Data-driven landslide susceptibility models formally integrate spatial landslide information with explanatory environmental variables that describe predisposing factors of slope instability. Well-performing models are commonly utilized to identify landslide-prone terrain or to understand the causes of slope instability. In most cases, however, the available landslide data is affected by spatial biases (e.g. underrepresentation of landslides far from infrastructure or in forests) and does therefore not perfectly represent the spatial distribution of past slope instabilities. Literature shows that implications of such data flaws are frequently ignored. This study was built upon landslide information that systematically relates to damage-causing and infrastructure-threatening events in South Tyrol, Italy (7400 km2). The created models represent three conceptually different strategies to deal with biased landslide information. The aims were to demonstrate why an inference of geomorphic causation from apparently well-performing models is invalid under common landslide data bias conditions (Model 1), to test a novel bias-adjustment approach (Model 2) and to exploit the underlying data bias to model areas likely affected by potentially damaging landslides (Model 3; intervention index), instead of landslide susceptibility. The study offers a novel perspective on how biases in landslide data can be considered within data-driven models by focusing not only on the process under investigation (landsliding), but also on the circumstances that led to the registration of landslide information (data collection effects). The results were evaluated in terms of statistical relationships, variable importance, predictive performance, and geomorphic plausibility. The results revealed that none of the models reflected landslide susceptibility. Despite partly high predictive performances, the models were unable to create geomorphically plausible spatial predictions. The impact-oriented intervention index, however, enabled to identify damage-causing landslides with high accuracy. We conclude that the frequent practice of inferring geomorphic causation from well-performing models without accounting for data limitations is invalid.

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.

There is unanimous agreement that a precise spatial representation of past landslide occurrences is a prerequisite to produce high quality statistical landslide susceptibility models. Even though perfectly accurate landslide inventories rarely exist, investigations of how landslide inventory-based errors propagate into subsequent statistical landslide susceptibility models are scarce. The main objective of this research was to systematically examine whether and how inventory-based positional inaccuracies of different magnitudes influence modelled relationships, validation results, variable importance and the visual appearance of landslide susceptibility maps. The study was conducted for a landslide-prone site located in the districts of Amstetten and Waidhofen an der Ybbs, eastern Austria, where an earth-slide point inventory was available. The methodological approach comprised an artificial introduction of inventory-based positional errors into the present landslide data set and an in-depth evaluation of subsequent modelling results. Positional errors were introduced by artificially changing the original landslide position by a mean distance of 5, 10, 20, 50 and 120 m. The resulting differently precise response variables were separately used to train logistic regression models. Odds ratios of predictor variables provided insights into modelled relationships. Cross-validation and spatial cross-validation enabled an assessment of predictive performances and permutation-based variable importance. All analyses were additionally carried out with synthetically generated data sets to further verify the findings under rather controlled conditions. The results revealed that an increasing positional inventory-based error was generally related to increasing distortions of modelling and validation results. However, the findings also highlighted that interdependencies between inventory-based spatial inaccuracies and statistical landslide susceptibility models are complex. The systematic comparisons of 12 models provided valuable evidence that the respective error-propagation was not only determined by the degree of positional inaccuracy inherent in the landslide data, but also by the spatial representation of landslides and the environment, landslide magnitude, the characteristics of the study area, the selected classification method and an interplay of predictors within multiple variable models. Based on the results, we deduced that a direct propagation of minor to moderate inventory-based positional errors into modelling results can be partly counteracted by adapting the modelling design (e.g. generalization of input data, opting for strongly generalizing classifiers). Since positional errors within landslide inventories are common and subsequent modelling and validation results are likely to be distorted, the potential existence of inventory-based positional inaccuracies should always be considered when assessing landslide susceptibility by means of empirical models.

In recent decades, data-driven landslide susceptibility models (DdLSM), which are based on statistical or machine learning approaches, have become popular to estimate the relative spatial probability of landslide occurrence. The available literature is composed of a wealth of published studies and that has identified a large variety of challenges and innovations in this field. This review presents a comprehensive up-to-date overview focusing on the topic of DdLSM. This research begins with an introduction of the theoretical aspects of DdLSM research and is followed by an in-depth bibliometric analysis of 2585 publications. This analysis is based on the Web of Science, Clarivate Analytics database and provides insights into the transient characteristics and research trends within published spatial landslide assessments. Following the bibliometric analysis, a more detailed review of the most recent publications from 1985 to 2020 is given. A variety of different criteria are explored in detail, including research design, study area extent, inventory characteristics, classification algorithms, predictors utilized, and validation technique performed. This section, dealing with a quantitative-oriented review expands the time-frame of the review publication done by Reichenbach et al. in 2018 by also accounting for the four years, 2017–2020. The originality of this research is acknowledged by combining together: (a) a recap of important theoretical aspects of DdLSM; (b) a bibliometric analysis on the topic; (c) a quantitative-oriented review of relevant publications; and (d) a systematic summary of the findings, indicating important aspects and potential developments related to the DdLSM research topic. The results show that DdLSM are used within a wide range of applications with study area extents ranging from a few kilometers to national and even continental scales. In more than 70% of publications, a combination of the predictors, slope angle, aspect and geology are used. Simple classifiers, such as, logistic regression or approaches based on frequency ratio are still popular, despite the upcoming trend of applying machine learning algorithms. When analyzing validation techniques, 38% of the publications were not clear about the validation method used. Within the studies that included validation techniques, the AUROC was the most popular validation metric, being used accounting for 44% of the studies. Finally, it can be concluded that the application of new classification techniques is often cited as a main research scope, even though the most relevant innovation could also lie in tackling data-quality issues and research designs adaptations to fit the input data particularities in order to improve prediction quality.