Modelling rules of landslide susceptibility prediction considering the suitability of linear environmental factors and different machine learning models

Abstract

For linear environmental factors such as river, road and geological fault networks, buffer analysis in GIS is commonly used to extract the buffer distances to the river and/or road networks. However, the line distances are discrete variables with random fluctuations of different grid sizes and are more sensitive to the errors of point and/or line elements, leading to a reduction in the accuracy of landslide susceptibility prediction (LSP). This study aims to use continuous environmental factors, such as the spatial density of river and road networks, to improve the suitability of linear environmental factors. Taking An'yuan County of Jiangxi Province as an example, 14 environmental factors, such as elevation, topographic relief, distances to river and road networks (original factors), are selected. Then, the two linear environmental factors of distances to river and road networks are improved to river and road density (improved factors). Based on machine learning models such as logistic regression (LR), multilayer perceptron (MLP), support vector machine (SVM) and C5.0 decision tree, original factor-based and improved factor-based machine learning models are built to carry out the LSP. The receiver operating characteristic (ROC) curves and the distribution characteristics of landslide susceptibility indexes are used to evaluate the LSP modelling rules. The results show that ① the LSP accuracy of the improved factor-based models are higher than those of the original factor-based models, indicating that the spatial density is more suitable for LSP; ② the C5.0 model has the best LSP performance among the four machine learning models, followed by the SVM, MLP and LR models; and ③ river and road factors are of great significance for landslide evolution, and their importance does not decrease underimproved factor-based machine learning models.

Keywords

• landslide susceptibility prediction
• linear environmental factor
• spatial density
• machine learning
• geographic information system