Evaluating deep-learning models for debris-covered glacier mapping. (December 2021)

• Record Type: Journal Article
• Authors: Xie, Zhiyuan; Asari, Vijayan K.; Haritashya, Umesh K.
• Is Part Of: Applied computing and geosciences. Volume 12(2021)
• Publication Date: 2021-12
• Subjects: Satellite imagery -- Glacier mapping -- Deep-learning -- Convolutional neural network -- Image segmentation; Earth sciences -- Data processing -- Periodicals; 550.285
• DOI: 10.1016/j.acags.2021.100071
• Languages: English
• ISSNs: 2590-1974

Abstract

Abstract: In recent decades, mountain glaciers have experienced the impact of climate change in the form of accelerated glacier retreat and other glacier-related hazards such as mass wasting and glacier lake outburst floods. Since there are wide-ranging societal consequences of glacier retreat and hazards, monitoring these glaciers as accurately and repeatedly as possible is important. However, the accurate glacier boundary, especially the debris-covered glacier (DCG) boundary, which is one of the primary inputs in many glacier analyses, remains a challenge even after many years of research using conventional remote sensing methods or machine-learning methods. The GlacierNet, a deep-learning-based approach, utilized the convolutional neural network (CNN) segmentation model to delineate DCG at a high level of accuracy. In this study, the performance of GlacierNet's CNN is compared with several advanced CNN segmentation models, including Mobile-UNet, Res-UNet, FCDenseNet, R2UNet, and DeepLabV3+, to identify the most salient features that could improve the DCG segmentation accuracy. The experimental evaluation shows the highest intersection over union (IOU) of 0.8623 for the DeepLabV3+ and, therefore, is recommended for the regional and large-scale DCG mapping. Moreover, GlacierNet's CNN with the second-highest IOU of 0.8599 is a suitable and light structure for regional DCG mapping.