Regional Glacier Mapping and Change Analysis in the Gangshika Region of Qilian Mountain Based on Deep Learning and Satellite Data Time-Series

Time-series glacier change analysis that is closely linked to global climate patterns has drawn increasing attention. However, few studies focused on regional glacier mapping with high spatial resolution and accuracy. This study combined a lightweight DeeplabV3+ model and activation threshold optimization strategy for regional glacier mapping and change analysis in the Gangshika region of Qilian Mountain, mainly using high-resolution satellite images from 2012 to 2023. For the extraction of glaciers with different scales, multiscale contextual information was represented through Atrous spatial pyramid pooling module. The predicted probability maps were first obtained. And the optimized activation thresholds were selected to determine which areas were glaciers. The accuracies of extracted glacier boundaries from 2012 to 2023 fluctuated between 82% and 96%, with an averaged value of 93%. Considering that the quality of images in 2012–2014 and corresponding glacier boundary accuracy were not very good, the change analysis was just conducted from 2015 to 2023. As of 2023, the Gangshika region hosts 45 glaciers, with a total area of 20.652 km² and a total perimeter of 148.8 km. From 2015 to 2023, they reduce by 9.2% and 1.7% respectively. Glaciers were predominantly concentrated in medium elevation zones (4200–4500 m) with a proportion of 46.6% and largest ablation rate of 1.68% between 2021 and 2023. The thickness and ice reserves were calculated by an empirical formula. From 2015 to 2023, the Gangshika glaciers showed reduction rates in thickness and ice reserves by 1.15% and 12.5%, respectively.