Stefan Steger

Evaluating the Effect of Modelling Methods and Landslide Inventories Used for Statistical Susceptibility Modelling

Landslide susceptibility maps can be elaborated using a variety of methodological approaches. This study investigates quantitative and qualitative differences between two statistical modelling methods, taking into account the impact of two different response variables (landslide inventories) for the Rhenodanubian Flysch zone of Lower Austria. Quantitative validation of the four generated susceptibility maps is conducted by calculating conventional accuracy statistics for an independent random landslide subsample. Qualitative geomorphic plausibility is estimated by comparing the final susceptibility maps with hillshades of a high resolution Airborne Laser Scan Digital Terrain Model (ALS-DTM). Spatial variations between the final susceptibility maps are displayed by difference maps and their densities. Although statistical quality criterions reveal similar qualities for all maps, difference maps and geomorphic plausibility expose considerable differences between the maps. Given that, this conclusion could only be drawn by evaluating additionally the geomorphic plausibility and difference maps. Therefore, we indicate that conventional statistical quality assessment should be combined with qualitative validation of the maps.

The influence of systematically incomplete shallow landslide inventories on statistical susceptibility models and suggestions for improvements

Complete landslide inventories are rarely available. The objectives of this study were to (i) elaborate the influence of incomplete landslide inventories on statistical landslide susceptibility models and to (ii) propose suitable modelling strategies that can reduce the effects of inventory-based incompleteness. In this context, we examined whether the application of a novel statistical approach, namely mixed-effects models, enables predictions that are less influenced by such inventory-based errors.The study was conducted for (i) an area located in eastern Austria and (ii) a synthetically generated data set. The applied methodology consisted of a simulation of two different inventory-based biases and an in-depth evaluation of subsequent modelling results. Inventory-based errors were simulated by gradually removing landslide data within forests and selected municipalities. The resulting differently biased inventories were introduced into logistic regression models while we considered the effects of including or excluding predictors that are directly related to the respective inventory-based bias. Mixed-effects logistic regression was used to account for variation that was due to an inventory-based incompleteness.The results show that most erroneous predictions, but highest predictive performances, were obtained from models generated with highly incomplete inventories and predictors that were able to directly describe the respective incompleteness. An exclusion of such bias-describing predictors led to systematically confounded relationships. The application of mixed-effects models proved valuable to produce predictions that were least affected by inventory-based errors.This paper highlights that the degree of inventory-based incompleteness is only one of several aspects that determine how an inventory-based bias may propagate into the final results. We propose a four-step procedure to deal with incomplete inventories in the context of statistical landslide susceptibility modelling.

Landslide Susceptibility Mapping at National Scale: A First Attempt for Austria

Numerous publications that addressing landslide susceptibility were published over the past decades, also due to an increasing demand of spatial information regarding potentially endangered areas. However, studies that provide an overview on landslide susceptibility at national scale are still scarce. This research presents a first attempt to generate a national scale landslide susceptibility map for Austria based on statistical techniques. Binary logistic regression has been applied to delineate susceptible areas using three different predictor sets. The initial predictor set relates to topographic variables only (model A), and was gradually expanded with the factors geology (model B) and land cover (model C). The Area Under the Receiver Operating Characteristic Curve (AUROC) was used to validate the predictions by means of a k-fold cross-validation. The obtained acceptable prediction performances (mean AUROC of model A: 0.76, B: 0.81 and C: 0.82) suggest a relatively high predictive performance of all models. However, during this study, several limitations of the conducted analysis (e.g. limited landslide data, bias propagation, overoptimistic performance estimates) became evident. The main drawbacks and further steps towards a more reliable representation of landslide susceptibility at national scale are discussed.

Integration of Geometrical Root System Approximations in Hydromechanical Slope Stability Modelling

Spatially distributed physically based slope stability models are commonly used to assess landslide susceptibility of hillslope environments. Several of these models are able to account for vegetation related effects, such as evapotranspiration, interception and root cohesion, when assessing slope stability. However, particularly spatial information on the subsurface biomass or root systems is usually not represented as detailed as hydropedological and geomechanical parameters. Since roots are known to influence slope stability due to hydrological and mechanical effects, we consider a detailed spatial representation as important to elaborate slope stability by means of physically based models. STARWARS/PROBSTAB, developed by Van Beek (2002), is a spatially distributed and dynamic slope stability model that couples a hydrological (STARWARS) with a geomechanical component (PROBSTAB). The infinite slope-based model is able to integrate a variety of vegetation related parameters, such as evaporation, interception capacity and root cohesion. In this study, we test two different approaches to integrate root cohesion forces into STARWARS/PROBSTAB. Within the first approach, the spatial distribution of root cohesion is directly related to the spatial distribution of land use areas classified as forest. Thus, each pixel within the forest class is defined by a distinct species related root cohesion value where the potential maximum rooting depth is only dependent on the respective species. The second method represents a novel approach that approximates the rooting area based on the location of single tree stems. Maximum rooting distance from the stem, maximum depth and shape of the root system relate to both tree species and external influences such as relief or soil properties. The geometrical cone-shaped approximation of the root system is expected to represent more accurately the area where root cohesion forces are apparent. Possibilities, challenges and limitations of approximating species-related root systems in infinite slope models are discussed.

The Challenge of “Trivial Areas” in Statistical Landslide Susceptibility Modelling

Open image in new window Landslide susceptibility maps are frequently produced by fitting multiple variable statistical models that generate a relationship between a binary response variable (presence and absence of past landslides) and a set of predisposing environmental factors. Within this study, we investigated the hypothesis that an inclusion of a high portion of “trivial areas” (e.g. flat areas) affects modelled relationships, quantitative validation results and the appearance of the final maps. This assumption was tested by systematically comparing logistic regression models that were based on data sets which ignored respectively included a high portion of “trivial areas”. Modelled relationships were evaluated by estimating odds ratios for all predictors. The Area under the Receiver Operating Characteristic Curve (AUROC) provided information on the prediction skill of each model. This performance measure was assessed by applying non-spatial and spatial partitioning techniques. Each analysis was additionally performed with artificial samples to confirm our observations. The results showed that the delineation of the study area affected modelled relationships and consequently the spatial pattern of landslide susceptibility maps as well. AUROC values confirmed that the apparent prediction skill of a model may increase whenever a high portion of easily classifiable areas (e.g. flat area) is included. Therefore we concluded that an interpretation of modelled relationships and prediction skills should always consider the spatial extent to which the respective statistical landslide susceptibility analysis was carried out. The apparent prediction performance of a geomorphic meaningless model can be enhanced by including a high portion of easily classifiable areas.