Identifying landslides from massive seismic data and machine learning: the case of the European Alps

Recent large landslides in many parts of the World (Nuugaatsiaq, Greenland; Taan-Tyndall, US; Culluchaca, Peru) as well as the increase in the frequency of mass movements in the European Alps (e.g. collapse of the Drus, Mont Blanc Massif, France; Piz Cengalo, Switzerland) revealed the threat of such events to human activity. Seismology provides continuous recordings of landslide activity at long distances. The objective of this work is to present a method to identify and construct instrumental landslide catalogs from massive seismological data. The method is developed and applied for the period 2000-2022 at the scale of the European Alps (~ 900 x 300 km). 

The detection method applied to the seismological observations consists of computing the energy of the signal between 2 and 10 Hz. Then, a supervised Random Forest classifier is trained to identify the source of the event (earthquakes or landslides). To implement  the seismological detection and identification methods, we compiled a database of 65 landslides and 4515 earthquakes (of MLv > 0.1). The dataset is composed of 2221 seismological traces of landslides and 17353 traces of earthquakes. Tests of the Random Forest identification method gave us a rate of good identification of around 100% for landslides and 96% for earthquakes. Tests on continuous data of the 65 days of the reference landslide events allow finding 235 new landslides including 61 over 65 reference events.

The trained model is then applied on continuous seismic data (~ 400 stations) acquired over the European Alps since 2000. To reject as many noise detections as possible, a first sorting of all detections is performed by looking at SNR ratio, number of stations involved in the detection in a small area and probability scores given by the Random Forest. The instrumental catalog is composed of ~ 183.000 possible landslides. In order to review the catalog, reject possible false detections and interpret the inventory, we developed a localization method. A first order of the localization is given by the spatial clusters of seismological stations that have detected the landslide signals. Then, to refine localizations, we compute travel times from seismological stations to all points of the area with a fast marching method and we perform the inversion by using NonLinLoc software (Lomax et al. 2000). The final landslide instrumental catalog will be presented and discussed.