Improving the Understanding of Alpine Mass Movements by leveraging AI on Spaceborne InSAR Data

Alpine environments are shaped by slow-moving mass movements that can accelerate suddenly, potentially leading to catastrophic failures that threaten human life, infrastructure, and ecosystems. Recent events which occurred without recognized prior warning signs highlight the need for systematic regional-scale monitoring, aimed at improving our understanding of landslide dynamics and associated risks. Spaceborne interferometric synthetic aperture radar (InSAR) provides high-resolution surface displacement data, making it a powerful tool for observing slope activity at different spatial and temporal scales. However, the interpretation of InSAR data remains time-intensive and subjective, limiting its utility for large-scale, continuous assessment. Artificial Intelligence (AI) may offer a solution to these challenges, by enabling automated analysis of InSAR data. Deep learning models, such as convolutional neural networks (CNNs), can be exploited to extract information on the location and activity status of mass movements from interferograms. Moreover, such an approach would reduce subjectivity of expert interpretation while increasing scalability and maximizing spatial coverage.

This work combines AI-driven surface displacement detection with geomorphological assessments to identify correlations between mass movement behaviour and driving factors across different types of mass movements in space and time. Mass movements in the canton of Valais, Switzerland, were manually mapped on Sentinel-1 wrapped interferograms acquired from two ascending and two descending tracks, spanning 12- to 18-day baselines. Classification was performed by considering an internationally established landslide classification scheme – with the addition of the rock glacier class. A specifically developed U-Net model trained on this dataset is applied and evaluated against expert mapping on previously unseen imagery. Performance, assessed via Intersection over Union metric, indicates that AI results are comparable to expert manual mapping. Future iterations aim to incorporate activity status detection and then also automated process classification using optical imagery and digital elevation models. This will allow us to focus on uncovering the underlying mechanisms of landslides through extensive spatio-temporal analyses that integrate geomorphological factors such as geological conditions, topography, and climate variables.