Quantifying the Uncertainty In Future Flood Hazard Assessment Assuming Climate Change Uncertainties

Floods have an enormous impact worldwide. In the UK the 2015/16 floods caused more than £1.3 billion worth of damage. The consensus of current research on climate change and its effect on future flooding hazards is that in the UK floods are going to become more frequent and more extreme. The ability to properly safeguard against such natural disasters will reduce the economical and social impacts occurring and protect future generations.

In modelling flood hazards the process takes regional climate models (RCM) as input for hydrological models. These are then analysed using extreme value theorem giving inputs into the hydraulic model. Finally, the hydraulic model output is produces probabilistic maps of flooding in a given area. At each stage there is uncertainty in the form of input (RCM's), structure (hydraulic model), and parameters (friction coefficient). This results in a cascade of uncertainty through the system. The aim of uncertainty quantification is to identify sources of uncertainty and understand how it cascades through the model framework.

When decisions are being made it is important to fully understand the total uncertainty associated with the system. Probabilistic maps are used to create regions in which there is a given probability of flooding. This results in a better understanding of possible future scenarios compared to deterministic methods. Floods are inherently uncertain and as such flood risk management must take uncertainty into account. Investigating and quantifying the uncertainty associated within probabilistic mapping has become an area of interest in flood inundation modelling. However, industrial and academic researchers alike have been validating and verifying hydraulic models but running a limited number of simulations. This does not account for structural or input uncertainty in the hydraulic model. Therefore, there is a need for Monte Carlo type approaches.

Increasing the dimensionality of the hydraulic model reduces the structural uncertainty and creates a more realistic output but takes more time to run. Thus, a comparison of 1D/2D hybrid and fully 2D hydraulic model outputs and structural uncertainties will help with deciding which is best suited for different studies.

Different sampling techniques and computational cost reduction methods can be implemented to speed up run time: Latin hypercube sampling, Multi-level Monte Carlo, Markov chain Monte Carlo, and Multi-level Markov chain Monte Carlo will all be investigated and compared. These approaches reduce the computational cost due to an increase in convergence rates while increasing accuracy compared to the simple Monte Carlo method. Utilising surrogate models will be explored as they can run faster while maintaining the required accuracy by extrapolating model outputs. The surrogate models can also be used for uncertainty propagation through the system and sensitivity analysis.

The application of these improvements to the hydraulic modelling process and the variance reduction techniques will result in a more accurate probabilistic map of case studies in the UK, while being more accessible to practitioners who have limited computational time. The hydraulic model and total uncertainty will be quantified; creating a framework that sufficiently assesses total model uncertainties. With a changing and more unpredictable climate the ability to have a more clear understanding of uncertainties could prevent economical, social, and environmental losses. This research can be used for further studies across the UK and abroad.