Refine
Has Fulltext
- yes (2)
Document Type
- Doctoral Thesis (2)
Language
- English (2)
Is part of the Bibliography
- yes (2) (remove)
Keywords
- flood risk (2) (remove)
Institute
- Extern (2) (remove)
Technological progress allows for producing ever more complex predictive models on the basis of increasingly big datasets. For risk management of natural hazards, a multitude of models is needed as basis for decision-making, e.g. in the evaluation of observational data, for the prediction of hazard scenarios, or for statistical estimates of expected damage. The question arises, how modern modelling approaches like machine learning or data-mining can be meaningfully deployed in this thematic field. In addition, with respect to data availability and accessibility, the trend is towards open data. Topic of this thesis is therefore to investigate the possibilities and limitations of machine learning and open geospatial data in the field of flood risk modelling in the broad sense. As this overarching topic is broad in scope, individual relevant aspects are identified and inspected in detail.
A prominent data source in the flood context is satellite-based mapping of inundated areas, for example made openly available by the Copernicus service of the European Union. Great expectations are directed towards these products in scientific literature, both for acute support of relief forces during emergency response action, and for modelling via hydrodynamic models or for damage estimation. Therefore, a focus of this work was set on evaluating these flood masks. From the observation that the quality of these products is insufficient in forested and built-up areas, a procedure for subsequent improvement via machine learning was developed. This procedure is based on a classification algorithm that only requires training data from a particular class to be predicted, in this specific case data of flooded areas, but not of the negative class (dry areas). The application for hurricane Harvey in Houston shows the high potential of this method, which depends on the quality of the initial flood mask.
Next, it is investigated how much the predicted statistical risk from a process-based model chain is dependent on implemented physical process details. Thereby it is demonstrated what a risk study based on established models can deliver. Even for fluvial flooding, such model chains are already quite complex, though, and are hardly available for compound or cascading events comprising torrential rainfall, flash floods, and other processes. In the fourth chapter of this thesis it is therefore tested whether machine learning based on comprehensive damage data can offer a more direct path towards damage modelling, that avoids explicit conception of such a model chain. For that purpose, a state-collected dataset of damaged buildings from the severe El Niño event 2017 in Peru is used. In this context, the possibilities of data-mining for extracting process knowledge are explored as well. It can be shown that various openly available geodata sources contain useful information for flood hazard and damage modelling for complex events, e.g. satellite-based rainfall measurements, topographic and hydrographic information, mapped settlement areas, as well as indicators from spectral data. Further, insights on damaging processes are discovered, which mainly are in line with prior expectations. The maximum intensity of rainfall, for example, acts stronger in cities and steep canyons, while the sum of rain was found more informative in low-lying river catchments and forested areas. Rural areas of Peru exhibited higher vulnerability in the presented study compared to urban areas. However, the general limitations of the methods and the dependence on specific datasets and algorithms also become obvious.
In the overarching discussion, the different methods – process-based modelling, predictive machine learning, and data-mining – are evaluated with respect to the overall research questions. In the case of hazard observation it seems that a focus on novel algorithms makes sense for future research. In the subtopic of hazard modelling, especially for river floods, the improvement of physical models and the integration of process-based and statistical procedures is suggested. For damage modelling the large and representative datasets necessary for the broad application of machine learning are still lacking. Therefore, the improvement of the data basis in the field of damage is currently regarded as more important than the selection of algorithms.
Floods continue to be the leading cause of economic damages and fatalities among natural disasters worldwide. As future climate and exposure changes are projected to intensify these damages, the need for more accurate and scalable flood risk models is rising. Over the past decade, macro-scale flood risk models have evolved from initial proof-of-concepts to indispensable tools for decision-making at global-, nationaland, increasingly, the local-level. This progress has been propelled by the advent of high-performance computing and the availability of global, space-based datasets. However, despite such advancements, these models are rarely validated and consistently fall short of the accuracy achieved by high-resolution local models. While capabilities have improved, significant gaps persist in understanding the behaviours of such macro-scale models, particularly their tendency to overestimate risk. This dissertation aims to address such gaps by examining the scale transfers inherent in the construction and application of coarse macroscale models. To achieve this, four studies are presented that, collectively, address exposure, hazard, and vulnerability components of risk affected by upscaling or downscaling.
