Even though quite different in occurrence and consequences, from a modeling perspective many natural hazards share similar properties and challenges. Their complex nature as well as lacking knowledge about their driving forces and potential effects make their analysis demanding: uncertainty about the modeling framework, inaccurate or incomplete event observations and the intrinsic randomness of the natural phenomenon add up to different interacting layers of uncertainty, which require a careful handling. Nevertheless deterministic approaches are still widely used in natural hazard assessments, holding the risk of underestimating the hazard with disastrous effects. The all-round probabilistic framework of Bayesian networks constitutes an attractive alternative. In contrast to deterministic proceedings, it treats response variables as well as explanatory variables as random variables making no difference between input and output variables. Using a graphical representation Bayesian networks encode the dependency relations between the variables in a directed acyclic graph: variables are represented as nodes and (in-)dependencies between variables as (missing) edges between the nodes. The joint distribution of all variables can thus be described by decomposing it, according to the depicted independences, into a product of local conditional probability distributions, which are defined by the parameters of the Bayesian network. In the framework of this thesis the Bayesian network approach is applied to different natural hazard domains (i.e. seismic hazard, flood damage and landslide assessments). Learning the network structure and parameters from data, Bayesian networks reveal relevant dependency relations between the included variables and help to gain knowledge about the underlying processes. The problem of Bayesian network learning is cast in a Bayesian framework, considering the network structure and parameters as random variables itself and searching for the most likely combination of both, which corresponds to the maximum a posteriori (MAP score) of their joint distribution given the observed data. Although well studied in theory the learning of Bayesian networks based on real-world data is usually not straight forward and requires an adoption of existing algorithms. Typically arising problems are the handling of continuous variables, incomplete observations and the interaction of both. Working with continuous distributions requires assumptions about the allowed families of distributions. To "let the data speak" and avoid wrong assumptions, continuous variables are instead discretized here, thus allowing for a completely data-driven and distribution-free learning. An extension of the MAP score, considering the discretization as random variable as well, is developed for an automatic multivariate discretization, that takes interactions between the variables into account. The discretization process is nested into the network learning and requires several iterations. Having to face incomplete observations on top, this may pose a computational burden. Iterative proceedings for missing value estimation become quickly infeasible. A more efficient albeit approximate method is used instead, estimating the missing values based only on the observations of variables directly interacting with the missing variable. Moreover natural hazard assessments often have a primary interest in a certain target variable. The discretization learned for this variable does not always have the required resolution for a good prediction performance. Finer resolutions for (conditional) continuous distributions are achieved with continuous approximations subsequent to the Bayesian network learning, using kernel density estimations or mixtures of truncated exponential functions. All our proceedings are completely data-driven. We thus avoid assumptions that require expert knowledge and instead provide domain independent solutions, that are applicable not only in other natural hazard assessments, but in a variety of domains struggling with uncertainties.

Flash floods are caused by intense rainfall events and represent an insufficiently understood phenomenon in Germany. As a result of higher precipitation intensities, flash floods might occur more frequently in future. In combination with changing land use patterns and urbanisation, damage mitigation, insurance and risk management in flash-flood-prone regions are becoming increasingly important. However, a better understanding of damage caused by flash floods requires ex post collection of relevant but yet sparsely available information for research. At the end of May 2016, very high and concentrated rainfall intensities led to severe flash floods in several southern German municipalities. The small town of Braunsbach stood as a prime example of the devastating potential of such events. Eight to ten days after the flash flood event, damage assessment and data collection were conducted in Braunsbach by investigating all affected buildings and their surroundings. To record and store the data on site, the open-source software bundle KoBoCollect was used as an efficient and easy way to gather information. Since the damage driving factors of flash floods are expected to differ from those of riverine flooding, a post-hoc data analysis was performed, aiming to identify the influence of flood processes and building attributes on damage grades, which reflect the extent of structural damage. Data analyses include the application of random forest, a random general linear model and multinomial logistic regression as well as the construction of a local impact map to reveal influences on the damage grades. Further, a Spearman's Rho correlation matrix was calculated. The results reveal that the damage driving factors of flash floods differ from those of riverine floods to a certain extent. The exposition of a building in flow direction shows an especially strong correlation with the damage grade and has a high predictive power within the constructed damage models. Additionally, the results suggest that building materials as well as various building aspects, such as the existence of a shop window and the surroundings, might have an effect on the resulting damage. To verify and confirm the outcomes as well as to support future mitigation strategies, risk management and planning, more comprehensive and systematic data collection is necessary.

