Reducing Uncertainties in Supraglacial Lake Area Quantification through the Incorporation of Floating Ice and Cloud Segmentation

Supraglacial lakes (SGLs), which form on glacier surfaces due to the melting of ice, contribute significantly to local hydrological systems, while also having global impacts. These lakes drain either through a network of surface channels on the glacier or by rapid drainages through moulins. In both scenarios, a substantial amount of ice mass is lost and, in the case of rapid drainages, can lead to short-term glacier flow speed-up due to the lowered friction between the glacier and the bedrock. The quantification of the amount of meltwater contained in these lakes is essential to understanding the role SGLs have in the cryosphere.

Remote sensing techniques, particularly the use of multispectral images, are pertinent in monitoring these lakes and their accompanied rapid drainages. One main hindrance of these images, however, is that the presence of clouds can interfere with data interpretation. While cloud segmentation tools exist to identify the presence of clouds in a scene, they are often incapable of differentiating between the clouds and the underlying ice and snow due to the similarities in optical spectral reflectances. This severely limits the capabilities of automated SGL detection algorithms because it is difficult for the model to determine whether a lake has drained, leaving the bare ice underneath exposed, or whether a cloud is currently covering the lake. Additionally, SGLs can go through periods where they either partially or fully refreeze, leaving portions of the surface area covered with floating ice. With traditional segmentation methods, this area is typically not included in the detected lake area, resulting in an underestimation of the actual lake area present.

To overcome these issues, this work employs the strengths of artificial intelligence to create deep learning-based models capable of segmenting SGLs, floating ice and clouds in Sentinel-2 images. A U-Net convolutional neural network was trained to identify the lakes and floating ice, whereas a self-training network was fed with Fmask labels to iteratively learn to identify cloud pixels in polar regions more accurately than the standard cloud-masking algorithms of Fmask and Sen2Cor. This processing chain aims at improving lake area quantification and aiding in the identification of true rapid lake drainages.