Towards an unsupervised generic seismic detector for hazardous mass-movements: a data-driven approach

Hazardous mass-movements pose a great danger to the population and critical infrastructure, especially in alpine areas. Monitoring and early-warning systems can potentially save many lives and improve the resilience of mountain communities to catastrophic events. Increasing coverage of seismic networks recording hazardous mass-movements opens up new warning perspectives as long as efficient algorithms screening the seismic data streams in real-time are available.

We propose to combine physical and statistical properties of seismic ground velocity recordings from geophones and seismometers as a foundation for an unsupervised workflow for mass movement detection. We evaluate the performance, consistency, and generalizability of unsupervised clustering algorithms like K-means and Bayesian Gaussian Mixture Models against supervised methods like the Random Forest classifier. Focusing on debris-flow records at the Illgraben torrent in Switzerland, we present a generic mass-movement detector with high accuracy and early-warning capability. We apply this detector to other datasets form other sites to investigate its transferability.

Since our results aim to enable mass-movement monitoring and early-warning worldwide, Open Research Data principles like Findability, Accessibility, Interoperability and Reusability (FAIR) are of high importance for this project. We discuss how using the Renku (renkulab.io) platform of the Swiss Data Science Center ensures FAIR data science principles in our investigation. This is a key step towards our ultimate goal to enable seismology-based early warning of mass-movements wherever it may be required.