Spatial correlation between landslides and geotechnical factors using Random Forest and SHAP

The activation as well as the consequences of landslides are difficult to predict, as they depend on factors characterized by large variability and uncertainties. The aim of this study is to establish a correlation between geological, geotechnical and geomorgpohlogical characteristics and the spatial distribution of recorded landslides.

The study area is located in the southwestern (SW) part of the island of Cyprus, covering an area of 552km². During the past years, more than 1800 landslides, active and inactive (dormant and relict), have been recorded within this area through detailed mapping based on field observations, rendering the area an ideal test bed. At the beginning of this research study, all recorded landslides were digitized in raster format. Consequently, the study area was partitioned into 15 x 15m size cells having three classes: no landslides, inactive landslides and active landslides. Additionally, regarding the geological aspect, polygons encompassing 100% rock mass formations within recorded landslides were categorized as rock mass landslides, while the rest were characterized as landslides in argillaceous (soft rock and soil) materials. A series of correlation analyses were conducted using the Random Forest and SHAP (Shapley Additive explanation) methods.

Considering the outcomes of the Random Forest method in argillaceous materials, it turns out that the most important factors for both active and inactive landslides are the Plasticity Index (PI) and the clay fraction (CF), followed by the factors associated with the geomorphology and the bedding structure (e.g. slope angle and bedding dip). The ranking results for inactive and active landslides in rock mass show that the most important factor is the Uniaxial Compressive Strength (UCS), followed by the Geological Strength Index (GSI). Furthermore, the orientation (azimuth) difference between slope and bedding dip (dip direction difference) appears to be more important than the slope angle.

Similar ranking results were obtained using the SHAP method for argillaceous materials. Regarding the contribution of each factor in the inactive landslides, it appears that the PI and the slope angle increase proportionally to the possibility of landslide occurrence, while the CF does not exhibit a specific trend. Regarding the dip direction difference, small values contribute more to the occurrence of landslides. The active landslides show a similar picture, but with the CF exhibiting a stronger correlation than in the case of inactive landslides. According to the SHAP analysis for rock mass, the parameters of importance in both inactive and active landslides are UCS and GSI, followed by the slope angle and the dip direction difference.

This research was funded by the European Commission (Marie Sklodowska-Curie Actions, Hybland-Society and Enterprise panel, Project No.: 101027880).