Towards detection and understanding of paraglacial landslides in Southern Alaska using surface velocity data and deep learning

Glacier mass loss results in destabilization of adjacent slopes. In Southern Alaska in particular, the combination of amplified glacier mass loss and fjord environments pose a potential cascading tsunami hazard that may intensify. This study utilizes the ITSLIVE glacier velocity dataset, which covers glacier-adjacent areas, to detect landslides. The re search involves manual mapping and investigates the potential of deep learning techniques for landslide detection using this dataset, while also analyzing the environmental factors that influence landslide occurrence. A total of 182 paraglacial landslides were identified, confirmed, and manually mapped across Southern Alaska. However, the detection likely only captures landslides with pronounced activity, and data quality issues may are present. Despite these limitations, the dataset proves valuable due to its accessibility and broad coverage. An initial deep learning-based landslide detection algorithm shows promising results in some areas but performs less effectively in others (Precision: 0. 25, Recall: 0. 53, F1Score: 0. 29, Mean IoU: 0. 34). Key challenges include noise, a limited training dataset, andtheneedforcarefulselectionofinputdatachannels. Analysisofthemappedlandslides reveals a preferential occurrence in areas where glaciers are experiencing above-average mass loss, suggesting a potential link between glacial dynamics and landslide activity. Additionally, landslides are more common in regions with higher annual precipitation, in dicating a possible connection between wet climate and slope destabilization. Permafrost does not appear to be a major factor at most sites, while the lithology aligns with known susceptibilities.