Precipitation and Erosion Threshold for Quick Clay Landslides Using Machine Learning Models

Abstract

Quick clay landslides may have major consequences on human civilization and are responsible for some of the most damaging natural disasters in Norway. Quick clay hazard zones in Norway are mapped as either low-, medium-, or high-hazard, but are not updated based on any regular monitoring. The main goal of the study was to investigate if precipitation and erosion data gathered from previous quick clay landslides in Oslo and Viken in southeastern Norway could be used to create a threshold for when conditions of quick clay reach the critical levels that may trigger a landslide event. The ultimate goal is that these thresholds can be used as part of a method to periodically update the hazard zone categories. The method used to create thresholds was based on machine learning models, including two different ensemble models (RUSBoosted decision trees and Bagged decision trees) and two Support Vector Machine (SVM) models (Cubic and Quadratic kernels). The models’ ability to classify landslides correctly were evaluated using area under the receiver operating characteristics curve (AUC) and confusion matrix to measure the false-negative and the false positive rates. The results showed that the Quadratic SVM model and the Bagged decision trees had the highest AUC (0.74, 0.85) and the lowest false-negative rate (54.5 %, 63.6 %) of the models trained when the models were trained with a combination of precipitation and erosion data. Training with only precipitation data did not change the results much, though a minor improvement was seen when including erosion data. The high false-negative rates suggest that the method as used here is unsuitable as part of a monitoring system. One major problem is lack of erosion data, and an improvement of the method will probably require yearly gathering of erosion data in addition to testing of other predictor variables.