Unsupervised Machine Learning Algorithms for Seismic Detection of Catastrophic Mass Movements

Catastrophic mass movements threaten mountain communities worldwide. Rockfalls, avalanches, debris flows and sediment pulses in rivers are common geomorphological processes but can destroy homes and infrastructure with little warning. Population pressure, thawing permafrost and other climatic affects will likely exacerbate this threat in the near future requiring new risk management strategies and monitoring tools.

In recent years, seismology has emerged as an efficient observational method to capture rapid mass movements and study their dynamics as well as variations in event activity. Multi-million cubic meter rock-ice avalanches like the 2025 event destroying parts of the village of Blatten, Switzerland, are often detected by national seismic networks primarily designed to monitor earthquake activity. Smaller events like rockfalls and debris flows require denser seismic networks with station spacing of a few kilometres or less. Nevertheless, their seismic signature is usually clear when seismic stations are close enough.

The straightforward detection of mass movements using seismic instrumentation has motivated new monitoring approaches. However, the challenge remains to automatically identify the seismic mass movement signature in continuous data streams given a wealth of other signals like anthropogenic noise and earthquakes, which are recorded at the same time and may mask the sought-after mass movement signals. Recent applications of machine learning algorithms have provided promising first results and allowed for mass movement detection in cases where empirical threshold-based triggering rules yield impermissible amounts of false positives.

Here we present a new approach to detect mass movements signals in continuous seismic catalogues. To tackle the challenge of algorithm transferability between sites with different seismic background noise we treat mass movement signals as anomalies given their catastrophic nature and rare occurrence. We use the isolation forest algorithm to quantify the degree of anomaly (‘anomaly score’) associated with any recorded signal. Using data from polar fjord systems, our results show that anomaly detection can efficiently reduce continuous seismic data sets to a handful of signals, which are likely related to rock avalanches and glacier break-off events. On smaller scales, anomaly scores can be processed to identify general characteristics of debris flow seismograms recorded near active torrents. The anomaly score approach thus facilitates systematically searching for large-scale mass movement seismograms in earthquake monitoring data and may be a stepping stone for flexible and transferable detection algorithms for monitoring and warning purposes.