@phdthesis{Schmidt2024,
  author    = {Schmidt, Lena Katharina},
  title     = {Altered hydrological and sediment dynamics in high-alpine areas - Exploring the potential of machine-learning for estimating past and future changes},
  doi       = {10.25932/publishup-62330},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-623302},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xxi, 129},
  year      = {2024},
  abstract  = {Climate change fundamentally transforms glaciated high-alpine regions, with well-known cryospheric and hydrological implications, such as accelerating glacier retreat, transiently increased runoff, longer snow-free periods and more frequent and intense summer rainstorms. These changes affect the availability and transport of sediments in high alpine areas by altering the interaction and intensity of different erosion processes and catchment properties. Gaining insight into the future alterations in suspended sediment transport by high alpine streams is crucial, given its wide-ranging implications, e.g. for flood damage potential, flood hazard in downstream river reaches, hydropower production, riverine ecology and water quality. However, the current understanding of how climate change will impact suspended sediment dynamics in these high alpine regions is limited. For one, this is due to the scarcity of measurement time series that are long enough to e.g. infer trends. On the other hand, it is difficult - if not impossible - to develop process-based models, due to the complexity and multitude of processes involved in high alpine sediment dynamics. Therefore, knowledge has so far been confined to conceptual models (which do not facilitate deriving concrete timings or magnitudes for individual catchments) or qualitative estimates ('higher export in warmer years') that may not be able to capture decreases in sediment export. Recently, machine-learning approaches have gained in popularity for modeling sediment dynamics, since their black box nature tailors them to the problem at hand, i.e. relatively well-understood input and output data, linked by very complex processes. Therefore, the overarching aim of this thesis is to estimate sediment export from the high alpine {\"O}tztal valley in Tyrol, Austria, over decadal timescales in the past and future - i.e. timescales relevant to anthropogenic climate change. This is achieved by informing, extending, evaluating and applying a quantile regression forest (QRF) approach, i.e. a nonparametric, multivariate machine-learning technique based on random forest. The first study included in this thesis aimed to understand present sediment dynamics, i.e. in the period with available measurements (up to 15 years). To inform the modeling setup for the two subsequent studies, this study identified the most important predictors, areas within the catchments and time periods. To that end, water and sediment yields from three nested gauges in the upper {\"O}tztal, Vent, S{\"o}lden and Tumpen (98 to almost 800 km² catchment area, 930 to 3772 m a.s.l.) were analyzed for their distribution in space, their seasonality and spatial differences therein, and the relative importance of short-term events. The findings suggest that the areas situated above 2500 m a.s.l., containing glacier tongues and recently deglaciated areas, play a pivotal role in sediment generation across all sub-catchments. In contrast, precipitation events were relatively unimportant (on average, 21 \% of annual sediment yield was associated to precipitation events). Thus, the second and third study focused on the Vent catchment and its sub-catchment above gauge Vernagt (11.4 and 98 km², 1891 to 3772 m a.s.l.), due to their higher share of areas above 2500 m. Additionally, they included discharge, precipitation and air temperature (as well as their antecedent conditions) as predictors. The second study aimed to estimate sediment export since the 1960s/70s at gauges Vent and Vernagt. This was facilitated by the availability of long records of the predictors, discharge, precipitation and air temperature, and shorter records (four and 15 years) of turbidity-derived sediment concentrations at the two gauges. The third study aimed to estimate future sediment export until 2100, by applying the QRF models developed in the second study to pre-existing precipitation and temperature projections (EURO-CORDEX) and discharge projections (physically-based hydroclimatological and snow model AMUNDSEN) for the three representative concentration pathways RCP2.6, RCP4.5 and RCP8.5. The combined results of the second and third study show overall increasing sediment export in the past and decreasing export in the future. This suggests that peak sediment is underway or has already passed - unless precipitation changes unfold differently than represented in the projections or changes in the catchment erodibility prevail and override these trends. Despite the overall future decrease, very high sediment export is possible in response to precipitation events. This two-fold development has important implications for managing sediment, flood hazard and riverine ecology. This thesis shows that QRF can