Seismic Monitoring and Machine Learning for Gravitational Hazards in Alpine Terrain

Gravitational hazards, such as rockfalls, landslides, and debris flows in steep mountain terrain, have a high destructive potential, and play a key role in erosion and landscape evolution. Furthermore, in response to climate change, such Alpine mass movements are likely to increase in number and volume in the future. Therefore, monitoring has to be improved to warn against mass movements and mitigate more significant damage. It is crucial to obtain an insight into both the process cascade and the step-wise formation of such events. As an alternative to many conventional approaches, seismology allows monitoring of mass movements at safe distances, provides estimates of event location and timing, and gives insights into flow dynamics and rheology of granular masses. This doctoral thesis aims to establish automated detection and classification of seismic signals and explores the extent to which seismic signatures contain information on mass movement parameters, including flow height and volume. This first part of the thesis focuses on seismic data recorded between 2017 and 2020 at Illgraben, a steep canyon located in Switzerland, which is known for its high debris-flow and rockfall activity. High rainfall intensities in summer regularly trigger mass wasting with rock-slope failure and debris flows. The frequent debris-flow occurrence (up to ten events per year) makes the Illgraben an ideal site for cross-validating seismically derived mass-movement characteristics with ground-truth data, such as digital terrain models, and other in-torrent measurements. I first focus on the high-frequency (> 1 Hz) spectral content of the seismic signal and use an impact model to estimate particle sizes of debris flows. I show that by focusing on small time windows of the recorded signal, the power spectral density does correlate with particle sizes estimated from video footage. Next, I investigate the long period (> 5 s) seismic data recorded by a broadband seismometer at Illgraben and seismic data from a large-scale debris-flow experiment to estimate ground tilt caused by the weight of the passing mass. I correlate the observed tilt with bulk debris-flow parameters such as volume, vertical force, and maximum flow height and find a Pearson correlation coefficient larger than 0.76 for flow height and maximum tilt for all stations. This indicates that such parameters can successfully be inferred from sensors placed at a safe distance. I then use the solution of the Boussinesq problem for the elastic response of the ground to a surface force to model the elastic ground deformation induced by a distributed mass. I find that despite uncertainties in the model’s elastic sub-surface properties, the tilt obtained from the seismic recordings lies within the model’s predicted range. To demonstrate the transferability of the findings, I then use a simplified version of the model to make predictions of flow height of two debris flows that occurred at Volcán de Fuego, Guatemala, with sparse auxiliary data coverage. To investigate monitoring and near-real-time warning capabilities, I use seismic data recorded near Pizzo Cengalo, a mountain located in eastern Switzerland that recently witnessed partial slope collapses up to 3.5 x 10ˆ6 mˆ3 . To train a machine-learning classifier intended to detect seismic signals of such slope failures, I use windowed seismic data of an imbalanced event catalog containing only five slope failure events, which leads to a highly biased training data set. By lowering the classification threshold for the minority class (seismic signals labeled as slope failures), I manage to obtain a prediction accuracy of 80% for the slope failure class. In a real-world simulation, I demonstrate that by including a seismic reference station at a great distance to the area of interest (> 100 km), a misclassification of earthquakes as slope failures can be prevented. Moreover, after 1.5 years of operational use, I reevaluate the classifier’s performance, during which six slope-failure alarms were issued. While one of the alarms successfully warned against the only reported event during this period, five alarms referred to Magnitude 3.4 - 4.4 tectonic earthquakes in northern Italy. I thus retrain the machine-learning classifier with the new events, which results in a 4% increase in performance for the slope failure class. Finally, the approach from my Pizzo Cengalo study was transferred to Illgraben by training a classifier to detect slope-failure signals at the channel head. The channel deposits are typically remobilized during debris-flow events. Data from summer 2017 is used to train the classifier, which is subsequently applied to summer 2019. I correlate the detected events with meteorological data, such as temperature, precipitation, and atmospheric pressure. In 2019, neither atmospheric pressure nor temperature variations can be identified as clear slope failure triggers. However, precipitation and slope failure frequency show a trend indicating a non-trivial relationship. I propose that freeze-thaw cycles in winter provide loose sediment to the gullies, which gets evacuated due to pore-pressure increase during the snowmelt season and intense rainfall events. Towards the end of the season, the availability of loose sediment limits the slope failure activity. In summary, I evaluated the capability of seismic measurements in monitoring and real-time warning of gravitational hazards and found that seismology can close the gap between traditional point-in-space and point-in-time measurements. This enhances the understanding of mass-movement triggers and dynamics. Moreover, through careful parameter choices, I showed that machine-learning techniques could give reliable slope- failure detections, enhancing operational monitoring and early warning