Comparison of expert, deterministic and Machine Learning approach for landslide susceptibility assessment in Ljubovija Municipality, Serbia

Landslide Susceptibility Assessment is becoming a very productive research area, wherein different modeling approaches are practiced to delineate zones of the high-low likelihood of landslide occurrence. However, there is no strong consensus on which approach is the most adequate. The reason behind the lack of the general view on the performance of different approaches could be partially explained by the particularity of each study. To evaluate the efficiency of different approaches they need to be applied under the same conditions for the same study area. Herein, we examined three different approaches, including expert, deterministic and Machine Learning, on the example of Ljubo-vija Municipality in western Serbia. The study area has been known as susceptible to landslides, and represents good ground for assessing the chosen methods. It is represented by complex geology, prone to landslides that are commonly hosted in thick weathering crust of Paleozoic formations, composed of schists and meta-sediments. Under extreme triggering conditions, such as the one that unfolded in May 2014, these thick weathering crusts saturate, and give way to a variety of landslide and flash-flood processes that we will be focusing on in this study. The application of the expert-approach, through Analytical Hierarchy Process provided a rough assessment map. The deterministic model, which couples simple infinite slope and hydrological model, provided us with lower quality results, when compared to the expert-based one. This could be explained by the assumptions used in the model are too simplistic to generically model a wide range of landslide typology. Finally, Machine Learning approach, using the Random Forest algorithm, provided significantly better results and showed that it can cope with versatile landslide typology over larger scales. Its AUC performance is about 0.75 which is considerably outperforming the AUC values of the other two models, which were up to 0.55, i.e. at the level of random guess.