be a very useful tool to model sediment export in high-alpine areas. Several validations in the second study showed good performance of QRF and its superiority to traditional sediment rating curves - especially in periods that contained high sediment export events, which points to its ability to deal with threshold effects. A technical limitation of QRF is the inability to extrapolate beyond the range of values represented in the training data. We assessed the number and severity of such out-of-observation-range (OOOR) days in both studies, which showed that there were few OOOR days in the second study and that uncertainties associated with OOOR days were small before 2070 in the third study. As the pre-processed data and model code have been made publically available, future studies can easily test further approaches or apply QRF to further catchments.},
  language  = {en}
}
@phdthesis{Luna2023,
  author    = {Luna, Lisa Victoria},
  title     = {Rainfall-triggered landslides: conditions, prediction, and warning},
  doi       = {10.25932/publishup-60092},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-600927},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xix, 119},
  year      = {2023},
  abstract  = {Rainfall-triggered landslides are a globally occurring hazard that cause several thousand fatalities per year on average and lead to economic damages by destroying buildings and infrastructure and blocking transportation networks. For people living and governing in susceptible areas, knowing not only where, but also when landslides are most probable is key to inform strategies to reduce risk, requiring reliable assessments of weather-related landslide hazard and adequate warning. Taking proper action during high hazard periods, such as moving to higher levels of houses, closing roads and rail networks, and evacuating neighborhoods, can save lives. Nevertheless, many regions of the world with high landslide risk currently lack dedicated, operational landslide early warning systems. The mounting availability of temporal landslide inventory data in some regions has increasingly enabled data-driven approaches to estimate landslide hazard on the basis of rainfall conditions. In other areas, however, such data remains scarce, calling for appropriate statistical methods to estimate hazard with limited data. The overarching motivation for this dissertation is to further our ability to predict rainfall-triggered landslides in time in order to expand and improve warning. To this end, I applied Bayesian inference to probabilistically quantify and predict landslide activity as a function of rainfall conditions at spatial scales ranging from a small coastal town, to metropolitan areas worldwide, to a multi-state region, and temporal scales from hourly to seasonal. This thesis is composed of three studies. In the first study, I contributed to developing and validating statistical models for an online landslide warning dashboard for the small town of Sitka, Alaska, USA. We used logistic and Poisson regressions to estimate daily landslide probability and counts from an inventory of only five reported landslide events and 18 years of hourly precipitation measurements at the Sitka airport. Drawing on community input, we established two warning thresholds for implementation in the dashboard, which uses observed rainfall and US National Weather Service forecasts to provide real-time estimates of landslide hazard. In the second study, I estimated rainfall intensity-duration thresholds for shallow landsliding for 26 cities worldwide and a global threshold for urban landslides. I found that landslides in urban areas occurred at rainfall intensities that were lower than previously reported global thresholds, and that 31\% of urban landslides were triggered during moderate rainfall events. However, landslides in cities with widely varying climates and topographies were triggered above similar critical rainfall intensities: thresholds for 77\% of cities were indistinguishable from the global threshold, suggesting that urbanization may harmonize thresholds between cities, overprinting natural variability. I provide a baseline threshold that could be considered for warning in cities with limited landslide inventory data. In the third study, I investigated seasonal landslide response to annual precipitation patterns in the Pacific Northwest region, USA by using Bayesian multi-level models to combine data from five heterogeneous landslide inventories that cover different areas and time periods. I quantitatively confirmed a distinctly seasonal pattern of landsliding and found that peak landslide activity lags the annual precipitation peak. In February, at the height of the landslide season, landslide intensity for a given amount of monthly rainfall is up to ten times higher than at the season onset in November, underlining the importance of antecedent seasonal hillslope conditions. Together, these studies contributed actionable, objective information for landslide early warning and examples for the application of Bayesian methods to probabilistically quantify landslide hazard from inventory and rainfall data.},
