From ground to space: joint assimilation of glaciological and remote sensing data for glacier mass balance modelling with machine learning

– Alban Gossard, Post-doctorant, IGE –

Résumé :

Glaciers play a role in freshwater storage, especially during drought periods, yet their response to climate change remains uncertain due to the limitations of current mass balance models. While all mass balance models approximate accumulation and ablation processes, traditional approaches rely on site-specific calibration using remote sensing data, with sparse in-situ glaciological measurements reserved for validation. This requirement limits their transferability, particularly in regions where glaciological data are limited or unavailable.

In this presentation we introduce Mass Balance Machine, a machine learning-based model designed to overcome these calibration challenges. By training a single statistical model on diverse glaciological datasets, Mass Balance Machine generalizes to unmonitored glaciers without requiring site-specific calibration. We also present ongoing work on the joint assimilation of glaciological and remote sensing data, enabling decadal mass balance forecasts at high spatio-temporal resolution.

Finally, we end this presentation by demonstrating how this framework can be integrated into ODINN.jl, a hybrid glacier modelling framework developed in collaboration between IGE and Stanford. This model uses Universal Differential Equations (UDEs) to combine traditional iceflow dynamics with machine learning. This hybrid approach bridges the gap between physical laws and data-driven methods, offering a powerful tool to explore new physical parametrizations using sparse and heterogeneous datasets.