Automatic Extraction of Supraglacial Rivers and Lakes on the Greenland Ice Sheet from High-Resolution Worldview Imagery Using a Deep Learning Approach

Abstract:

Supraglacial rivers abundantly cover the Greenland Ice Sheet ablation zone throughout the melt season, transporting large volumes of meltwater to the ice edge and the ocean, contributing to sea level rise. Meltwater transport off the surface of the ice sheet through this supraglacial hydrologic network is not incorporated in numerical models of ice sheet runoff, partly due to the lack of high-resolution observations. Common methods of delineating supraglacial rivers and lakes (SRL) from satellite imagery usually require intervention of manual digitization via image interpretation, limiting widespread application.

Here, a deep learning approach is proposed for pixel-wise segmentation of SRL in high-resolution multispectral Worldview (WV) imagery. The model is trained and validated using WV02/3 images collected over the Southwest Greenland ablation zone along with their corresponding labels (water/non-water), and tested on an 84 km² area of a WV02 image in which SRL were previously marked by human experts. The proposed deep learning framework yields pixel-wise predictions at the original input resolution, greatly improves time efficiency (assuming just one GPU), and is highly flexible to extend to other polar regions and various satellite images. Continued algorithm development with more detailed training sets will no doubt improve classification accuracy.