Spatio-temporal reconstruction of winter glacier mass balance in the Alps, Scandinavia, Central Asia and western Canada (1981-2019) using climate reanalyses and machine learning

Spatio-temporal reconstruction of winter glacier mass balance is important for assessing long-term impacts of climate change. However, high-altitude regions significantly lack reliable observations, which is limiting the calibration of glaciological and hydrological models. Reanalysis products provide estimates of snow precipitation also for remote high-mountain regions, but this data come with inherent uncertainty, and assessing their biases is difficult given the low quantity and quality of available (long-term) in situ observations.In this study, we aim at improving knowledge on the spatio-temporal variations in winter glacier mass balance by exploring the combination of data from reanalyses and direct snow accumulation observations on glaciers using machine learning. We use the winter mass balance data of 95 glaciers distributed over the European Alps, western Canada, Central Asia and Scandinavia and compare them with the total precipitation from the ERA5 and the MERRA-2 reanalysis products during the snow accumulation seasons from 1981 until 2019. We develop and apply a machine learning model to adjust the precipitation from the reanalysis products along the elevation profile of the glaciers and consequently to reconstruct the winter mass balance in both space (for glaciers without observational data) and time (filling observational data gaps). The employed machine learning model is a gradient boosting regressor (GBR), which combines several meteorological variables from the reanalyses (e.g. air temperature, relative humidity) with topographical parameters. These GBR-derived estimates are evaluated against the winter mass balance data using (i) independent glaciers (site-independent GBR) and (ii) independent accumulation seasons (season-independent GBR). Both approaches resulted in reduced biases and increased correlation between the precipitation of the original reanalyses and the winter mass balance data of the glaciers. Generally, the GBR models have also shown a good representation of the spatial (vertical elevation intervals) and temporal (years) variability of the winter mass balance on individual glaciers.

Automated summary

In this study, we improved the understanding of the spatio-temporal variations in winter glacier mass balance by combining reanalysis data and direct snow accumulation observations with machine learning. By adjusting precipitation data from reanalysis products, we successfully reconstructed the winter mass balance of glaciers without observational data. The machine learning model showed good representation of the spatial and temporal variability of the winter mass balance.