Application of Deep Learning in Glacier Boundary Extraction: A Case Study of the Tomur Peak Region, Tianshan, Xinjiang

Glaciers are one of the most important water resources in the arid regions of Xinjiang, making it crucial to accurately monitor glacier changes for the region’s sustainable development. However, due to their typical distribution in remote, high-altitude areas, large-scale and long-term field observations are often constrained by the high costs of manpower, resources, and finances. Globally, fewer than 40 glaciers have been monitored for more than 20 years, and, in China, only Glacier No. 1 at the headwaters of the Urumqi River has monitoring records exceeding 50 years. To address these challenges, this study analyzed glacier changes in the Tomur Peak region of the Tianshan Mountains over the past 35 years using Landsat satellite imagery. Through experiments with deep learning models, the results show that the 3-4-5 band combination performed best for glacier boundary extraction. The DeepLabV3+ model, with MobileNetV2 as the backbone, achieved an overall accuracy of 90.44%, a recall rate of 82.75%, and a mean Intersection over Union (IoU) that was 1.6 to 5.94 percentage points higher than other models. Based on these findings, the study further analyzed glacier changes in the Tomur Peak region, revealing an average annual glacier reduction rate of 0.18% and a retreat rate of 6.97 km2·a−1 over the past 35 years. This research provides a more precise and comprehensive scientific reference for understanding glacier changes in arid regions, with significant implications for enhancing our understanding of the impacts of climate change on glaciers, optimizing water resource management, and promoting regional sustainable development.