Counteracting flawed landslide data in statistically based landslide susceptibility modelling for very large areas: a national-scale assessment for Austria

The reliability of input data to be used within statistically based landslide susceptibility models usually determines the quality of the resulting maps. For very large territories, landslide susceptibility assessments are commonly built upon spatially incomplete and positionally inaccurate landslide information. The unavailability of flawless input data is contrasted by the need to identify landslide-prone terrain at such spatial scales. Instead of simply ignoring errors in the landslide data, we argue that modellers have to explicitly adopt their modelling design to avoid misleading results. This study examined different modelling strategies to reduce undesirable effects of error-prone landslide inventory data, namely systematic spatial incompleteness and positional inaccuracies. For this purpose, the Austrian territory with its abundant but heterogeneous landslide data was selected as a study site. Conventional modelling practices were compared with alternative modelling designs to elucidate whether an active counterbalancing of flawed landslide information can improve the modelling results. In this context, we compared widely applied logistic regression with an approach that allows minimizing the effects of heterogeneously complete landslide information (i.e. mixed-effects logistic regression). The challenge of positionally inaccurate landslide samples was tackled by elaborating and comparing the models for different terrain representations, namely grid cells, and slope units. The results showed that conventional logistic regression tended to reproduce incompleteness inherent in landslide training data in case the underlying model relied on explanatory variables directly related to the data bias. The adoption of a mixed-effects modelling approach appeared to reduce these undesired effects and led to geomorphologically more coherent spatial predictions. As a consequence of their larger spatial extent, the slope unit–based models were able to better cope with positional inaccuracies of the landslide data compared to their grid-based equals. The presented research demonstrates that in the context of very large area susceptibility modelling (i) ignoring flaws in available landslide data can lead to geomorphically incoherent results despite an apparent high statistical performance and that (ii) landslide data imperfections can actively be diminished by adjusting the research design according to the respective input data imperfections.