Identifying the essential conditioning factors of landslide susceptibility models under different grid resolutions using hybrid machine learning: A case of Wushan and Wuxi counties, China

This study attempts to identify the essential conditioning factors of landslides to increase the predictive ability of landslide susceptibility models and explore the effects of different grid resolutions (i.e., 30 m, 300 m, 1000 m, 2000 m, and 3000 m) on landslide susceptibility assessment. Firstly, taking Wushan and Wuxi counties in Chongqing as an example, a geospatial dataset comprising 1137 historical landslide locations and preliminary 28 conditioning factors was randomly divided into training (70%) and testing (30%) datasets at each grid resolution. Secondly, spearman correlation coefficient (SCC), recursive feature elimination (RFE) and their combination (SCC-RFE) were chosen to identify the essential conditioning factors out of 28 original factors at five grid resolutions. Subsequently, random forest (RF) model was used to construct landslide susceptibility model with the original and essential conditioning factors, respectively. Finally, the reasonableness of the essential conditioning factors was verified by comparing the receiver operation characteristic (ROC) curves (AUC) and other statistical signifiers in multiple grid resolutions. Results show that: (1) Average annual rainfall, elevation, lithology and POI have a significant impact on the occurrence of landslides, while NDVI and land cover has little effect on the occurrence of landslides in Wushan and Wuxi counties. (2) The primary essential factors (i.e., elevation, rainfall) are less affected by the grid resolution, while the subdominant factors (i.e., DEM-derived factors, human activity factors) are strongly influenced. (3) SCC-RFE-RF model performs best with the screened essential conditioning factors at a grid resolution smaller than 2000 m, which indicates choosing the essential conditioning factors or optimum grid resolution can guarantee greater accuracy of landslide susceptibility models. This study provides a reference for future analysis in selecting landslide conditioning factors and grid resolutions.