Machine learning can improve ice sheet bed mapping

What the study found: The authors report that machine learning (ML) can enhance the value of airborne radio-echo sounding (RES) data for mapping the bed beneath ice sheets. They identify two main uses already in practice: denoising and automated picking of radar returns, and improved spatial interpolation and uncertainty quantification of flightline data.

Why the authors say this matters: The authors conclude that using ML throughout the workflow may help maximize the observational value of past RES campaigns and guide future strategic efforts in the Polar Regions. They also state that this is relevant because ice sheet thickness measurements underpin models used to improve projections of sea-level rise.

What the researchers tested: This is an overview article on recent ML advances relevant to ice sheet RES research. The authors present examples from the Antarctic and Greenland Ice Sheets and discuss ML applications in denoising, automated radar-return picking, interpolation, uncertainty quantification, and planning future surveys.

What worked and what didn't: The article says ML-driven approaches can outperform traditional methods for interpolation of basal topography, and that progress has been made in automated extraction of reflecting horizons from radargrams. It also notes two areas where ML may have a role in planning future surveys, but the abstract does not specify which areas those are.

What to keep in mind: The abstract does not provide detailed performance metrics, sample sizes, or comparisons beyond the general statement that ML-driven interpolation can outperform traditional methods. It is also an overview article, so the summary of evidence is limited to the examples and advances the authors chose to highlight.

Key points

• Machine learning can enhance the value of airborne radio-echo sounding data for ice sheet bed mapping.
• The authors highlight two current ML uses: radar-data denoising/automated picking and better interpolation with uncertainty quantification.
• Examples from Antarctica and Greenland show ML-driven interpolation can outperform traditional methods for basal topography.
• The article says progress has been made in automated extraction of reflecting horizons from radargrams.
• The abstract suggests ML may also help plan future surveys, but it does not specify how.