Modelling Uncertainties and Sensitivity Analysis of Landslide Susceptibility Prediction under Different Environmental Factor Connection Methods and Machine Learning Models

Abstract

The utilization of different connection methods between landslides and environmental factors introduces uncertainty in landslide susceptibility prediction (LSP). Investigating and identifying the characteristics of this uncertainty and determining more suitable connection methods are of significant importance for LSP. This study uses original 12 environmental factors data as well as calculated data from five connection methods, namely, probability statistics (PS), frequency ratio (FR), information volume (IV), index of entropy (IOE), and weight of evidence (WOE), as model input variables. Then, four machine learnings logistic regression (LR), Bayesian networks (BN), support vector machines (SVM) and C5.0 Decision Trees (C5.0 DT) are combined with the calculated data and the original data to create 24 unique combinations of connection methods and models. Under these 24 combinations, the uncertainty analysis of LSP modeling is conducted, using Huichang County of China as an example. The analysis entails accuracy assessment, statistical analysis of landslide susceptibility indexes (LSIs), distribution patterns of LSIs and sensitivity analysis of the two uncertainty issues. The results show that: 1) LSP accuracies of the FR-, IV- and IOE-based models are comparable, but are lower than those of the WOE-based models, with those of the PS-based models being the worst. WOE has better nonlinear connection performance than the other methods. 2) LSP accuracies of individual models are slightly lower than those of coupled models, but their modeling efficiencies are higher than those of coupled models. 3) The machine learning is more sensitive than the connection method for LSP. In short, WOE-C5.0 DT has the lowest LSP uncertainty while a single machine learning can produce satisfied LSP results with high modelling efficiency.