Uncertainty analysis of non-landslide sample selection in landslide susceptibility prediction using slope unit-based machine learning models

Abstract

The selection of non-landslide samples has a great impact on the machine learning modelling for landslide susceptibility prediction (LSP). This study presents a novel framework for studying the uncertainty of non-landslide samples selection on the LSP results through the slope unit-based machine learning models. In this framework, the non-landslide samples are randomly selected from the non-landslide areas by multiple times (N = 1, 10, 100, 500, 1000, 5000) to construct LSP models and calculate N types of landslide susceptibility indexes (LSIs). Afterwards, the statistical analysis is used to represent the uncertainty of LSIs under each non-landslide selection. The maximum probability analysis (MPA) is applied to reduce the uncertainty of non-landslide samples selection in LSP, which calculates the probability of N types of LSIs falling into very high, high, moderate, low and very low landslide susceptibility levels and selects the optimal landslide susceptibility level with the highest probability for each slope unit. Chongyi County in China is selected as the example, slope unit-based logistic regression (LR) and support vector machine (SVM) models are constructed with 16 conditioning factors. The area under the receiver operating features curve (AUC) and frequency ratio (FR) accuracy are used to evaluate the LSP performance. Results show that the N types of LSIs in each slope unit exhibit a normal distribution rather than one constant value. The uncertainties of LSIs caused by non-landslide samples selection are well represented by statistical analysis. The AUC values of slope unit-based LR/SVM models range from 0.714/ 0.711 (N = 1) to 0.787/0.775 (N = 5000) and increase to 0.867/0.848, meanwhile, the FR accuracies range from 0.772/ 0.763 (N = 1) to 0.815/0.826 (N = 5000) and increase to 0.843/0.861 by the MPA method. It is concluded that some more scientific and accurate landslide susceptibility results are obtained by selecting non-landslide samples multiple times and using the MPA method.