Abstract
This work aims at evaluating the sensitivity of landslide susceptibility mapping (LSM) to sampling design in lithologically-heterogeneous areas. We hypothesize that random sampling of the landslide absence data in such areas can be biased by statistical aggregation of the explanatory variables, which impact the model outputs. To test this hypothesis, we train a Random Forest (RF) model in two different domains, as follows: (1) in lithologically heterogeneous areas, and (2) in lithologically homogeneous domains of the respective areas. Two heterogeneous areas are selected in Japan (125 km(2)) and Romania (497 km(2)), based on existing landslide inventories that include 371 and 577 scarps, respectively. These areas are divided into two, respectively three domains, defined by lithological units that reflect relatively homogeneous topographies. Fourteen terrain attributes are derived from a 30 m SRTM digital elevation model and employed as explanatory variables. Results show that LSM is sensitive to a random sampling of the absence data in lithologically heterogeneous areas. Accuracy measures improve significantly when sampling and LSM are conducted in lithologically homogeneous domains, as compared to heterogeneous areas, reaching an increase of 9% in AUC and 17% in the Kappa index.