  language  = {en}
}
@phdthesis{Golly2017,
  author    = {Golly, Antonius},
  title     = {Formation and evolution of channel steps and their role for sediment dynamics in a steep mountain stream},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-411728},
  school      = {Universit{\"a}t Potsdam},
  pages     = {180},
  year      = {2017},
  abstract  = {Steep mountain channels are an important component of the fluvial system. On geological timescales, they shape mountain belts and counteract tectonic uplift by erosion. Their channels are strongly coupled to hillslopes and they are often the main source of sediment transported downstream to low-gradient rivers and to alluvial fans, where commonly settlements in mountainous areas are located. Hence, mountain streams are the cause for one of the main natural hazards in these regions. Due to climate change and a pronounced populating of mountainous regions the attention given to this threat is even growing. Although quantitative studies on sediment transport have significantly advanced our knowledge on measuring and calibration techniques we still lack studies of the processes within mountain catchments. Studies examining the mechanisms of energy and mass exchange on small temporal and spatial scales in steep streams remain sparse in comparison to low-gradient alluvial channels. In the beginning of this doctoral project, a vast amount of experience and knowledge of a steep stream in the Swiss Prealps had to be consolidated in order to shape the principal aim of this research effort. It became obvious, that observations from within the catchment are underrepresented in comparison to experiments performed at the catchment's outlet measuring fluxes and the effects of the transported material. To counteract this imbalance, an examination of mass fluxes within the catchment on the process scale was intended. Hence, this thesis is heavily based on direct field observations, which are generally rare in these environments in quantity and quality. The first objective was to investigate the coupling of the channel with surrounding hillslopes, the major sources of sediment. This research, which involved the monitoring of the channel and adjacent hillslopes, revealed that alluvial channel steps play a key role in coupling of channel and hillslopes. The observations showed that hillslope stability is strongly associated with the step presence and an understanding of step morphology and stability is therefore crucial in understanding sediment mobilization. This finding refined the way we think about the sediment dynamics in steep channels and motivated continued research of the step dynamics. However, soon it became obvious that the technological basis for developing field tests and analyzing the high resolution geometry measured in the field was not available. Moreover, for many geometrical quantities in mountain channels definitions and a clear scientific standard was not available. For example, these streams are characterized by a high spatial variability of the channel banks, preventing straightforward calculations of the channel width without a defined reference. Thus, the second and inevitable part of this thesis became the development and evaluation of scientific tools in order to investigate the geometrical content of the study reach thoroughly. The developed framework allowed the derivation of various metrics of step and channel geometry which facilitated research on the a large data set of observations of channel steps. In the third part, innovative, physically-based metrics have been developed and compared to current knowledge on step formation, suggested in the literature. With this analyses it could be demonstrated that the formation of channel steps follow a wide range of hydraulic controls. Due to the wide range of tested parameters channel steps observed in a natural stream were attributed to different mechanisms of step formation, including those based on jamming and those based on key-stones. This study extended our knowledge on step formation in a steep stream and harmonized different, often time seen as competing, processes of step formation. This study was based on observations collected at one point in time. In the fourth part of this project, the findings of the snap-shot observations were extended in the temporal dimension and the derived concepts have been utilized to investigate reach-scale step patterns in response to large, exceptional flood events. The preliminary results of this work based on the long-term analyses of 7 years of long profile surveys showed that the previously observed channel-hillslope mechanism is the responsible for the short-term response of step formation. The findings of the long-term analyses of step patterns drew a bow to the initial observations of a channel-hillslope system which allowed to join the dots in the dynamics of steep stream. Thus, in this thesis a broad approach has been chosen to gain insights into the complex system of steep mountain rivers. The effort includes in situ field observations (article I), the development of quantitative scientific tools (article II), the reach-scale analyses of step-pool morphology (article III) and its temporal evolution (article IV). With this work our view on the processes within the catchment has been advanced towards a better mechanistic understanding of these fluvial system relevant to improve applied scientific work.},
