Characterisation of precursory seismic activity towards early warning of landslides via semi-supervised learning.

This study demonstrates that machine learning from seismograms, obtained from commonly deployed seismometers, can identify the early stages of slope failure in the field. Landslide hazards negatively impact the economy and public through disruption, damage of infrastructure and even loss of life. Triggering factors leading to landslides are broadly understood, typically associated with rainfall, geological conditions and steep topography. However, early warning at slope scale of an imminent landslide is more challenging. Through semi-supervised learning for seismic event detection from continuous seismic recordings over a period of about 10 years, we demonstrate that timely landslide induced displacement prediction is possible, providing the basis for landslide early warning systems. Our proposed methodology detects and characterises seismic precursors to landslide events making use of seismic recordings near an active slow moving earth slide-flow using a semi-supervised Siamese network. This data driven methodology identifies increase in microseismicity, and the change in the frequency spectrum of that microseismicity which identify key stages prior to a failure: 'rest', 'precursor' and 'active'. Due to the semi-supervised nature of Siamese networks, the methodology is adaptable to discovering new types of distinct events, making it an ideal solution for precursor detection at new sites.