A Progressive Learning Strategy for Large-Scale Glacier Mapping

In recent years, the worldwide temperature increase has resulted in rapid deglaciation and a higher risk of glacier-related natural hazards such as flooding and debris flow. Due to the severity of these hazards, continuous observation and detailed analysis of glacier fluctuations are crucial. Many such analyses require an accurately delineated glacier boundary. However, the complexity and heterogeneity of glaciers, particularly debris-covered glaciers (DCGs), poses a challenge for glacier mapping when using conventional remote sensing or machine-learning techniques. Some examples exist about small-scale automated glacier mapping, but large or regional-scale mapping is challenging. Previously, a deep-learning-based approach named GlacierNet2 had been developed to accurately delineate the complete DCG outlines on the regional scope via taking advantage of multiple models. This paper uses a modified version of GlacierNet2 to study the feasibility and effectiveness of large-scale glacier mapping in Nepal Himalaya, Karakoram, and parts of western Himalaya. Also, we propose a large-scale mapping strategy to progressively enhance the network familiarity to varied types of glaciers via systematically repeating the training process. This strategy allows the network to delineate a large number of glaciers while only requiring a small proportion of initial training data. Thus, resulting in a significant drop in labor and expert intervention, which are required for selecting and labeling the training data. Our results show a successful and accurate generation of glacier boundaries with an intersection over union (IOU) score of 0.8115 in the Karakoram and parts of western Himalaya and an IOU of 0.7525 in the Nepal Himalaya. Our work outlines how future efforts of large and global scale mapping can be developed to monitor and analyze glacier dynamics.