The uncertainty of landslide susceptibility prediction modeling: suitability of linear conditioning factors

For linear conditioning factors such as rivers, roads, and geological faults, existing studies mainly use buffer analysis in Geographic Information System to obtain discrete variables such as distance to rivers and roads. These discrete variables have random fluctuations and are sensitive to the errors of point or line elements, leading to a decrease of landslide susceptibility prediction (LSP) accuracy. This study proposes continuous conditioning factors such as river density and road density to improve the suitability of the linear factors. Xunwu County in China is taken as an example; 337 historical landslides and 12 conditioning factors are acquired. First, the distance to rivers and roads and other 10 conditioning factors together constitute the original factors of LSP. Second, the distance to rivers and distance to roads are replaced by the road density and river density, respectively, to constitute the improved factors. Third, based on the support vector machine (SVM), logistic regression (LR), and random forest (RF), original factor- and improved factor-based SVM, LR, and RF models are constructed for comparisons. Finally, the LSP uncertainty is evaluated. Results show that (1) the improved factor-based models have higher LSP accuracies than original factor-based models, indicating that density factors are more feasible than linear factors with more explicit physical meaning; (2) landslide susceptibility indexes distribution features indicate that improved factor-based models reduce the uncertainty of LSP; (3) spatial density factors do not reduce the importance of conditioning factors in both original factor- and improved factor-based models.