One mass balance model to rule them all: joint assimilation of remote sensing and glaciological data into MassBalanceMachine

Glacier surface mass balance observations are highly heterogeneous, with point glaciological measurements spanning decadal timescales being limited to a few well-monitored regions, while geodetic observations are available globally with a multi-annual baseline and almost full glacier coverage. Existing surface mass balance modelling approaches can only calibrate their parameters for individual glaciers or regions with available observations. This often implies that a big part of the available observations per glacier (generally glaciological data) cannot be exploited for calibration since they play the role of independent data for validation. There is a need to move towards flexible surface mass balance models, capable of leveraging in a coherent fashion both glaciological and remote sensing data.

To address this challenge, we introduce a new version of MassBalanceMachine, a neural network-based model that predicts monthly surface mass balance using topographical features and monthly climate forcing as inputs. The pointwise nature of the model, combined with the global availability of input features thanks to OGGM and ERA5, enables the implementation of custom loss functions to train the model. This loss function allows the model to align with both glaciological measurements and geodetic mass balance observations over multiple decades.

By leveraging the interannual variability and point-wise nature from glaciological data and correcting long term biases with geodetic data, MassBalanceMachine can generate reliable high-resolution monthly mass balance predictions for unmonitored glaciers, and correct existing predictions where previous training strategies are known to fail, e.g. in the Alps. This approach exploits data from data-rich sparse regions to make predictions for other unmonitored regions or glaciers, offering a scalable solution for global glacier mass balance estimation.