EAGLE

Motivation

‍Slow-moving mass movements pose a major concern in the Swiss Alps, since they can result in significant hazards such as landslides, rock avalanches, and debris flows, endangering both human lives and infrastructure. Currently, in Switzerland, the detection and monitoring of these processes heavily rely on manual methods, where experts analyse Earth Observation (EO) data or conduct field measurements. Large constellations of satellites are constantly circling the Earth, capturing and storing high-resolution EO data, leading to an exponentially growing archive of data. Thus, the core objective of our project is to harness the power of machine learning to unlock the potential of this rich repository of EO data. By developing innovative algorithms and techniques, we aim to automate the detection and monitoring processes, providing more accurate countrywide mass movement catalogues for the Swiss Alps.

Proposed Approach / Solution

We aim to develop deep learning-based approaches for mapping alpine landforms of interest using various Earth observation (EO) data, with a particular focus on Interferometric Synthetic Aperture Radar (InSAR) data (see Figure 1). While some models have demonstrated promising mapping capabilities using optical images (Prakash et al., 2021; Prakash et al., 2020) or stacked InSAR data (Liang et al., 2023), only a few have explored the direct mapping of mass movements from wrapped InSAR data (Bralet et al., 2024).

‍

‍

Our approach involves creating a comprehensive dataset of manually mapped mass movements using Sentinel-1 wrapped interferograms from both ascending and descending tracks, covering a range of temporal baselines in the Swiss Alps. This dataset is then used to train advanced segmentation models specifically designed for InSAR data, including U-Net, DeepLabv3, SegFormer, and deep residual shrinkage models. These models are evaluated against expert annotations on previously unseen imagery. Additionally, our work includes the integration of other EO data, such as optical imagery and digital elevation model derivatives. Finally, we are exploring novel techniques to classify the dynamics and activities associated with these processes.

Impact

‍Automating the mapping of slow-moving mass movements will significantly improve the accuracy, consistency, and timeliness of hazard detection in the Swiss Alps. Our approach enables the creation of large-scale, up-to-date catalogues, supporting early warning systems, infrastructure planning, and disaster response. This contributes to more effective risk management and increased resilience in alpine regions.

‍