Glacier Blueprints: Deriving Spatial Surface Mass Balance from Remote Sensing at a Regional Scale with Machine Learning

The surface mass balance (SMB) of glaciers represents the link between glaciers and their local climate. Quantifying the SMB is essential for calibrating glacier mass-balance models, improving our understanding of the glacier’s response to a changing climate, which affects freshwater availability, sea-level rise, and the risk of natural hazards, among others.

While the SMB cannot be measured directly from space, it can be derived from observations of elevation change, ice velocity, and ice thickness (gradients). Such approaches have been successfully applied in detailed studies of individual glaciers with high spatial and temporal data coverage. However, extending these efforts to regional or global scales present significant challenges due to inconsistent temporal data coverage, coarse spatial resolution, and large uncertainties linked to regional datasets. As a result, many regional glacier evolution models continue to rely on single glacier-wide average mass-balance estimates from long-term geodetic elevation change measurements for model calibration. However, this can lead to model overparameterization and equifinality problems, which are major sources of uncertainty in projections. With the advent of extensive remote sensing datasets and machine learning approaches, there is now an unprecedented opportunity to estimate spatial SMB patterns across glaciers, on regional to global scales.

In this study, we estimate spatial SMB patterns on glaciers in the Swiss Alps with a generalised approach that does not rely on high spatial coverage from in-situ measurements, but rather on datasets with a regional availability. More specifically, we use observational datasets of ice thickness and ice velocity fields derived from remote sensing to calculate the ice flux divergence and combine this with the continuity equation for ice thickness and observations of elevation change to estimate spatial SMB patterns. To optimize the calculation of the ice flux divergence, which relies on non-local ice flow behaviour, we employ a machine learning approach to determine the best filtering (smoothing) parameters for the spatial velocity and thickness gradients. The performance of the method is assessed by comparing SMB estimates with in-situ SMB values derived from stake measurements. This study aims at providing a scalable framework for estimating spatially resolved SMB patterns, with potential applications at the global scale.