You are using an old browser with security vulnerabilities and can not use the features of this website.

Here you will see how you can easily upgrade your browser.

PhD-Project I8: Learning landslide triggers from complex networks

Timescale: Oct.2018 – Sept.2021


Prof. Dr. Oliver Korup, University of Potsdam;

Prof. Dr. Jürgen Kurths, PIK Potsdam


A large number of all landslides worldwide are triggered by regional disturbances such as rainstorms or strong earthquakes. Such triggers can mobilise between tens to more than hundred thousand of slope failures in a matter of several minutes to several days, exclusive of any reactivated slope failures. Hence, the success of predicting such regional landsliding episodes hinges on how well we can detect the timing and magnitude of their triggers. While earthquake prediction (or forecasting) remains challenging, more progress has been made in the field of forecasting rainstorms. Novel methods such as event synchronisation based on spatially embedded complex networks, for example, have helped to objectively map out the trajectories of extreme precipitation brought by tropical cyclones. This approach was at the core of NatRiskChange’s Project P1 and has delivered promising results that are largely consistent with eye-of-the-storm tracks obtained from satellite imagery at the scale of entire nations or island arcs (in our case the Japanese archipelago and the surrounding seas).

How these rainfall patterns map to the distribution of nearly 400,000 landslides documented in Japan, remains to be examined, however. We propose to develop methods to identify from existing landslide catalogues those landslides that have the highest likelihood of having had a rainfall trigger, and thus separate them from co-seismic landslides. Part of this approach rests on new insights into landslide characteristics largely indicative of seismic triggers (for example, directivity effects also investigated during the Kumamoto Task Force in NatRiskChange). Also, the new extreme rainfall trajectories offer uncharted opportunities for detecting probabilistically those slope failures that originated during rainstorms, taking in effects of slope exposure, hillslope hydrology, and land-use patterns. Reliably distinguishing meteorological from co-seismic landslides is important for modern landslide prediction and especially for robustly informing vulnerability, and eventually, risk appraisals. We will develop statistical tests for distinguishing these two types of landslides. For a deeper understanding of the occurrences of such rainfalls, further meteorological parameters, in particular wind field, will be included in a network-based analysis. From the practical landslide forecasting side, preliminary results also support the idea that the most extreme rainstorms are not necessarily responsible for the bulk of landslide volume, let alone the related damage. Our study will add new insights to this proposition and thus create highly demanded data on the magnitude and frequency of landslides stratified by their major triggers.

Objectives and Methods

  • to separate rainfall- from earthquake-triggered landslides using meteorological and topographic predictors such as extreme rainfall trajectories and seismic directivity effects (developed and derived in NatRiskChange Phase I)
  • Methods: complex spatially embedded networks, probabilistic classification methods using Bayesian Reasoning and Machine Learning