Leveraging Deep Learning to Enhance Satellite-Derived Ice Sheet Elevation Measurements
The Greenlandic and Antarctic Ice Sheets are critical indicators of global climate change, with their mass loss significantly contributing to rising sea levels. Obtaining precise, long-term observations of surface elevation changes across these vast ice masses is essential for accurately assessing their role in the Earth’s climate system. Satellite radar altimetry has been a valuable tool for monitoring ice sheet elevations since the early 1990s, providing continuous measurements over the majority of the polar regions. However, one of the most significant challenges in utilizing radar altimetry data for ice sheet research lies in the accurate retrieval of the true surface elevation due to the complex interaction between the radar signal and the snow/ice surface.
The radar signal can penetrate the upper layers of the snowpack, resulting in an inaccurate measurement of the actual surface height. This radar penetration effect introduces a bias in the elevation estimates, which can vary significantly over time due to changes in the physical properties of the snow and firn. Conventional retracking algorithms, such as the Threshold First-Maximum Retracker Algorithm (TFMRA) and the European Space Agency’s (ESA) ICE1 and ICE2 products, have attempted to mitigate this issue by targeting the lower part of the radar waveform’s leading edge. However, residual biases often remain, requiring further post-processing corrections based on waveform parameters like the leading-edge width (LEW) and backscatter.
To address these challenges, researchers at the Alfred Wegener Institute (AWI) have developed a novel approach – the AWI-ICENet1 convolutional neural network (ConvNet) retracker. This data-driven solution aims to directly minimize the effects of time-variable radar penetration, reducing the need for extensive post-processing corrections.
Simulating a Comprehensive Training Dataset
The key to the success of the AWI-ICENet1 retracker lies in the creation of a comprehensive and representative training dataset. Unlike previous machine learning approaches that have relied on limited ground-truth measurements, the AWI team generated a large, synthetic dataset of simulated radar waveforms. This dataset, comprising 3.8 million waveforms, was designed to accurately model the complex interaction between the radar signal and the ice sheet surface, taking into account factors such as:
- Varying surface topography and slopes
- Differences in bulk attenuation rates, which influence the volume scattering component
- Realistic radar system characteristics, including the Ku-band frequency, antenna gain pattern, and range resolution
By incorporating these elements, the simulated waveforms closely resemble the actual CryoSat-2 Low Resolution Mode (LRM) waveforms observed over the Greenland and Antarctic ice sheets. The synthetic dataset serves as a robust reference for training the AWI-ICENet1 ConvNet, enabling the network to learn the underlying relationships between the waveform shape and the true surface elevation.
Advancing Ice Altimetry with ConvNet Retracking
The AWI-ICENet1 ConvNet architecture consists of a series of 1D convolutional layers that process the radar waveform as sequential data, extracting relevant features and ultimately producing an estimate of the retracked range – a proxy for the true surface elevation. This approach differs from previous attempts that utilized 2D ConvNets to process radar imagery, which can be susceptible to issues related to the variable positioning of consecutive waveforms within the radar range window.
The training of the AWI-ICENet1 model was performed using the simulated waveform dataset, with the network optimized to minimize the mean-squared-error between the retracked range and the known reference range. The resulting model exhibits excellent performance, as demonstrated by the low cross-validation errors and the model’s ability to generalize to unseen data.
Validating AWI-ICENet1 against Conventional Retrackers
To assess the effectiveness of the AWI-ICENet1 retracker, the researchers conducted a comprehensive evaluation, comparing its performance to that of other widely used retrackers, including the TFMRA and the ESA’s ICE1 and ICE2 products.
Improved Cross-Point Accuracy
One of the key metrics used to assess the quality of the retracked elevations is the cross-point error (CPE), which measures the difference between elevation estimates at crossover points between successive satellite tracks. The analysis revealed that the AWI-ICENet1 retracker consistently outperforms the other methods, exhibiting significantly lower median and standard deviation of the CPEs across various regions of the Antarctic ice sheet.
The results demonstrate that AWI-ICENet1 is particularly effective in areas with complex topography and higher surface slopes, where conventional retrackers tend to struggle. Furthermore, the new retracker shows a remarkable reduction in the temporal variability of the CPEs, indicating a substantial decrease in the impact of time-variable radar penetration compared to other approaches.
Minimizing Transient Penetration Biases
Another critical aspect of the evaluation focused on the ability of the retrackers to mitigate the effects of transient radar penetration biases, which can introduce significant errors in surface elevation change (SEC) estimates. The researchers analyzed the time-dependent variability in elevation anomalies (Δh) for selected regions in Greenland and Antarctica.
Their findings show that the AWI-ICENet1 retracker substantially reduces the magnitude and temporal variability of the Δh, effectively capturing the true elevation changes due to changes in surface mass balance or firn compaction. In contrast, the ESA ICE2 retracker and, to a lesser extent, the TFMRA and ESA ICE1 retracker exhibit much larger Δh, indicating a stronger sensitivity to changes in the electromagnetic properties of the snowpack.
Improved Ice Sheet Mass Change Estimates
The researchers further assessed the impact of the different retrackers on ice sheet mass change estimates by comparing the results to independent elevation change data from the ICESat-2 laser altimeter mission. For the period from January 2019 to December 2021, the volume change estimates derived from the AWI-ICENet1 retracked CryoSat-2 data show excellent agreement with the ICESat-2 results, with differences of less than 20 km³/yr for both the Greenland and Antarctic ice sheets.
In contrast, the other retrackers, even when combined with empirical corrections based on waveform parameters, exhibit much larger biases and uncertainties in the mass change estimates, highlighting the superior performance of the AWI-ICENet1 approach.
Extending the Reach of AWI-ICENet1
The versatility of the AWI-ICENet1 retracker extends beyond the CryoSat-2 mission, as it can be applied to historical, current, and future satellite radar altimetry missions, such as Envisat, ERS-1, ERS-2, Sentinel-3, and the upcoming Copernicus Polar Ice and Snow Topography Altimeter (CRISTAL) mission. This will enable the creation of a consistent, high-accuracy elevation change product across multiple decades, providing invaluable insights into the long-term behavior of the Greenland and Antarctic ice sheets.
Furthermore, the researchers are exploring the potential of expanding the waveform simulator to include synthetic-aperture radar (SAR) waveforms, which would allow the application of the AWI-ICENet1 approach to the SAR mode of CRISTAL and Sentinel-3, further enhancing the utility of this innovative retracking solution.
Conclusion
The development of the AWI-ICENet1 convolutional neural network retracker represents a significant advancement in the field of satellite radar altimetry for ice sheet monitoring. By leveraging a comprehensive, simulated training dataset and a tailored 1D ConvNet architecture, the researchers have addressed the long-standing challenge of accurately retrieving true surface elevations from radar waveforms, overcoming the limitations of conventional retracking methods.
The superior performance of AWI-ICENet1, as demonstrated by its reduced cross-point errors, minimized transient penetration biases, and improved agreement with independent laser altimetry data, underscores the potential of this data-driven approach to revolutionize our understanding of ice sheet mass balance and its contribution to global sea level rise. As the scientific community continues to explore the impacts of climate change, tools like AWI-ICENet1 will be instrumental in providing the high-quality, long-term observations needed to inform critical policy decisions and advance our knowledge of the cryosphere.
