Identifying the essential influencing factors of landslide susceptibility models based on hybrid-optimized machine learning with different grid resolutions: a case of Sino-Pakistani Karakorum Highway.

Abstract

This study attempts to explore the essential influencing factors of landslides and explores the effects of different datasets on landslide susceptibility mapping (LSM) at six grid resolutions (i.e., 10 m, 30 m, 300 m, 1000 m, 2000 m, and 3000 m). Firstly, the geospatial dataset of 21 influencing factors was extracted from 1847 historical landslide InSAR (Interferometric Synthetic Aperture Radar) points, which were taken as a sample for the Sino-Pakistani Karakorum Highway. Secondly, Spearman correlation coefficient (SCC), random forest feature selection (RFFS), and their combinations (SCC-RFFS) were selected at different grid resolutions to identify the essential influencing factors from the 21 original factors. A random division into training set (70%) and test set (30%) was performed. Then, the LSM models for the original influencing factors and the selected influencing factors were constructed separately using machine learning models. Finally, the reasonableness of the essential influencing factors was verified by comparing the accuracy of the models under different grid resolutions. The results show that (1) relief degree of land surface (RDLS), SPI, and rainfall have significant effects on landslide occurrence. (2) The primary elements (i.e., RDLS, slop, rainfall) are less affected by the grid resolution, while the secondary elements (TWI) are more affected by the grid resolution. (3) At 30 m, the SCC-RFFS-RF model can get the highest landslide susceptibility model accuracy. The prediction will also provide scientific guidance for the allocation of land resources on a regional and global scale, and minimize the human and economic costs along the highway, while ensuring safe highway operations.