, present, and future missions.