Continental-scale mapping of Antarctic supraglacial hydrology using machine learning

Surface meltwater influences Antarctic ice-shelf stability by enhancing melt, lowering albedo, and promoting hydrofracture. Although supraglacial hydrology is now recognised as widespread across Antarctica, existing meltwater feature records vary in spatial resolution and temporal sampling. Recent work tends to only classify supraglacial lakes while inconsistently representing more transient features such as slush, underestimating the full extent of surface meltwater. Furthermore, traditional threshold-based supraglacial meltwater mapping approaches require extensive manual post-processing and are difficult to scale; machine learning offers a promising alternative but requires systematic evaluation to ensure classifiers generalise reliably across time and space.

Here, as part of the ESA 5D Antarctica project, we present a new continent-wide, high-resolution record of supraglacial hydrology across all Antarctic ice shelves from 2016 to 2026. The dataset is derived from ~135,100 Sentinel-2 images using supervised machine learning implemented in Google Earth Engine. To systematically evaluate machine-learning approaches, which remain underexplored in glaciological applications and are often applied without rigorous validation, we compared five algorithms: Random Forest, Gradient Boosting Decision Trees, Classification and Regression Trees, Support Vector Machines, and k-Nearest Neighbours. Each was assessed using five complementary validation experiments: repeated cross-validation to assess internal consistency, independent validation against expert-labelled data to test external accuracy, leave-one-year-out cross-validation to evaluate temporal transferability, leave-one-region-out testing to assess spatial transferability, and controlled label corruption to quantify sensitivity to annotation error. Random Forest achieved the most consistent performance and ranked first overall with a mean macro-F1 score of 0.992 and was selected for continent-wide deployment.

The resulting dataset provides monthly classifications of supraglacial hydrology, distinguishing open meltwater features, including lakes, channels, and water-filled crevasses, from non-open meltwater features such as saturated firn and slush. The dataset is delivered alongside an interactive cloud-based application that enables users to visualise, classify, and export products on demand. By resolving Antarctic surface hydrology at unprecedented spatial and temporal scales and enabling on-demand delineation of meltwater features through a publicly available application, this work supports the assessment of processes relevant to ice-shelf stability and provides constraints for climate and ice-sheet modelling. This capability is increasingly important for understanding the role of surface meltwater in Antarctic ice-shelf systems under future warming.