Three times accelerated glacier area loss in Svalbard revealed by deep learning

Konstantin Maslov, Thomas Schellenberger, Claudio Persello, Alfred Stein

The rapid warming in polar regions highlights the need to monitor climate change impacts such as glacier retreat and related global sea level rise. Glacier area is an essential climate variable but its tracking is complicated by the labour-intensive manual digitisation of satellite imagery. Here we introduce ICEmapper, a deep learning model that maps glacier outlines from Sentinel-1 time series with accuracy on par with human experts. We used this model to retrieve Svalbard glacier outlines for 2016–2024 and found a tripling of the glacier area loss rate (-260 km2 p.a.) in the last decade as compared to that previously reported for 1980–2010 (-80 km2 p.a.). This acceleration is largely driven by increased calving at tidewater glaciers and the climatic warming signal impacting land-terminating glaciers. Additionally, our analysis shows significant area changes related to glacier surging, namely, the Nathorstbreen system and Austfonna, Basin-3 surges. These two surges collectively added to the area change in 2006–2016 (+194.30 km2 or +0.59%), thus delaying the regionwide area loss by approximately two years. Our results indicate a significant acceleration in glacier area loss in Svalbard, and we anticipate broader applications of our method to track glacier changes on larger scales.