In regions where levees make up a critical part of flood risk management, it is imperative to understand how reliable the levees are. Reliability estimates are typically founded on complex probabilistic computations, the details of which are usually beyond the expertise of risk managers. This can be problematic. When reliability estimates and intuition conflict, risk managers tend to become skeptical of the methods used, and they reach a roadblock in decision making. Furthermore, probabilistic methods generally rely on mathematical descriptions of levee failures. With multifunctional levees - even such simple ones as a bench or tree existing on the levee, the failure mechanisms are not yet perfectly understood, and thus cannot be brought into a traditional probabilistic analysis. This is often a reason for not considering multifunctional flood defense alternatives.
The research in this project addresses the skepticism about traditional reliability analysis for levee systems (as used in the Netherlands), and proposes alternative methods for handling the uncertainty surrounding multifunctional flood defenses. We achieve this via a two-pronged approach.
First, we investigate aspects of the reliability model that has been used in the Netherlands for national risk studies, and will be used for future assessment of Dutch levees. We begin with the accuracy of the most ‘magical’ aspect of the model - its ability to iteratively combine failure probabilities of different levee sections and failure mechanisms. We also investigate the goodness of the model’s approach to accounting for the length effect (the phenomenon by which the failure probability of a levee section increases as a levee section gets longer). These two aspects of the model elicit the most skepticism; by investigating them in detail, we hope to alleviate some of this skepticism and to remove some of the mystery surrounding the methods.
Second, we investigate the use of Bayesian networks as an alternative to the methods currently in use in the Netherlands in cases where it is warranted. Bayesian networks are powerful probabilistic modeling tools that have a couple of attractive advantages: (1) you can use ‘evidence’ (observations or strong intuition) to update the distributions of the random variables contributing to levee failure (e.g. thickness of a blanket layer), and subsequently to update the reliability estimate, (2) uncertain failure mechanisms (as is the case with multifunctional flood defenses) can be incorporated via (structured) expert elicitation when no mathematical description is available, and (3) the model is a ‘graphical’ model, which means the relationships between variables is pictorially described, making it less abstract to managers who would like to understand the model, but lack the probabilistic training.
The research in this project addresses the skepticism about traditional reliability analysis for levee systems (as used in the Netherlands), and proposes alternative methods for handling the uncertainty surrounding multifunctional flood defenses. We achieve this via a two-pronged approach.
First, we investigate aspects of the reliability model that has been used in the Netherlands for national risk studies, and will be used for future assessment of Dutch levees. We begin with the accuracy of the most ‘magical’ aspect of the model - its ability to iteratively combine failure probabilities of different levee sections and failure mechanisms. We also investigate the goodness of the model’s approach to accounting for the length effect (the phenomenon by which the failure probability of a levee section increases as a levee section gets longer). These two aspects of the model elicit the most skepticism; by investigating them in detail, we hope to alleviate some of this skepticism and to remove some of the mystery surrounding the methods.
Second, we investigate the use of Bayesian networks as an alternative to the methods currently in use in the Netherlands in cases where it is warranted. Bayesian networks are powerful probabilistic modeling tools that have a couple of attractive advantages: (1) you can use ‘evidence’ (observations or strong intuition) to update the distributions of the random variables contributing to levee failure (e.g. thickness of a blanket layer), and subsequently to update the reliability estimate, (2) uncertain failure mechanisms (as is the case with multifunctional flood defenses) can be incorporated via (structured) expert elicitation when no mathematical description is available, and (3) the model is a ‘graphical’ model, which means the relationships between variables is pictorially described, making it less abstract to managers who would like to understand the model, but lack the probabilistic training.
Michael Nevins, U.S. Army Corps of Engineers. [Aerial view of a broken levee under repair on the Feather River near Nicolaus, CA]
The proposed research is of critical importance to Dutch policy decisions about flood defence strategies. This research is directly relevant for Rijkswaterstaat, who has the task of making policy decisions, as well as the individual water boards, who are required to assess the safety of their defences and carry out improvements based on the results of the probabilistic model. The research is relevant for the Dutch market in international flood risk projects. The proposed research is also relevant for the U.S. Army Corps of Engineers (USACE). They have expressed keen interest in application of the probabilistic methods used in the Netherlands to a pilot area in the U.S. Involvement of the USACE is anticipated and will ensure sustained contact and knowledge transfer.
