im PublikationsserverTitelangaben
Himmelstoss, Toni
; Mikolka-Flöry, Sebastian ; Haas, Florian ; Becht, Michael ; Pfeifer, Norbert ; Heckmann, Tobias:
Evaluation of convolutional neural networks (CNNs) for the automatic generation of geomorphological maps in high alpine environments.
2025
Veranstaltung: EGU General Assembly 2025, Wien.
(Veranstaltungsbeitrag: Kongress/Konferenz/Symposium/Tagung, Poster)
Volltext
|
Link zum Volltext (externe URL): https://doi.org/10.5194/egusphere-egu25-8600 |
Kurzfassung/Abstract
Geomorphological maps are essential tools for understanding landscape evolution and natural hazards in mountain environments. However, their creation requires substantial time investment and expert knowledge, limiting the availability of up-to-date mapping products. While automated mapping approaches have been developed for selected individual landforms, no spatially exhaustive method exists for the complex terrain of high alpine environments. To overcome this, we evaluated different CNN models, achieving state-of-the-art results for semantic segmentation, for automated geomorphological mapping. For the training we created homogenized maps of three alpine valleys covering more than 170 km² and comprising 20 landform classes. As input layer we tested various three-band raster composites, consisting of various combinations of digital elevation model (DEM) derivatives like topographic openness or terrain wetness index.
Preliminary results show that several models achieve F1 scores exceeding 0.8 for the most relevant geomorphological features, including ground moraine, rock glaciers, lateral moraines, and fluvial terraces. Lower performance was observed for narrow and shallow landforms and anthropogenic features like streets and buildings. However, anthropogenic features are often underrepresented in high alpine valleys explaining their worse performance. Hence, our results indicate that additional manual correction is necessary to use these automatically derived maps in downstream tasks. However, the time required to create geomorphological maps of consistent quality, on the basis of these automatically derived maps, can significantly be reduced. This enables rapid geomorphological mapping in previously unmapped high alpine catchments and facilitates the creation of multitemporal maps within single study areas. The latter application opens new possibilities for quantifying structural changes in alpine geomorphic systems over time, contributing to our understanding of landscape evolution and response to climate change.
Forschungsprojekte
Weitere Angaben
| Publikationsform: | Veranstaltungsbeitrag (unveröffentlicht): Kongress/Konferenz/Symposium/Tagung, Poster |
|---|---|
| Schlagwörter: | Deep Learning, CNN, Geomorphological Maps, Neural Networks, Automated Mapping |
| Sprache des Eintrags: | Englisch |
| Institutionen der Universität: | Mathematisch-Geographische Fakultät > Geographie > Lehrstuhl für Physische Geographie |
| DOI / URN / ID: | 10.5194/egusphere-egu25-8600 |
| Titel an der KU entstanden: | Ja |
| KU.edoc-ID: | 36431 |
Letzte Änderung: 19. Mär 2026 07:29
URL zu dieser Anzeige: https://edoc.ku.de/id/eprint/36431/
