From Decision Trees to Deep Learning: Enhanced Supraglacial Lake Detection in Antarctica

With the ongoing effects of global warming, supraglacial meltwater in polar regions plays a critical role in ice sheet dynamics, influencing global sea levels. In Antarctica, the accumulation of meltwater on ice surfaces not only reduces albedo—accelerating melting in a self-reinforcing cycle—but also drives processes such as meltwater injection and basal lubrication, with possible destabilizing effects for ice sheets. Monitoring the seasonal evolution and dynamics of supraglacial lakes is essential for understanding these processes, yet the vast and remote nature of the Antarctic ice sheet presents significant challenges. Spaceborne remote sensing offers the best solution, providing continuous, large-scale, and long-term observations. However, extracting reliable information from optical and synthetic aperture radar (SAR) data remains complex due to limitations in spatial transferability, cloud cover, polar night, and the spectral similarities of frozen lakes with surrounding ice. The Sentinel mission bridges these gaps, enabling the combination of optical and SAR data to achieve the best possible accuracy for mapping and monitoring supraglacial lakes.

This study evaluates whether a deep learning-based mapping approach outperforms a pixel-based Random Forest (RF) classification algorithm for supraglacial lake (SGL) detection in Antarctica. As a benchmark, we utilized an RF model trained on 14 regions and 24 input channels, including Sentinel-2 spectral bands, spectral indices, and topographic variables. To work toward a circum-Antarctic, operational SGL mapping product, we reduced the input channels by selecting the four most important features identified by the RF approach and trained a convolutional neural network (CNN) on partially labeled data from 16 Sentinel-2 scenes, including more images with cloud cover. Both models were validated using the same 16 test areas across eight Antarctic ice shelves.

The RF approach achieved a producer’s accuracy, user’s accuracy, and F1 score of 0.750, 0.945, and 0.837, respectively, whereas the CNN-based workflow achieved scores of 0.915, 0.912, and 0.913, respectively. In scene-specific comparisons, the CNN outperformed the RF approach in 13 of the 16 validation scenes. Key advantages of the CNN approach include its ability to detect lakes under thin clouds and over floating ice, resulting in less fragmented lake area estimates and requiring fewer input features. However, challenges persist in transition zones between lakes and slush, where spectral details outweigh the benefits of shape-based detection.