The first study focuses on a type of downscaling where coarse flood hazard inundation grids are enhanced to a finer resolution. While such inundation downscaling has been employed in numerous global model chains, ours is the first study to focus specifically on this component, providing an evaluation of the state of the art and a novel algorithm. Findings demonstrate that our novel algorithm is eight times faster than existing methods, offers a slight improvement in accuracy, and generates more physically coherent flood maps in hydraulically challenging regions. When applied to a case study, the algorithm generated a 4m resolution inundation map from 30m hydrodynamic model outputs in 33 s, a 60-fold improvement in runtime with a 25% increase in RMSE compared with direct hydrodynamic modelling. All evaluated downscaling algorithms yielded better accuracy than the coarse hydrodynamic model when compared to observations, demonstrating similar limits of coarse hydrodynamic models reported by others. The substitution of downscaling into flood risk model chains, in place of high-resolution modelling, can drastically improve the lead time of impactbased forecasts and the efficiency of hazard map production. With downscaling, local regions could obtain high resolution local inundation maps by post-processing a global model without the need for expensive modelling or expertise.
The second study focuses on hazard aggregation and its implications for exposure, investigating implicit aggregations commonly used to intersect hazard grids with coarse exposure models. This research introduces a novel spatial classification framework to understand the effects of rescaling flood hazard grids to a coarser resolution. The study derives closed-form analytical solutions for the location and direction of bias from flood grid aggregation, showing that bias will always be present in regions near the edge of inundation. For example, inundation area will be positively biased when water depth grids are aggregated, while volume will be negatively biased when water elevation grids are aggregated. Extending the analysis to effects of hazard aggregation on building exposure, this study shows that exposure in regions at the edge of inundation are an order of magnitude more sensitive to aggregation errors than hazard alone. Among the two aggregation routines considered, averaging water surface elevation grids better preserved flood depths at buildings than averaging of water depth grids. The study provides the first mathematical proof and generalizeable treatment of flood hazard grid aggregation, demonstrating important mechanisms to help flood risk modellers understand and control model behaviour.
The final two studies focus on the aggregation of vulnerability models or flood damage functions, investigating the practice of applying per-asset functions to aggregate exposure models. Both studies extend Jensen’s inequality, a well-known 1906 mathematical proof, to demonstrate how the aggregation of flood damage functions leads to bias. Applying Jensen’s proof in this new context, results show that typically concave flood damage functions will introduce a positive bias (overestimation) when aggregated. This behaviour was further investigated with a simulation experiment including 2 million buildings in Germany, four global flood hazard simulations and three aggregation scenarios. The results show that positive aggregation bias is not distributed evenly in space, meaning some regions identified as “hot spots of risk” in assessments may in fact just be hot spots of aggregation bias. This study provides the first application of Jensen’s inequality to explain the overestimates reported elsewhere and advice for modellers to minimize such artifacts.
In total, this dissertation investigates the complex ways aggregation and disaggregation influence the behaviour of risk models, focusing on the scale-transfers underpinning macro-scale flood risk assessments. Extending a key finding of the flood hazard literature to the broader context of flood risk, this dissertation concludes that all else equal, coarse models overestimate risk. This dissertation goes beyond previous studies by providing mathematical proofs for how and where such bias emerges in aggregation routines, offering a mechanistic explanation for coarse model overestimates. It shows that this bias is spatially heterogeneous, necessitating a deep understanding of how rescaling may bias models to effectively reduce or communicate uncertainties. Further, the dissertation offers specific recommendations to help modellers minimize scale transfers in problematic regions. In conclusion, I argue that such aggregation errors are epistemic, stemming from choices in model structure, and therefore hold greater potential and impetus for study and mitigation. This deeper understanding of uncertainties is essential for improving macro-scale flood risk models and their effectiveness in equitable, holistic, and sustainable flood management.