Landslide susceptibility mapping using statistical and machine learning models (case study: Austria)

Landslides are a common natural hazard around the world, but the accuracy of the maps produced can be impacted by various adverse effects and uncertainties. Researchers have continuously sought to improve the accuracy of landslide susceptibility maps. This study aims to create a landslide susceptibility map for Austria using t-test and random forest models. Nine criteria for landslide occurrence, including elevation, slope, aspect, distance to drainages, distance to faults, distance to roads, land cover, lithology, and precipitation, were used. In the t-test model, the weight of each criterion was calculated using the t-statistical test and then combined with each other using the Simple Additive Weighting technique to draw the final landslide susceptibility map. The random forest model was trained using multiple decision trees and based on the landslide occurrence points and criterion layers, the relative weight of each layer was calculated, resulting in the final landslide susceptibility map. The receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) were used to compare the two models, with the results showing that the random forest model performed better with an AUC of 0.893, compared to the t-test model with an AUC of 0.852. The importance of different criteria was assessed, and it was found that slope and precipitation were the most important factors in the occurrence of landslides in both models. The results showed that both models have unique advantages in landslide susceptibility mapping. Accordingly, the higher accuracy of the random forest model, and the possibility of weighting the criteria and sub-criteria in the t-test model, make both models practical in this field.