In the last century, approximately 100,000 people lost their lives during a flood event and over 1.4 billion people were affected. As the population and economic activities grow in flood-prone areas and the frequencies and intensities of flood events increase due to climate change, damage due to flooding is expected to increase. To limit and control the potentially increased risks, in many locations flood defences, such as levees, are built and existing flood defences are reinforced. It is thus important to be able to properly estimate the reliability of these levees and to understand their potential failure mechanisms. In particular, there are significant uncertainties associated with the occurrence of geotechnical failure mechanisms such as slope instability and piping.
The hindcasting of levee failures can provide valuable information about the factors and uncertainties that dominate levee performance and reliability. Systematic forensic engineering approaches to evaluate failed structures and methods of hindcasting have been developed in the field of structural engineering, but these are not well applicable to failed levees. This is mostly due to the scarcity of relevant information prior to, during or after the levee failure, which leaves multiple scenarios and alternative model choices possible to characterize the event.
This thesis proposes and demonstrates methods for systematic analysis of levee failures at the individual section and system level. These methods of hindcasting are expected to contribute to the overall quality, repeatability, transferability, transparency and recognisability of the analysis of levee failures. In this thesis, existing approaches for evaluating structural failures have been adapted to analyse levee failures using both deterministic and probabilistic techniques. This thesis focuses on the levee failure mechanism of slope instability of the inner slope.
Firstly a deterministic method is proposed which is applied to a slope failure. In this method, the uncertainties in possible causes and computational models are modelled by defining possible scenarios explaining the failure based on all the information available. The influence of the identified scenarios and possible alternatives in model choices are analysed through a sensitivity analysis. Results of the computations are confirmed or refuted by observation information of the failure such as the shape of the failure surface. To illustrate the method, it is applied to the levee failure near Breitenhagen (2013), in Germany in Chapter 2 of this dissertation. The levee near Breitenhagen is located at the intersection of the Saale and the Elbe and it failed due to instability of the slope at the polder side of the levee. Unexpected saturation of the levee, steep slope of the levee, and the influence of the tree roots were identified to cause of the levee failure by previous reports. However, in the present study, an old breach was found to be there (the first proxy was a pond likely caused by this old breach next to the levee; the old breach was later confirmed with archive research). This old breach and pond resulted in a scenario with low strength and high water pressures in both levee and the aquifer and was identified to be the most likely scenario explaining the failure. The results indicate that locally low values of shear strength (low values of pre-overburden pressure or cohesion) explain the failure. Other scenarios that were evaluated resulted in a situation that was not likely to fail or, resulted in a slip surface that differs from the observed failure surface.
The deterministic method does not quantify uncertainties explicitly. That makes it difficult to uniquely identify the most likely scenario to explain the failure. Therefore the deterministic method is advanced by making it probabilistic and by including Bayesian techniques in Chapter 3. Thereby a better insight is provided into the relative likelihoods of the various scenarios explaining the failure. Failure observations (water level at failure, the shape of the slip surface, etc.) and a-priori levee information (soil layering, shear strength etc) are systematically taken into account to quantitatively identify the most likely scenario explaining the failure and the most representative model choices to most accurately characterise the failure. The Bayesian techniques are also used for updating the scenario and possible alternatives in model choices using the observations of the actual failure (if present) such as the shape of the slip surface. To illustrate the method, it is also applied to the levee failure near Breitenhagen (2013) in Germany. Similar to the deterministic method, the old breach resulting in a scenario with locally weak soil and aquifer connection is found to be the most likely scenario. Further, the Limit equilibrium using Spencer’s approach and undrained soil response is identified to be the most representative model choices. The shear strength ratio is identified as the 6 most dominant contributor to the failure. Compared to the “deterministic method” introduced in chapter 2, the probabilistic method adds the possibility to quantitatively substantiate the identification of the most likely scenario explaining the failure as well as the most representative model choices.
Both methods of hindcasting have had little application and validation. Therefore both methods have been applied to a large-scale levee failure experiment. The levee of the Leendert de Boerpolder, in the Netherlands, was brought to failure under controlled circumstances. As a result, very detailed information is available. The levee was brought to failure by gradually lowering the water level in an excavated ditch at the polder side of the levee. Since the water level drawdown is known at the time of failure, this information is used to validate the outcome of both methods of hindcasting. The available levee information was used in two steps. In the first instance, only basic information was used in the hindcasting. In the second step, the geometry of the observed slip surface is also included. The probabilistic method using Bayesian techniques required some adjustment, to account for the survival of previous load phases during a stepwise increase of the load. Both methods of hindcasting identified the same water level drawdown at the moment of failure, but different model choices. In addition, the identified water level drawdown is confirmed by the observed water level drawdown at the time of the failure, i.e. 1.6 m.
Finally, this thesis introduces a method to quantify the influence of deviating conditions on the failure rate of a levee by looking at failures on a system level. The annual failure rate of a levee section is assessed based on information from historical floods. The return period of past events is also taken into account. The presence of deviating conditions at failed and survived levee sections is analysed based on satellite observations. Bayesian techniques and likelihood ratios are used to update the average failure rate as a function of the presence of a deviation. The river system of Sachsen-Anhalt, Germany, is used as a case study. It experienced severe floods with many levee failures in the years 2002 and 2013 resulting in the failure of 41 levee sections due to internal erosion, instability or overflow. It is found that the presence of geological deviations has a significant influence on the observed failure rate and that the failure rate increases with the magnitude of the hydraulic loading. The results show that in the case of the occurrence of a visually identifiable geological deviation in the subsurface, the updated failure rate of a section is about 14 times high than when there is no visually identifiable deviation. The presence of other deviations, such as bushes or trees, or permanent water near the levee also results in a somewhat higher failure rate (20–30% higher) than the calculated average annual failure rate. It is also discussed how the expected number of failures in a system during a high water event with a certain magnitude can be estimated. The results of this research can be used to further optimize soil investigations, calibrate the results of more advanced reliability analyses, and complement risk assessments. The method offers opportunities in particular in environments where little data is available.
Overall, the methods and insights developed in this thesis can contribute to a better understanding of the performance and reliability of flood defence systems.