Abstract:

Landslide pseudo-absence samples play an important role in applying statistical methods for landslide susceptibility mapping. It can constrain the overestimation of landslide susceptibility. The reliability of pseudo-absence samples is unknown in existing methods of pseudo-absence sample generation in landslide susceptibility mapping. The absence samples they generate may contain sites very susceptible to landslides but have not yet failed in the past simply because of the lack of triggering factors. These false absence samples lower the quality of absence samples and in turn lower the quality of the entire training sample set, and then decrease the accuracy of the mapping using statistical methods. In this article, a new method of measuring the reliability of pseudo-absence samples is proposed. The basic principle of the method is that the more similar the geographic environment is between two locations, the closer the geographic features are between them. Accordingly, the more similar the geographic environment of a location is with landslide presence samples, the more likely is the location prone to landslides and the less reliable is this location as an absence sample. According to the inference above, the reliability of a grid as an absence sample can be calculated by the similarity of its geographic environment with the typical environmental conditions under which landslides occurred. The Youfang catchment in southern Gansu Province—an area with high risk for landslides—was used as the study area to apply the proposed method and map the reliability of each grid chosen as pseudo-absence sample. The landslide initial zones in the Youfang catchment were used to validate the effectiveness of the proposed method. The results show that the mean reliability of grids in the landslide initial zones chosen as pseudo-absence samples is 0.26 and the reliability of more than 95% of the grids is lower than 0.5. This indicates that the proposed method of measuring the reliability of pseudo-absence samples is effective.

Key words: reliability of pseudo-absence samples, geographic environment, similarity, landslide susceptibility mapping