Can deep learning combined with satellite altimetry enable accurate mapping of sea ice thickness and snow depth from Sentinel-1 SAR at high resolution, and are the resulting uncertainty estimates robust enough for operational applications?

Andreas is a postdoctoral researcher at DTU Space, Technical University of Denmark. He studied electrical engineering at DTU, specialising in space technology, and completed his PhD at DTU Space in collaboration with ESA, developing deep learning methods to map Arctic sea ice in satellite radar imagery. He led the AutoICE Challenge, an international competition on automated sea ice charting from satellite data.Andreas’s research focuses on applying AI to Earth observation for the cryosphere, combining satellite radar imagery with altimetry to monitor Arctic sea ice. He took part in the SWIDA-RINGS circumnavigation of Antarctica in 2024/25, an airborne survey of the continent’s grounding line. He also investigates event-based cameras for geophysical and space applications, including observations of lightning above thunderstorms from the International Space Station.

Research abstract and objectives

Satellite altimeters such as ESA’s CryoSat-2 and Sentinel-3, and NASA’s ICESat-2 can measure sea ice thickness, but only along narrow tracks, giving complete Arctic coverage roughly once a month at tens of kilometers resolution. Sentinel-1 radar images cover up to 400 km, in darkness and through cloud, but cannot measure thickness directly.
AI4S1SIT will train deep learning models on tens of thousands of scenes where the altimeters and Sentinel-1 observe the same ice, learning the relationship between the model “sees” in radar image and what the altimeter measured. Applied across full Sentinel-1 scenes, this transfers the altimeters’ measuring capability to the imagery, producing sea ice thickness maps at sub-100 meter resolution with up to weekly coverage.

Scientific questions addressed by the project:

Is it possible to map sea ice thickness accurately from Sentinel-1 SAR imagery using deep learning trained on satellite altimetry? Can snow depth on sea ice be derived in the same radar imagery at the same fine resolution, by combining radar and laser freeboard? And how reliable are the resulting AI-derived products – can the uncertainty quantified well enough for operational use?