Assigned Session: #AGM28: Generic Meeting Session
Glacier mapping using Deep Neural Networks in the Tropical Andes
Abstract ID: 28.7399 | Accepted as Poster | Poster | 2025-02-28 12:45 - 14:15 | Ágnes‐Heller‐Haus/Small Lecture Room
Diego Pacheco Ferrada (0)
Seehaus, Thorsten (1)
Diego Pacheco Ferrada ((0) Friedrich-Alexander-Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Bayern, DE)
Seehaus, Thorsten (1)
(0) Friedrich-Alexander-Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Bayern, DE
(1) Friedrich-Alexander-Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Bayern, DE
Tropical Andes Glaciers have experienced a significant and accelerated decrease over the last decades, mainly driven by climatic variables affected by climate change. Despite their importance in high-altitude environments, only a few studies have evaluated the area and volume changes on regional and multitemporal scales, focusing mainly on specific areas in Perú or Bolivia. Furthermore, the potential growth of debris-covered glaciers extent imposes new challenges for mapping, especially with conventional threshold methods. Therefore, this study aims to generate updated and temporally consistent outlines of the Tropical Andes Glacier by implementing a fully automatic routine supported by machine-learning approaches, which can be suitable to evaluate the ice volume change over the last decade in the tropics. A deep learning model using state-of-art architectures was trained to map the glacier extent across the Tropical Andes. Here, the Glacier-VisionTransformer-U-Net (GlaViTU) -a hybrid deep learning model composed by a segmentation transformer inline with a convolutional neural network- was trained for a large-scale glacier delineation considering the most recent Peruvian glacier inventory (INAIGEM) from 2020, including debris-free and debris-cover glaciers. The model was fed with diverse remote sensing data: optical (Sentinel-2), topographic features (Copernicus DEM elevation and slope), and synthetic aperture radar (SAR) data (Sentinel-1 backscatter and coherence). Once trained, the model has successfully reproduced the overall glacier’s extent of the Peruvian Andes. While the model performs best (IoU) when segmenting debris-free glaciers, the most challenging areas remain debris cover glaciers. Nonetheless, coherence maps from repeat-pass and multiple orbits improve differentiation between debris-cover and debris-free glacier areas, as well as mitigate the impact of shadowed and overlayed. Moreover, results show that clouds partially occluding glaciers do not affect the delineation, which is a crucial improvement in regions with limited cloud-free optical imagery. This study underscores the importance of combining remote sensing data to improve automated glacier mapping. These results highlight the potential for multitemporal glacier monitoring in the entire Tropical Andes; enabling periodic glacier mapping to evaluate temporal evolution and, volume changes over the last decade in combination with remote sensed DEMs, e.g. TanDEM-X acquisitions.
![]() |
||||||||
|