Potential of Transfer Learning for Mapping Supraglacial Stream Drainage Patterns in Northwest Greenland

Supraglacial meltwater drains the Greenland Ice Sheet surface through complex stream networks observable in WorldView satellite images. Mapping these constantly evolving networks, with channel sizes as small as 1-20m wide, is critical to understanding ice sheet drainage efficiency and source areas contributing to sea level rise. However, manual digitization is laborious and infeasible over large spatial domains. Here we employ machine learning to digitize these networks by using a VGG16 architecture pre-trained on ImageNet as an encoder, followed by decoder blocks and convolutional filter layers trained with multiple NW Greenland supraglacial stream networks manually digitized from WorldView imagery. We find that transfer learning-based stream mapping is able to successfully identify many stream stretches and may be a valuable tool to aid in delineation of supraglacial drainage networks. The model was most successful in mapping areas with higher contrast and fewer relict channels. In addition to meltwater routing applications, transfer learning-based algorithm mapping may also allow for increased accuracy of supraglacial catchment delineations used to assess regional climate model outputs of ocean-going runoff. Overall, we anticipate that transfer learning-based models may improve the ease and accuracy of mapping supraglacial stream networks, thus improving understanding of ice sheet surface drainage patterns and climate model performance.