Estimating Supraglacial Lake Area for Greenland using Sentinel-2 Images and Deep Learning

Abstract

Ice loss from the Greenland ice sheet is one of the main sources of global sea level rise. Surface meltwater is one of the drivers of Greenland ice sheet mass loss. Supraglacial lakes are formed when meltwater accumulates in topographic depressions on glaciers or ice sheets during the melt season. The development and rapid drainage of supraglacial lakes in Greenland have been linked to the collapse of floating ice shelves. This can then lead to increased discharge of ice from outlet glaciers and increased ice velocity. Supraglacial lakes in Greenland are studied using Sentinel-2 images with daily observation intervals and high spatial resolution. The objective of this study is to estimate the maximum area of supraglacial lakes using Sentinel-2 L1C images between 2019 and 2022 in the months of July and August. After pre-processing Sentinel-2 L1C images, the detection and semantic segmentation of supraglacial lakes is carried out using a deep learning algorithm. As large labeled Sentinel-2 images are not available and labeling the training data is time-consuming, the F-mask algorithm is used for the training data labels. The deep learning algorithm consists of several stages, and the model is validated with manually labeled data at every stage. The training labels for the next stages are generated from the most successful model of its previous stage. After that, labels are generated for all acquired images, and the maximum area of the lakes in the months of July and August between 2019 and 2022 is calculated for supraglacial lakes in Greenland sub-regions.