GrIS-MDM: A Hydrology Knowledge-Based Framework Combining Deep Learning Network for Moulin Detection Using Ultrahigh-Resolution UAV Imagery

Moulins serve as critical hydrological conduits on the Greenland Ice Sheet (GrIS), facilitating the transfer of surface meltwater to the subglacial bed and directly modulating basal lubrication and ice velocity. Despite their significance, automated detection remains difficult; moulins are often sub-pixel scale in standard satellite imagery and are frequently misidentified as crevasses or inactive stream segments due to spectral overlap. This study introduces GrIS-MDM (GrIS Moulin Detection Model), a novel hydrology-informed framework designed to automate moulin extraction using ultra-high-resolution (0.06 m) unmanned aerial vehicle (UAV) imagery.

The GrIS-MDM framework synergises topographic data from Digital Elevation Models (DEMs) with spectral information from Digital Orthophotography Maps (DOMs) through a sequential tripartite workflow. The process begins by identifying tubular depressions using a contour-derived K-index to effectively eliminate shallow noise and surface artifacts. Subsequently, a multistage attention ResU-Net (MAResU-Net) is implemented to segment supraglacial river networks, utilising an automated sample collection protocol that substantially reduces manual labelling requirements. Finally, topological constraints are applied to isolate true moulins at river termini, distinguishing them from spurious depressions within the river interiors.

Validation conducted in the Sermeq Avannarleq region yielded a recall of 0.795 and a precision of 0.729. Experimental results demonstrate that GrIS-MDM achieves a 20.4% improvement in F1-score over traditional depth-based sink-filling methods. Integrating these detected moulins into hydrological models increased the spatial consistency of reconstructed stream networks by 5.8%. Furthermore, drainage analysis confirmed the model’s accuracy, with simulated water capture (90.8%) closely aligning with ground-truth observations (92.3%). Sensitivity tests indicate the framework remains effective at 2-m resolution, suggesting strong potential for deployment with high-resolution satellite platforms such as WorldView or ArcticDEM. This research offers a robust tool for enhancing high-precision supraglacial hydrological modelling and refining GrIS mass balance assessments.