  language  = {en}
}
@phdthesis{Korzeniowska2017,
  author    = {Korzeniowska, Karolina},
  title     = {Object-based image analysis for detecting landforms diagnostic of natural hazards},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-402240},
  school      = {Universit{\"a}t Potsdam},
  pages     = {XV, 139},
  year      = {2017},
  abstract  = {Natural and potentially hazardous events occur on the Earth's surface every day. The most destructive of these processes must be monitored, because they may cause loss of lives, infrastructure, and natural resources, or have a negative effect on the environment. A variety of remote sensing technologies allow the recoding of data to detect these processes in the first place, partly based on the diagnostic landforms that they form. To perform this effectively, automatic methods are desirable. Universal detection of natural hazards is challenging due to their differences in spatial impacts, timing and longevity of consequences, and the spatial resolution of remote-sensing data. Previous studies have reported that topographic metrics such as roughness, which can be captured from digital elevation data, can reveal landforms diagnostic of natural hazards, such as gullies, dunes, lava fields, landslides and snow avalanches, as these landforms tend to be more heterogeneous than the surrounding landscape. A single roughness metric is often limited in such detections; however, a more complex approach that exploits the spatial relation and the location of objects, such as object-based image analysis (OBIA), is desirable. In this thesis, I propose a topographic roughness measure derived from an airborne laser scanning (ALS) digital terrain model (DTM) and discuss its performance in detecting landforms principally diagnostic of natural hazards. I further develop OBIA-based algorithms for the detection of snow avalanches using near-infrared (NIR) aerial images, and the size (changes) of mountain lakes using LANDSAT satellite images. I quantitatively test and document how the level of difficulty in detecting these very challenging landforms depends on the input data resolution, the derivatives that could be evaluated from images and DTMs, the size, shape and complexity of landforms, and the capabilities of obtaining the information in the data. I demonstrate that surface roughness is a promising metric for detecting different landforms in diverse environments, and that OBIA assists significantly in detecting parts of lakes and snow avalanches that may not be correctly assigned by applying only the thresholding of spectral properties of data and their derivatives. The curvature-based surface roughness parameter allows the detection of gullies, dunes, lava fields and landslides with a user's accuracy of 0.63, 0.21, 0.53, and 0.45, respectively. The OBIA algorithms for detecting lakes and snow avalanches obtained user's accuracy of 0.98, and 0.78, respectively. Most of the analysed landforms constituted only a small part of the entire dataset, and therefore the user's accuracy is the most appropriate performance measure that should be given in a such classification, because it tells how many automatically-extracted pixels in fact represent the object that one wants to classify, and its calculation does not take the second (background) class into account. One advantage of the proposed roughness parameter is that it allows the extraction of the heterogeneity of the surface without the need for data detrending. The OBIA approach is novel in that it allows the classification of lakes regardless of the physical state of their water, and also allows the separation of frozen lakes from glaciers that have very similar water indices used in purely optical remote sensing applications. The algorithm proposed for snow avalanches allows the detection of release zones, tracks, and deposition zones by verifying the snow heterogeneity based on a roughness metric evaluated from a water index, and by analysing the local relation of segments with their neighbouring objects. This algorithm contains few steps, which allows for the simultaneous classification of avalanches that occur on diverse mountain slopes and differ in size and shape. This thesis contributes to natural hazard research as it provides automatic solutions to tracking six different landforms that are diagnostic of natural hazards over large regions. This is a step toward delineating areas susceptible to the processes producing these landforms and the improvement of hazard maps.},
  language  = {en}
}
@phdthesis{Vogel2013,
  author    = {Vogel, Kristin},
  title     = {Applications of Bayesian networks in natural hazard assessments},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-69777},
  school      = {Universit{\"a}t Potsdam},
  year      = {2013},
  abstract  = {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.},
  language  = {en}
}