IceBoost v2.0, the AI ​​model for mapping the thickness of Earth’s glaciers

Devon Ice Sheet (Northern Canadian Arctic)

The model made it possible to reconstruct the distribution and thickness of each glacier included in the Randolph Glacier Inventory, excluding the polar ice caps of Antarctica and Greenland, and to make the data freely accessible through an interactive web-app accessible at https://nmaffe.github.io/iceboost_webapp/ for the exploration of global parameters. According to the data that emerged, the planet’s glaciers retain a total volume estimated at around 150 thousand cubic kilometers of ice, corresponding to a potential rise in average sea levels of 32.3 centimeters in the event of complete melting. Compared to previous mappings, IceBoost v2.0 offers up to 40% greater accuracy in adhering to field surveys; on the Geikie Plateau, in eastern Greenland, where the thickness reaches 2 kilometers, the new estimate found a volume almost double that of previous measurements.

On the methodological relevance of these parameters for climate modeling, Niccolò Maffezzoli explained: “The distribution of ice thickness is a fundamental variable for glaciological and climate models. To predict how glaciers will evolve by 2100 and their contribution to sea level rise, it is necessary to know their current state in as much detail as possible. Researchers from the Glacier Model Intercomparison Project (GlacierMIP4) who are working on the next simulations to inform the IPCC on the evolution of glaciers by 2100 will use IceBoost v2.0 as the only representation of the current situation”.

Svalbard archipelago

The dataset also identifies geographical areas with higher margins of uncertainty that require further survey campaigns, including the Himalayas, the Karakoram and the ice fields of Patagonia, providing useful indicators also for the management of freshwater reserves on which approximately 1.9 billion people depend globally. Analyzing the complexity of glacial systems and research prospects, Maffezzoli concluded: “Glaciers are complex systems. The physics that describes them is known. However, many parameters that enter the equations are unknown or the variables are difficult to measure. Machine learning-based models offer an alternative approach. They learn from data, making predictions without an a priori imposed physical description. Often, if we have enough data, this approach is successful. In the future we will move towards the development of hybrid models, in which physical description and learning from experimental measurements will work together to produce more accurate estimates. We must hurry, the glaciers in the middle latitudes, including the Alps, will disappear within a few decadesThe study was funded through a Marie Skłodowska-Curie Actions fellowship of the Horizon Europe program within the SKYNET project, in collaboration with the University of California Irvine, NASA’s Jet Propulsion Laboratory, Dartmouth College and the University of Copenhagen.