Refine
Has Fulltext
- yes (3) (remove)
Document Type
- Doctoral Thesis (3) (remove)
Language
- English (3)
Is part of the Bibliography
- yes (3)
Keywords
- Schadensmodellierung (3) (remove)
Institute
- Extern (3) (remove)
Natural extreme events are an integral part of nature on planet earth. Usually these events are only considered hazardous to humans, in case they are exposed. In this case, however, natural hazards can have devastating impacts on human societies. Especially hydro-meteorological hazards have a high damage potential in form of e.g. riverine and pluvial floods, winter storms, hurricanes and tornadoes, which can occur all over the globe. Along with an increasingly warm climate also an increase in extreme weather which potentially triggers natural hazards can be expected. Yet, not only changing natural systems, but also changing societal systems contribute to an increasing risk associated with these hazards. These can comprise increasing exposure and possibly also increasing vulnerability to the impacts of natural events. Thus, appropriate risk management is required to adapt all parts of society to existing and upcoming risks at various spatial scales. One essential part of risk management is the risk assessment including the estimation of the economic impacts. However, reliable methods for the estimation of economic impacts due to hydro-meteorological hazards are still missing. Therefore, this thesis deals with the question of how the reliability of hazard damage estimates can be improved, represented and propagated across all spatial scales. This question is investigated using the specific example of economic impacts to companies as a result of riverine floods in Germany.
Flood damage models aim to describe the damage processes during a given flood event. In other words they describe the vulnerability of a specific object to a flood. The models can be based on empirical data sets collected after flood events. In this thesis tree-based models trained with survey data are used for the estimation of direct economic flood impacts on the objects. It is found that these machine learning models, in conjunction with increasing sizes of data sets used to derive the models, outperform state-of-the-art damage models. However, despite the performance improvements induced by using multiple variables and more data points, large prediction errors remain at the object level. The occurrence of the high errors was explained by a further investigation using distributions derived from tree-based models. The investigation showed that direct economic impacts to individual objects cannot be modeled by a normal distribution. Yet, most state-of-the-art approaches assume a normal distribution and take mean values as point estimators. Subsequently, the predictions are unlikely values within the distributions resulting in high errors. At larger spatial scales more objects are considered for the damage estimation. This leads to a better fit of the damage estimates to a normal distribution. Consequently, also the performance of the point estimators get better, although large errors can still occur due to the variance of the normal distribution. It is recommended to use distributions instead of point estimates in order to represent the reliability of damage estimates.
In addition current approaches also mostly ignore the uncertainty associated with the characteristics of the hazard and the exposed objects. For a given flood event e.g. the estimation of the water level at a certain building is prone to uncertainties. Current approaches define exposed objects mostly by the use of land use data sets. These data sets often show inconsistencies, which introduce additional uncertainties. Furthermore, state-of-the-art approaches also imply problems of missing consistency when predicting the damage at different spatial scales. This is due to the use of different types of exposure data sets for model derivation and application. In order to face these issues a novel object-based method was developed in this thesis. The method enables a seamless estimation of hydro-meteorological hazard damage across spatial scales including uncertainty quantification. The application and validation of the method resulted in plausible estimations at all spatial scales without overestimating the uncertainty.
Mainly newly available data sets containing individual buildings make the application of the method possible as they allow for the identification of flood affected objects by overlaying the data sets with water masks. However, the identification of affected objects with two different water masks revealed huge differences in the number of identified objects. Thus, more effort is needed for their identification, since the number of objects affected determines the order of magnitude of the economic flood impacts to a large extent.
In general the method represents the uncertainties associated with the three components of risk namely hazard, exposure and vulnerability, in form of probability distributions. The object-based approach enables a consistent propagation of these uncertainties in space. Aside from the propagation of damage estimates and their uncertainties across spatial scales, a propagation between models estimating direct and indirect economic impacts was demonstrated. This enables the inclusion of uncertainties associated with the direct economic impacts within the estimation of the indirect economic impacts. Consequently, the modeling procedure facilitates the representation of the reliability of estimated total economic impacts. The representation of the estimates' reliability prevents reasoning based on a false certainty, which might be attributed to point estimates. Therefore, the developed approach facilitates a meaningful flood risk management and adaptation planning.
The successful post-event application and the representation of the uncertainties qualifies the method also for the use for future risk assessments. Thus, the developed method enables the representation of the assumptions made for the future risk assessments, which is crucial information for future risk management. This is an important step forward, since the representation of reliability associated with all components of risk is currently lacking in all state-of-the-art methods assessing future risk.
In conclusion, the use of object-based methods giving results in the form of distributions instead of point estimations is recommended. The improvement of the model performance by the means of multi-variable models and additional data points is possible, but small. Uncertainties associated with all components of damage estimation should be included and represented within the results. Furthermore, the findings of the thesis suggest that, at larger scales, the influence of the uncertainty associated with the vulnerability is smaller than those associated with the hazard and exposure. This leads to the conclusion that for an increased reliability of flood damage estimations and risk assessments, the improvement and active inclusion of hazard and exposure, including their uncertainties, is needed in addition to the improvements of the models describing the vulnerability of the objects.
