Reconstructing Annual Global Glacier Mass Balance using Bayesian Neural Fields

The accelerating loss of glacier mass is disrupting local ecosystems, reshaping hydrological regimes, increasing the likelihood of glacier-related hazards, and undermining the resilience of dependent communities. Quantifying annual glacier mass balance remains challenging because existing estimates rely on either sparse in situ and remote sensing observations or process-based models, each with inherent limitations. Observational approaches, while consistent at global scales, exhibit large regional variations. In contrast, modelling frameworks provide complete spatiotemporal fields but are typically deterministic, sensitive to calibration data, and constrained by assumptions that may overlook key energy balance drivers. Through our approach, we combine the strengths of models and various observational methods in a Bayesian neural network framework. Specifically, we use a Bayesian Neural Field architecture that is first pre-trained on fixed-geometry annual mass balance outputs from the Open Global Glacier Model (OGGM) and then finetuned using multimodal observations available at different spatial and temporal scales. The model uses static glacier characteristics and near-surface climate variables as predictors. We demonstrate through a blocked testing strategy that our framework can fill gaps for glaciers with missing or highly uncertain mass balance records as well as reconstruct meaningful long-term time series. In summary, this approach provides a scalable, uncertainty-aware method for generating spatially and temporally complete annual glacier mass balance estimates.