Landslide susceptibility mapping by using various selection strategies of landslide conditioning factors and XGBoost

Landslides are one of the most common and dangerous natural hazards that occur worldwide. Their occurrence may cause material losses and even death. Therefore, it is important to incorporate any mitigation action to ensure safety. One of the first steps can be generation of the landslide susceptibility maps which portrays the terrain probability to landsliding. There are numerous methods for creating landslide susceptibility maps, and machine learning methods are recently widely used. Therefore, in this study, the XGBoost machine learning algorithm was also implemented.

However, many scientists reported that the most critical step in any prediction model is the selection of the most appropriate features. In the case of landslide susceptibility modelling, they are called landslide conditioning factors (LCFs). LCFs are selected based on expert knowledge, literature review, or based on various statistical approaches for feature selection. Among statistical approaches, Symmetrical Uncertainty (SU), Analysis of variance (ANOVA) or Pearson correlation index (PI) can be applied.

Therefore, the objective of this experiment was to evaluate the effect of the feature selection method on the accuracy of the maps of susceptibility to landslides. For the experiment, two various areas of interest have been evaluated in the area of Polish Flysch Carpathians. Also, various accuracy measures were used to evaluate model performance among them Area Under the Curve (AUC), precision, Recall, and F1-score.

Accuracy measures indicated that the best method for feature selection is  Pearson correlation (F1 score on the level of 77.2 % and 79.4 %) for both study cases, however, the difference between these feature selection methods are not significant.