We present a global dataset of glacier ice thickness modeled with IceBoost v2.0, a gradient-boosted decision tree scheme trained on 7 million ice thickness measurements and informed by physical and geometrical predictors. We model the distributed ice thickness for every glacier in the two latest Randolph Glacier Inventory releases (RGI v6.0 and v7.0), totaling 215,547 and 274,531 glacier outlines, respectively, plus 955 ice masses contiguous with the Greenland Ice Sheet. IceBoost v2.0 represents the third existing RGI v6.0 global ice volume estimate, and the first for RGI v7.0. On RGI v6.0 we find a global glacier volume of (150 ± 38) × 103 km3, consistent with the two previous estimates of (141 ± 40) × 103 km3 and (158 ± 41) × 103 km3. The corresponding sea-level equivalent (SLE), 323 ± 91 mm, is likewise consistent with the two earlier values of 311 ± 100 mm and 324 ± 84 mm. On RGI v7.0 we find a global glacier volume of (149 ± 38) × 103 km3 and SLE of 323 ± 91 mm. Reconstructed ice thickness distributions can vary substantially across models for individual glaciers, ice caps, and even large glacier complexes. Compared to measurements, IceBoost v2.0 root mean square error is 20–45% lower than that of other models in the high Arctic, and comparable elsewhere. We examine major glaciated regions and compare results with the other models. Confidence in our estimates is highest at high latitudes, where abundant training data adequately sample the feature space. Over steep and mountainous terrain, small glaciers, and lower-latitude regions with limited training data, confidence is lower. IceBoost v2.0 is applicable to ice sheet margins. On the Geikie Plateau (East Greenland), we find nearly twice as much ice as previously reported, highlighting the potential for improved constraints on bed topography in this region. No physical laws are explicitly imposed during training, so sufficient and high-quality training data are crucial. The quality of the generated maps depends on the accuracy of the training data, the Digital Elevation Model, ice velocity fields, and glacier geometries, including nunataks. Using the Jensen Gap, we probe the model’s curvature with respect to input errors and find it is strongly concave over low-slope, thick-ice regions, implying a potential downward bias in predicted thickness under input uncertainty. The released dataset can be used to model future glacier evolution and sea-level rise, inform the design of glaciological surveys and field campaigns, as well as guide policies on freshwater management.

Machine-learned global glacier ice volumes

期刊论文
出版模式：

开放获取
发表日期：

2026年7月30日

查看全文

Niccolò Maffezzoli,
Eric Rignot,
Carlo Barbante,
Mathieu Morlighem,
Troels C. Petersen,
Sebastiano Vascon

摘要

We present a global dataset of glacier ice thickness modeled with IceBoost v2.0, a gradient-boosted decision tree scheme trained on 7 million ice thickness measurements and informed by physical and geometrical predictors. We model the distributed ice thickness for every glacier in the two latest Randolph Glacier Inventory releases (RGI v6.0 and v7.0), totaling 215,547 and 274,531 glacier outlines, respectively, plus 955 ice masses contiguous with the Greenland Ice Sheet. IceBoost v2.0 represents the third existing RGI v6.0 global ice volume estimate, and the first for RGI v7.0. On RGI v6.0 we find a global glacier volume of (150 ± 38) × 10³ km³, consistent with the two previous estimates of (141 ± 40) × 10³ km³ and (158 ± 41) × 10³ km³. The corresponding sea-level equivalent (SLE), 323 ± 91 mm, is likewise consistent with the two earlier values of 311 ± 100 mm and 324 ± 84 mm. On RGI v7.0 we find a global glacier volume of (149 ± 38) × 10³ km³ and SLE of 323 ± 91 mm. Reconstructed ice thickness distributions can vary substantially across models for individual glaciers, ice caps, and even large glacier complexes. Compared to measurements, IceBoost v2.0 root mean square error is 20–45% lower than that of other models in the high Arctic, and comparable elsewhere. We examine major glaciated regions and compare results with the other models. Confidence in our estimates is highest at high latitudes, where abundant training data adequately sample the feature space. Over steep and mountainous terrain, small glaciers, and lower-latitude regions with limited training data, confidence is lower. IceBoost v2.0 is applicable to ice sheet margins. On the Geikie Plateau (East Greenland), we find nearly twice as much ice as previously reported, highlighting the potential for improved constraints on bed topography in this region. No physical laws are explicitly imposed during training, so sufficient and high-quality training data are crucial. The quality of the generated maps depends on the accuracy of the training data, the Digital Elevation Model, ice velocity fields, and glacier geometries, including nunataks. Using the Jensen Gap, we probe the model’s curvature with respect to input errors and find it is strongly concave over low-slope, thick-ice regions, implying a potential downward bias in predicted thickness under input uncertainty. The released dataset can be used to model future glacier evolution and sea-level rise, inform the design of glaciological surveys and field campaigns, as well as guide policies on freshwater management.