Im Graduiertenkolleg NatRiskChange der Universität Potsdam und anderen Forschungseinrichtungen werden beobachtete sowie zukünftig mögliche Veränderungen von Naturgefahren untersucht. Teil des strukturierten Doktorandenprogramms sind sogenannte Task-Force-Einsätze, bei denen die Promovierende zeitlich begrenzt ein aktuelles Ereignis auswerten. Im Zuge dieser Aktivität wurde die Sturzflut vom 29.05.2016 in Braunsbach (Baden-Württemberg) untersucht. In diesem Bericht werden erste Auswertungen zur Einordnung der Niederschläge, zu den hydrologischen und geomorphologischen Prozessen im Einzugsgebiet des Orlacher Bachs sowie zu den verursachten Schäden beleuchtet. Die Region war Zentrum extremer Regenfälle in der Größenordnung von 100 mm innerhalb von 2 Stunden. Das 6 km² kleine Einzugsgebiet hat eine sehr schnelle Reaktionszeit, zumal bei vorgesättigtem Boden. Im steilen Bachtal haben mehrere kleinere und größere Hangrutschungen über 8000 m³ Geröll, Schutt und Schwemmholz in das Gewässer eingetragen und möglicherweise kurzzeitige Aufstauungen und Durchbrüche verursacht. Neben den großen Wassermengen mit einer Abflussspitze in einer Größenordnung von 100 m³/s hat gerade die Geschiebefracht zu großen Schäden an den Gebäuden entlang des Bachlaufs in Braunsbach geführt.

Modern natural hazards research requires dealing with several uncertainties that arise from limited process knowledge, measurement errors, censored and incomplete observations, and the intrinsic randomness of the governing processes. Nevertheless, deterministic analyses are still widely used in quantitative hazard assessments despite the pitfall of misestimating the hazard and any ensuing risks. In this paper we show that Bayesian networks offer a flexible framework for capturing and expressing a broad range of uncertainties encountered in natural hazard assessments. Although Bayesian networks are well studied in theory, their application to real-world data is far from straightforward, and requires specific tailoring and adaptation of existing algorithms. We offer suggestions as how to tackle frequently arising problems in this context and mainly concentrate on the handling of continuous variables, incomplete data sets, and the interaction of both. By way of three case studies from earthquake, flood, and landslide research, we demonstrate the method of data-driven Bayesian network learning, and showcase the flexibility, applicability, and benefits of this approach. Our results offer fresh and partly counterintuitive insights into well-studied multivariate problems of earthquake-induced ground motion prediction, accurate flood damage quantification, and spatially explicit landslide prediction at the regional scale. In particular, we highlight how Bayesian networks help to express information flow and independence assumptions between candidate predictors. Such knowledge is pivotal in providing scientists and decision makers with well-informed strategies for selecting adequate predictor variables for quantitative natural hazard assessments.

Flash floods and debris flows are iconic hazards inmountainous regions with steep relief, high rainfall intensities, rapid snowmelt events, and abundant sediments. The cuesta landscapes of southern Germany hardly come to mind when dealing with such hazards. A series of heavy rainstorms dumping up to 140mm in 2 h caused destructive flash floods and debris flows in May 2016. The most severe damage occurred in the Braunsbach municipality, which was partly buried by 42,000 m(3) of boulders, gravel, mud, and anthropogenic debris from the small catchment of Orlacher Bach (similar to 6 km(2)). We analysed this event by combining rainfall patterns, geological conditions, and geomorphic impacts to estimate an average sediment yield of 14,000 t/km(2) that mostly (similar to 95%) came from some 50 riparian landslides and channel-bed incision of similar to 2 m. This specific sediment yield ranks among the top 20% globally, while the intensity-duration curve of the rainstormis similarly in the upper percentile range of storms that had triggered landslides. Compared to similar-sized catchments in the greater region hit by the rainstorms, we find that the Orlacher Bach is above the 95th percentile in terms of steepness, storm-rainfall intensity, and topographic curvatures. The flash flood transported a sediment volume equal to as much as 20-40% of the Pleistocene sediment volume stored in the Orlacher Bach fan, andmay have had several predecessors in the Holocene. River control structures from 1903 and records of a debris flow in the 1920s in a nearby catchment indicate that the local inhabitants may have been aware of the debris-flow hazards earlier. Such recurring and destructive events elude flood-hazard appraisals in humid landscapes of gentle relief, and broaden mechanistic views of how landslides and debris flows contribute to shaping small and deeply cut tributaries in the southern Germany cuesta landscape.