These different behaviours and resulting biases can potentially induce important consequences in long-term glacier evolution projections. Both models agree around the average values seen during training (i.e. In recent years, shrinking glaciers have contributed to about 30% of global sea level rise 1. GloGEMflow relies on EURO-CORDEX ensembles26, whereas ALPGM uses ADAMONT25, an adjusted version of EURO-CORDEX specifically designed for mountain regions. https://doi.org/10.1016/B978-0-12-821575-3.00009-8. Hock, R. et al. Summer climate is computed between April 1st and September 30th and winter climate between October 1st and March 31st. Three different types of cross validation were performed: a Leave-One-Glacier-Out (LOGO), a Leave-One-Year-Out (LOYO) and a Leave-Some-Years-and-Glaciers-Out (LSYGO). Braithwaite, R. J. For such cases, we assumed that ice dynamics no longer play an important role, and the mass changes were applied equally throughout the glacier. South American Glaciers Melting Faster, Changing Sea Level These trends explored with energy balance models from the literature correspond to the behaviour captured by our deep learning MB model, with a clearly less sensitive response of glacier-wide MB to extreme climate forcings, particularly in summer (Fig. By Carol Rasmussen,NASA's Earth Science News Team. The Nature of Kinematic Waves in Glaciers and their Application to With a secondary role, glacier model uncertainty decreases over time, but it represents the greatest source of uncertainty until the middle of the century8. Paul, F. et al. (b) Climate predictors are based on climatic anomalies computed at the glaciers mean altitude with respect to the 19672015 reference period mean values. This creates a total of 34 input predictors for each year (7 topographical, 3 seasonal climate, and 24 monthly climate predictors). b, c, d and f, g, h annual glacier-wide MB probability distribution functions for all n scenarios in each RCP. Simulations for projections in this study were made by generating an ensemble of 60 cross-validated models based on LSYGO. This experiment enabled the exploration of the response to specific climate forcings of a wide range of glaciers of different topographical characteristics in a wide range of different climatic setups, determined by all meteorological conditions from the years 19672015 (Fig. melt and sublimation of ice, firn and snow; or calving)9; and (2) ice flow dynamics, characterized by the downward movement of ice due to the effects of gravity in the form of deformation of ice and basal sliding. The glacier ice volume in the French Alps at the beginning of the 21st century is unevenly distributed, with the Mont-Blanc massif accounting for about 60% of the total ice volume in the year 2015 (7.06 out of 11.64km3, Fig. contributed to the extraction of nonlinear mass balance responses and to the statistical analysis. Ice caps in the Canadian Arctic, the Russian Arctic, Svalbard, and parts of the periphery of Greenland are major reservoirs of ice, as well as some of the biggest expected contributors to sea level rise outside the two polar ice sheets7. With this cross-validation we determined a deep learning MB model spatiotemporal (LSYGO) RMSE of 0.59m.w.e. Steiner, D., Walter, A. Graphics inspired by Hock and Huss40. This implies that specific climatic differences between massifs can be better captured by ALPGM than GloGEMflow. MB rates only begin to approach equilibrium towards the end of the century under RCP 2.6, for which glaciers could potentially stabilize with the climate in the first decades of the 22nd century depending on their response time (Fig. The authors declare no competing interests. 14, 815829 (2010). "The Patagonia Icefields are dominated by so-called 'calving' glaciers," Rignot said. Marzeion, B. et al. A.R. Bartk, B. et al. The same was done with winter snowfall anomalies, ranging between 1500mm and +1500mm in steps of 100mm, and summer snowfall anomalies, ranging between 1000mm and +1000mm in steps of 100mm. We perform, to the best of our knowledge, the first-ever deep learning (i.e. Publishers note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. The 29 RCP-GCM-RCM combinations available, hereafter named climate members, are representative of future climate trajectories with different concentration levels of greenhouse gases (TableS1). Rev. Lett. ICCV (2015) https://doi.org/10.1109/iccv.2015.123. & Funk, M. A comparison of empirical and physically based glacier surface melt models for long-term simulations of glacier response. S5cf), except for the largest glaciers (e.g. Changes in DDFs with respect to air temperature also strongly depend on albedo, with ice presenting a substantially more nonlinear response than snow. . Moreover these three aspects of glacier behavior are inextricably interwoven: a high sensitivity to climate change goes hand-in-hand with a large natural variability. Some of these models use a single DDF, while others have separate DDFs for snow and ice, producing a piecewise function composed of two linear sub-functions that can partially account for nonlinear MB dynamics depending on the snowpack. Fundam. Evol. PubMedGoogle Scholar. On the other hand, ice caps present a different response to future warming, with our results suggesting a negative MB bias by models using linear PDD and accumulation relationships. The climatic forcing comes from high-resolution climate ensemble projections from 29 combinations of global climate models (GCMs) and regional climate models (RCMs) adjusted for mountain regions for three Representative Concentration Pathway (RCP) scenarios: 2.6, 4.5, and 8.525. A sensitivity analysis of both MB models revealed nonlinear relationships between PDDs, snowfall (in winter and summer) and glacier-wide MB, which the linear model was only able to approximate (r2=0.41 for the Lasso vs. r2=0.76 for deep learning in cross-validation31; Fig. Our results indicate that these uncertainties might be even larger than we previously thought, as linear MB models are introducing additional biases under the extreme climatic conditions of the late 21st and 22nd centuries. S5 and S6). Despite the existence of slightly different trends during the first half of the century, both the Lasso and the temperature-index model react similarly under RCP 4.5 and 8.5 during the second half of the century, compared to the deep learning model. Glacier surface mass changes are commonly modelled by relying on empirical linear relationships between PDDs and snow, firn or ice melt8,9,10,29. Strong Alpine glacier melt in the 1940s due to enhanced solar radiation. Massifs without glaciers by 2100 are marked with a cross, b Glacier ice volume distribution per massif, with its remaining fraction by 2100 (with respect to 2015), c Annual glacier-wide MB per massif, d Annual snowfall per massif, e Annual cumulative positive degree-days (CPDD) per massif. This method has the advantage of including glacier-specific dynamics in the model, encompassing a wide range of different glacier behaviours. a deep artificial neural network) or the Lasso (regularized multilinear regression)30. Res. Huss, M. & Hock, R. A new model for global glacier change and sea-level rise. Winter tourism under climate change in the Pyrenees and the French Alps: relevance of snowmaking as a technical adaptation. On the other hand, for flatter glaciers large differences between deep learning and Lasso are obtained for almost all climate scenarios (Fig. Sci. Nonetheless, a close inspection of the annual glacier-wide MB rates from both models reveals similar patterns to those found when comparing deep learning and Lasso approaches (Figs. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Data 12, 19731983 (2020). This results in a higher complexity of the Lasso compared to a temperature-index model.