Automated Mapping of Braided Palaeochannels From Optical Images With Deep Learning Methods

Vanzani, F.; Carbonneau, P.; Fontana, A.

Abstract

The increasing availability of remotely sensed data has provided an enormous quantity of information for studying the geomorphology of exposed surfaces of alluvial plains. In many cases, the key for reconstructing their formation lies in the recognition of optical traces related to abandoned palaeochannels and their morphometric characteristics. Abundant braided palaeohydrographic traces are documented in alluvial plains of northern Italy, where large sectors of the present surface correspond to landforms related to fluvioglacial systems supplied by Alpine glaciers during the Last Glacial Maximum (LGM). Nevertheless, the complexity of multichannel patterns, the overlapping field division systems and urbanization, hinder the efforts to manually map these traces. In this work, we used high‐resolution aerial photos of the proximal sector of the Friulian Plain (NE Italy) to train an Attention‐UNet deep learning algorithm to segment palaeohydrographic traces. The trained model was used to automatically recognize braided palaeochannels over 232 km2. The resulting map represents a significant step for investigating the long‐term alluvial dynamics. Moreover, we assessed the robustness of our method by deploying the model in three other areas in northern Italy with comparable characteristics, as well as in Montenegro, near Podgorica. In each case, the braided pattern was successfully mapped by the algorithm. This work highlights the breakthrough potential of deep learning methods to rapidly detect complex geomorphological traces in cultivated plains, taking into consideration advantages, challenges and limitations.

Citation

Vanzani, F., Carbonneau, P., & Fontana, A. (2025). Automated Mapping of Braided Palaeochannels From Optical Images With Deep Learning Methods. Journal of Geophysical Research: Earth Surface, 130(2), Article e2024JF008051. https://doi.org/10.1029/2024jf008051

• Journal Article Type: Article
• Acceptance Date: Jan 24, 2025
• Online Publication Date: Feb 20, 2025
• Publication Date: Feb 1, 2025
• Deposit Date: Jan 24, 2025
• Publicly Available Date: Feb 26, 2025
• Journal: Journal of Geophysical Research: Earth Surface
• Print ISSN: 2169-9003
• Electronic ISSN: 2169-9011
• Publisher: American Geophysical Union
• Peer Reviewed: Peer Reviewed
• Volume: 130
• Issue: 2
• Article Number: e2024JF008051
• DOI: https://doi.org/10.1029/2024jf008051
• Keywords: deep learning, alluvial plains, remote sensing, last glacial maximum
• Public URL: https://durham-repository.worktribe.com/output/3349152