Uncertainty quantification and attribution in flood risk modelling

Flood risk models are used to help manage the financial and societal risks associated with floods. They quantify risk (usually in terms of flood losses) by modelling the underlying physical hazard, the exposure and vulnerability of people, properties and/or other assets to the hazard. However, these models are subject to significant uncertainty. Knowing how much uncertainty exists in risk estimates and how it can be attributed to its various sources is essential to guide efforts for model improvement, as well as to help risk managers make better decisions. Currently, there is a lack of knowledge regarding how much uncertainty affects the inputs of flood risk models and how it can be quantified. Additionally, it remains unclear how such input uncertainties propagate and which uncertain input mostly controls the uncertainty in predicted losses. Flood risk models can now estimate risk at large scale which means they produce large and complex datasets that can be challenging to analyse. The application of Machine Learning methods (like Decision Trees) poses a set of specific challenges. In this Thesis I intend to improve our understanding of uncertainty quantification and attribution in flood risk modelling. I performed an uncertainty and sensitivity analysis of a flood risk model over a large and heterogeneous domain, the Rhine River Basin, and in doing so developed a new Machine Learning method (interactive Decision Tree) to analyse large datasets by incorporating scientific knowledge into machine learning. I then developed an approach to quantify the uncertainty in the inputs of a flood risk model, based on a systematic literature review. Using this approach, I reported variability ranges for residential buildings value, damage ratios of the vulnerability curves, and the return period of flood events in the Rhine River Basin. I used these input uncertainty ranges to perform a global sensitivity analysis on the flood risk model over the Rhine River Basin. First, I identified the dominant input uncertainties in each spatial unit of the study domain for two key model outputs: the Average Annual Losses (AAL) and the Loss Exceedance Curves (LEC). I found that AAL is dominated by uncertainty in vulnerability, while the dominant uncertainties for the LEC change with the return period, with vulnerability becoming increasingly important with increasing return period. Finally, I attempted at linking dominant input uncertainties with the hydrological and socio-economic characteristics of the spatial units using an interactive Decision Tree. I found that topography, degree of urbanization and residential buildings' value are key characteristics for determining how dominant uncertainties change over space within the study domain.