Sjursen et al. (2025) Machine learning improves seasonal mass balance prediction for unmonitored glaciers

Identification

• Journal: ˜The œcryosphere
• Year: 2025
• Date: 2025-11-17
• Authors: Kamilla Hauknes Sjursen, Jordi Bolíbar, Marijn van der Meer, Liss Marie Andreassen, Julian Peter Biesheuvel, Thorben Dunse, Matthias Huss, Fabien Maussion, David R. Rounce, B. S. Tober
• DOI: 10.5194/tc-19-5801-2025

Research Groups

• Department of Civil Engineering and Environmental Sciences, Western Norway University of Applied Sciences (HVL), Sogndal, Norway
• Univ. Grenoble Alpes, CNRS, IRD, G-INP, Institut des Géosciences de l’Environnement, Grenoble, France
• Faculty of Civil Engineering and Geosciences, Delft University of Technology, Delft, the Netherlands
• Laboratory of Hydraulics, Hydrology, and Glaciology (VAW), ETH Zürich, Zurich, Switzerland
• Swiss Federal Institute for Forest, Snow, and Landscape Research (WSL), Sion, Switzerland
• Norwegian Water Resources and Energy Directorate (NVE), Oslo, Norway
• Department of Geosciences, University of Fribourg, Fribourg, Switzerland
• Bristol Glaciology Centre, School of Geographical Sciences, University of Bristol, Bristol, UK
• Department of Atmospheric and Cryospheric Sciences, University of Innsbruck, Innsbruck, Austria
• Department of Civil and Environmental Engineering, Carnegie Mellon University, Pittsburgh, PA, USA

Short Summary

This study introduces the Mass Balance Machine (MBM), an XGBoost-based machine learning model, to accurately predict seasonal and annual glacier mass balance on unmonitored glaciers in Norway. MBM significantly outperforms traditional temperature-index models in seasonal mass balance predictions, demonstrating the potential of machine learning to generalize from sparse in situ data.

Objective

• To evaluate the ability of a machine learning model to generalize spatio-temporal information across glaciers using seasonal and annual point mass balance measurements.
• To provide accurate, high-resolution predictions of surface mass balance on unmonitored glaciers in regional-scale applications.

Study Configuration

• Spatial Scale: Regional (mainland Norway), individual glaciers (32), point scale (stake measurements), elevation-band scale, glacier-wide scale. Meteorological data at 9 km horizontal resolution (ERA5-Land), topographical data at 90 m resolution (Copernicus DEM GLO-90).
• Temporal Scale: Data from 1962–2021; model comparison for 1980–2019. Predictions at monthly, seasonal (winter/summer), annual, and decadal resolutions.

Methodology and Data

• Models used:
  • Mass Balance Machine (MBM): Data-driven mass balance model based on the XGBoost architecture.
  • GloGEM (Global Glacier Evolution Model)
  • OGGM (Open Global Glacier Model)
  • PyGEM (Python Glacier Evolution Model)
• Data sources:
  • Glaciological point mass balance measurements: Approximately 4000 seasonal and annual measurements from 32 glaciers in mainland Norway (1962–2021), collected by the Norwegian Water Resources and Energy Directorate (NVE).
  • Meteorological features: Monthly mean values from the ERA5-Land climate reanalysis dataset (2 m air temperature, total precipitation, downward surface solar radiation, forecast albedo, surface sensible and latent heat fluxes, net surface thermal radiation).
  • Topographical features: Aspect, slope, elevation difference between ERA5-Land cell and glacier surface, and skyview factor (derived from Copernicus DEM GLO-90).
  • Geodetic mass balance observations: Satellite-derived DEMs (Hugonnet et al., 2021) for 2000–2019, and photogrammetry/laser scanning (Andreassen et al., 2016, 2020) for specific sub-periods.

Main Results

• MBM successfully predicted annual and seasonal point mass balance on independent test glaciers with RMSE ranging from 0.59 m w.e. to 1.00 m w.e. and bias from -0.01 m w.e. to 0.04 m w.e.
• MBM outperformed traditional glacier evolution models in seasonal mass balance predictions across spatial scales, reducing RMSE by up to 46 % for glacier-wide winter mass balance and 25 % for glacier-wide summer mass balance.
• MBM exhibited very low biases in seasonal mass balance (-0.05 m w.e. for winter, +0.10 m w.e. for summer) compared to glacier evolution models, which showed considerable positive or negative biases (e.g., OGGM: +0.26 m w.e. winter, -0.27 m w.e. summer).
• MBM effectively captured mass balance gradients with elevation, particularly for glaciers in northern and western Norway.
• On annual mass balance, MBM showed similar performance to GloGEM and OGGM (RMSE of 0.97 m w.e. for point scale, 0.54 m w.e. for glacier-wide).
• MBM's monthly predictions differed from other models in transition months (May and September), showing a larger number of negative mass balances.
• MBM showed good agreement with decadal glaciological mass balance rates (RMSE of 0.26 m w.e.a⁻¹ and bias of -0.1 m w.e.a⁻¹).

Contributions

• Developed the Mass Balance Machine (MBM), a novel XGBoost-based machine learning model capable of high spatio-temporal resolution glacier mass balance reconstruction for unmonitored glaciers.
• Demonstrated the superior capability of machine learning models to generalize across diverse glaciers and climatic settings from relatively sparse seasonal and annual point mass balance data, a significant improvement over existing temperature-index models that struggle with transferability to unmonitored glaciers.
• Provided more accurate seasonal mass balance predictions compared to state-of-the-art glacier evolution models, highlighting MBM's potential for enhancing hydrological modeling in glacierized catchments where seasonal observations are scarce.
• Showcased the flexibility of ML approaches to integrate observations at different temporal resolutions (seasonal and annual targets from monthly predictions) and their adaptability for applications such as input to ice flow models and improved glaciological mass balance estimates.
• Proposed a pathway for reconciling glacier mass balance estimates by leveraging the data-driven nature and flexibility of ML to integrate diverse and complementary observational datasets, including remote sensing data, at various spatio-temporal scales.

Funding

• NASA Award 80NSSC24K1530
• DOI NPS Grant #P22AC02208-02

Citation

@article{Sjursen2025Machine,
  author = {Sjursen, Kamilla Hauknes and Bolíbar, Jordi and Meer, Marijn van der and Andreassen, Liss M. and Biesheuvel, Julian Peter and Dunse, Thorben and Huss, Matthias and Maussion, Fabien and Rounce, David R. and Tober, B. S.},
  title = {Machine learning improves seasonal mass balance prediction for unmonitored glaciers},
  journal = {˜The œcryosphere},
  year = {2025},
  doi = {10.5194/tc-19-5801-2025},
  url = {https://doi.org/10.5194/tc-19-5801-2025}
}