AI estimates the ice beneath the world’s glacier maps

IceBoost adds estimated ice thickness to glacier maps. The research was published on 30 July 2026.[1]

Researchers assembled 378,373 gridded measurements from 1,661 glaciers, fewer than 1% of the world’s glaciers. IceBoost averages two decision-tree models trained on terrain, ice velocity and climate information; regional evaluation kept entire glaciers separate between training and testing.[1]

Using Randolph Glacier Inventory v7 outlines, IceBoost estimates 149 ± 38 thousand cubic kilometres of ice for an approximately 2010–2018 reference period. Compared with the Millan and Farinotti models, root-mean-square prediction errors were similar outside polar regions and up to 20–45% lower at high latitudes. Small, steep glaciers remain harder to estimate.[1]

One revealing improvement involved removing inputs tied to how individual glaciers were outlined, including their area and perimeter. Those features could introduce artificial jumps where neighbouring glaciers meet. The revised model produces smoother thickness maps across such boundaries.[1]

A line drawn to organise a map should not become an invented step in the ice beneath it.[1]

Report an error or suggest a correction