Implementing landslide path dependency in landslide susceptibility modelling

Abstract

Landslide susceptibility modelling—a crucial step towards the assessment of landslide hazard and risk—has hitherto not included the local, transient effects of previous landslides on susceptibility. In this contribution, we implement such transient effects, which we term “landslide path dependency”, for the first time. Two landslide path dependency variables are used to characterise transient effects: a variable reflecting how likely it is that an earlier landslide will have a follow-up landslide and a variable reflecting the decay of transient effects over time. These two landslide path dependency variables are considered in addition to a large set of conditioning attributes conventionally used in landslide susceptibility. Three logistic regression models were trained and tested fitted to landslide occurrence data from a multi-temporal landslide inventory: (1) a model with only conventional variables, (2) a model with conventional plus landslide path dependency variables, and (3) a model with only landslide path dependency variables. We compare the model performances, differences in the number, coefficient and significance of the selected variables, and the differences in the resulting susceptibility maps. Although the landslide path dependency variables are highly significant and have impacts on the importance of other variables, the performance of the models and the susceptibility maps do not substantially differ between conventional and conventional plus path dependent models. The path dependent landslide susceptibility model, with only two explanatory variables, has lower model performance, and differently patterned susceptibility map than the two other models. A simple landslide susceptibility model using only DEM-derived variables and landslide path dependency variables performs better than the path dependent landslide susceptibility model, and almost as well as the model with conventional plus landslide path dependency variables—while avoiding the need for hard-to-measure variables such as land use or lithology. Although the predictive power of landslide path dependency variables is lower than those of the most important conventional variables, our findings provide a clear incentive to further explore landslide path dependency effects and their potential role in landslide susceptibility modelling.