You are using an old browser with security vulnerabilities and can not use the features of this website.
Timescale: Oct.2015 – Sept.2018
Prof. Dr. Oliver Korup, University of Potsdam
Prof. Dr. Jürgen Kurths, Potsdam Institute for Climate Impact Research (PIK)
Quantitative studies that aim at predicting where and when landslides occur in a given area of interest have taken centre stage in the research literature on mass movements. Over the last two decades, researchers have adopted multivariate statistical tools from the fields of data mining and machine learning to predict the spatial and temporal susceptibility to landslides fromdetailed inventories. These data sets are matrices containing n observed landslides with m characteristics each, including e. g., geographic position, landslide size, topography, rock type, soil and groundwater conditions, distance from fault lines, river or road networks. Some of the more widely used methods such as multivariate logistic regression, artificial neural networks, or Bayesian Weights-of-Evidence have routinely produced prediction success rates of >80% (Korup & Stolle, 2014). This optimistic prospect in spatial landslide prediction is at odds with recent re-assessments of global landslide damage: Nearly 32,000 people have died because of landslides between 2004 and 2010 alone, shadowing previous estimates, and inviting more reliable methods of landslide early warning.
Some of the key problems that limit the predictive capacity of these models include overfitting, poorly constrained data quality, and missing data. Further uncertainty comes from insufficiently time-stamped landslide data, let alone the many transient causes and triggers that compromise prediction of landslides in times of climate and global change. Hence, landslide prediction mainly deals with mapping susceptibility under static scenarios. Dynamic prediction of the timing of landslides, on the other hand, and whether they are clustered in time has remained challenging (Korup et al., 2012).
Objectives and Methods
To address this shortcoming, the key goals of this project are to
We will work on these objectives using several moderate to large landslide inventories. We will compare inventories that are linked to historic trigger events such as strong earthquakes or rainstorms with inventories that feature mostly prehistoric landslides of unknown trigger and age to test the applicability of various prediction algorithms. The expected outcome of this project is a statistically robust, yet versatile and applicable tool for landslide prediction under uncertainty and non-stationary boundary conditions. We will provide an objective ranking and benchmarking of the most popular landslide prediction methods and their potential pitfalls when applied to non-stationary time series with uncertainties. This project will work with various landslide inventories from the Himalayan and the Alpine regions.
Ugur Öztürk is based at the research team “Geohazards” of the University of Potsdam.