Today, more than half of the world’s population lives in urban areas. With a high density of population and assets, urban areas are not only the economic, cultural and social hubs of every society, they are also highly susceptible to natural disasters. As a consequence of rising sea levels and an expected increase in extreme weather events caused by a changing climate in combination with growing cities, flooding is an increasing threat to many urban agglomerations around the globe.
To mitigate the destructive consequences of flooding, appropriate risk management and adaptation strategies are required. So far, flood risk management in urban areas is almost exclusively focused on managing river and coastal flooding. Often overlooked is the risk from small-scale rainfall-triggered flooding, where the rainfall intensity of rainstorms exceeds the capacity of urban drainage systems, leading to immediate flooding. Referred to as pluvial flooding, this flood type exclusive to urban areas has caused severe losses in cities around the world. Without further intervention, losses from pluvial flooding are expected to increase in many urban areas due to an increase of impervious surfaces compounded with an aging drainage infrastructure and a projected increase in heavy precipitation events. While this requires the integration of pluvial flood risk into risk management plans, so far little is known about the adverse consequences of pluvial flooding due to a lack of both detailed data sets and studies on pluvial flood impacts. As a consequence, methods for reliably estimating pluvial flood losses, needed for pluvial flood risk assessment, are still missing.
Therefore, this thesis investigates how pluvial flood losses to private households can be reliably estimated, based on an improved understanding of the drivers of pluvial flood loss. For this purpose, detailed data from pluvial flood-affected households was collected through structured telephone- and web-surveys following pluvial flood events in Germany and the Netherlands.
Pluvial flood losses to households are the result of complex interactions between impact characteristics such as the water depth and a household’s resistance as determined by its risk awareness, preparedness, emergency response, building properties and other influencing factors. Both exploratory analysis and machine-learning approaches were used to analyze differences in resistance and impacts between households and their effects on the resulting losses. The comparison of case studies showed that the awareness around pluvial flooding among private households is quite low. Low awareness not only challenges the effective dissemination of early warnings, but was also found to influence the implementation of private precautionary measures. The latter were predominately implemented by households with previous experience of pluvial flooding. Even cases where previous flood events affected a different part of the same city did not lead to an increase in preparedness of the surveyed households, highlighting the need to account for small-scale variability in both impact and resistance parameters when assessing pluvial flood risk.
While it was concluded that the combination of low awareness, ineffective early warning and the fact that only a minority of buildings were adapted to pluvial flooding impaired the coping capacities of private households, the often low water levels still enabled households to mitigate or even prevent losses through a timely and effective emergency response.
These findings were confirmed by the detection of loss-influencing variables, showing that cases in which households were able to prevent any loss to the building structure are predominately explained by resistance variables such as the household’s risk awareness, while the degree of loss is mainly explained by impact variables.
Based on the important loss-influencing variables detected, different flood loss models were developed. Similar to flood loss models for river floods, the empirical data from the preceding data collection was used to train flood loss models describing the relationship between impact and resistance parameters and the resulting loss to building structures. Different approaches were adapted from river flood loss models using both models with the water depth as only predictor for building structure loss and models incorporating additional variables from the preceding variable detection routine.
The high predictive errors of all compared models showed that point predictions are not suitable for estimating losses on the building level, as they severely impair the reliability of the estimates. For that reason, a new probabilistic framework based on Bayesian inference was introduced that is able to provide predictive distributions instead of single loss estimates. These distributions not only give a range of probable losses, they also provide information on how likely a specific loss value is, representing the uncertainty in the loss estimate.
Using probabilistic loss models, it was found that the certainty and reliability of a loss estimate on the building level is not only determined by the use of additional predictors as shown in previous studies, but also by the choice of response distribution defining the shape of the predictive distribution. Here, a mix between a beta and a Bernoulli distribution to account for households that are able to prevent losses to their building’s structure was found to provide significantly more certain and reliable estimates than previous approaches using Gaussian or non-parametric response distributions.
The successful model transfer and post-event application to estimate building structure loss in Houston, TX, caused by pluvial flooding during Hurricane Harvey confirmed previous findings, and demonstrated the potential of the newly developed multi-variable beta model for future risk assessments. The highly detailed input data set constructed from openly available data sources containing over 304,000 affected buildings in Harris County further showed the potential of data-driven, building-level loss models for pluvial flood risk assessment.
In conclusion, pluvial flood losses to private households are the result of complex interactions between impact and resistance variables, which should be represented in loss models. The local occurrence of pluvial floods requires loss estimates on high spatial resolutions, i.e. on the building level, where losses are variable and uncertainties are high.
Therefore, probabilistic loss estimates describing the uncertainty of the estimate should be used instead of point predictions. While the performance of probabilistic models on the building level are mainly driven by the choice of response distribution, multi-variable models are recommended for two reasons:
First, additional resistance variables improve the detection of cases in which households were able to prevent structural losses.
Second, the added variability of additional predictors provides a better representation of the uncertainties when loss estimates from multiple buildings are aggregated.
This leads to the conclusion that data-driven probabilistic loss models on the building level allow for a reliable loss estimation at an unprecedented level of detail, with a consistent quantification of uncertainties on all aggregation levels. This makes the presented approach suitable for a wide range of applications, from decision support in spatial planning to impact- based early warning systems.
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.