Mapping the hydrology of Greenland’s crevasses with deep learning

Greenland’s crevasse fields transfer nearly half of the seasonal runoff to the bed of the ice sheet, with implications for ice rheology, subglacial hydrology, and subsequent feedbacks in ice dynamics. The hydrological behaviour of crevasses has been shown to be complex and spatially heterogenous, but the drainage mechanics are poorly understood, particularly in comparison to other water pathways (lake drainage, moulins, and finer-scale fractures). To better understand crevasse drainage processes at scale, we develop a convolutional neural network (CNN) to map water-filled crevasses from 10 metre resolution Sentinel-2 MSI imagery. Training and validation datasets are produced using NDWI-based approaches that are accurate but require time-consuming and scene-specific manual tuning. In contrast, our scaleable CNN approach allows for the seasonal and multiannual evolution of ponded crevasse fields to be efficiently monitored. After constructing a comprehensive, time-evolving dataset of crevasse field hydrology, we aim to quantify controls (strain rate, melt rate, etc.) on the time-evolving filling and drainage of crevasses. Our ultimate objective is to use these derived relationships to improve the parameterisation of spatially heterogenous crevasse hydrological behaviour into coupled models of Greenland Ice Sheet hydrology-dynamics.