Conventional data-driven landslide susceptibility models may only tell us half of the story: Potential underestimation of landslide impact areas depending on the modeling design

Abstract

Data-driven landslide susceptibility models (LSM) have been widely used to spatially predict areas likely affected by landslides. Even though such approaches are used to estimate the spatial likelihood of landslide occurrence, they are frequently trained with observations that rather represent landslide release zone conditions without explicitly considering the potential landslide downslope propagation into flat terrain. Landslide runout models, instead, are designed to route the downslope propagation of masses from given release zones but are seldom coupled with landslide release susceptibility models. This research compares the prediction pattern of three conceptually different models: (i) a model trained on landslide release observations only (Model R), (ii) a model trained on landslide body observations only (Model B), and, (iii) a model that explicitly couples landslide-release susceptibility and the subsequent runout zones (Model R + R). The aim is to explore and interpret the implications of differences in the susceptibility classes and associated model performance metrics. In this context, the interpretation focuses on the results observed for gentle terrain that is located downslope of the prevalent steeper hillsides. These partly built-up areas (i.e., buildings, linear infrastructures) are frequently within the reach of open-hillslope debris flows and, therefore of particular interest from a hazard and risk viewpoint.The study area for this research is the 54 km2 catchment Córrego Dantas, situated in the mountainous region of Rio de Janeiro, SE Brazil. After a heavy rainfall event in 2011, an inventory composed of 313 landslides was mapped and used as input for this work. First, a landslide release susceptibility model (Model R) based on an established data-driven method (Generalized Additive model, GAM) was created to explore the spatial likelihood of landslide release. Model R was then combined with r.randomwalk to compute the propensity of downslope regions to be affected by the released landslides, resulting in an integrated landslide impact susceptibility score (Model R + R). In order to achieve this, a constrained top-down random walk approach was used. Furthermore, a GAM model built upon observations from the entire landslide body (Model B) was used for comparison. Results reveal a potential underestimation of the landslide threat in flatter areas. This tendency is most pronounced for the release-trained model (Model R). The runout impact susceptibility scores produced by the random walk (Model R + R) were observed to better represent the actual impact threat of the whole geomorphological phenomena. From a statistical viewpoint, also the simpler model (Model B) exhibited a satisfactory performance to predict the full landslide area that includes release and deposit zones. The outcomes of this research could be of particular interest to modelers aiming to model landslide susceptibility of regions characterized by long-runout hillslope debris flows and for studies focusing on landslide impact and risk.