Abstract
Glaciers are identified as “Essential Climate Variables” by the World Meteorological Organization as changes in glacier area, elevation and mass are major indicators for climate change. Commonly employed methods for the estimation of glacier mass balance consist of in-situ measurements of snow accumulation and ice ablation, and repeat digital elevation model (DEM) differencing. With these techniques, due to the limited availability of data, glacier mass balance information is sparse and annual mass balance time series exist only for few glaciers in the world. Physically based models can be employed to overcome such limitations, as they make use of climatic variables, such as temperature and precipitation, to derive glacier mass balance time series. However, in high altitude the quality of climatic variables is often an issue due to the limited availability of weather stations in such areas, and ad-hoc calibration procedures are often employed to ensure reliable results. Alternative approaches are based on finding empirical correlations between snow cover variations and glacier mass balance. Indeed, it has been found that variations of snowline altitude at regional scale around the glacier is correlated to the winter mass balance, while summer snowline elevation is a proxy for summer and annual glacier mass balance. Up to now, these empirical methods have been only applied on small areas, by calibrating the empirical correlation function on a regional scale. In this work we tested the capability of machine learning and deep learning to derive a glacier mass balance estimation method which can be employed over large areas worldwide. In particular, we derive monthly regional winter snowline altitude time series extracted from MODIS data, to take into account the process of snow accumulation. Time series of summer glacier snowline elevation derived from high resolution sensors such as Sentinel-2 and Landsat, instead, are correlated to the ablation processes. With these we build a reference dataset which is used to train different machine learning and deep learning models calibrated with in-situ measurements of glacier mass balance. The glacier mass balance estimation method is then tested in different areas worldwide.