Threshold uncertainty for multivariate extreme value models for coastal flood risk analysis
Threshold uncertainty for multivariate extreme value models for coastal flood risk analysis.
Liu, Y.and Ross, E. and Jonathan, P. and Gouldby, B.P.
In: EVAN Conference 2017, 5-7 September 2017, Southampton, UK. (2017)
Full text not available from this repository.
|Abstract:||For the past few decades, extreme value statistical models have been widely used in the field of coastal flood risk analysis. In particular, the class of multivariate extreme value / extremal dependence models, such as Heffernan and Tawn (2004), have become one of the main approaches for simulating multivariate extreme coastal conditions as part of a joint probability study, e.g. simulating jointly the wave height, skew surge, wind speed and direction, etc. during storms. These variables serve as the boundary condition to other sophisticated physical or numerical models to ultimately generate the flood extent and depth. Such analysis enables the users to obtain a probabilistic view of possible flooding events as well as to understand the consequence of rarely observed or unobserved coastal conditions.
The conditional model from Heffernan and Tawn (2004) is reasonably parametrised, so that fitting the model is relatively easy, even for high dimensional problems where other models may suffer. It can also account for both asymptotic independence and dependence, and in an asymmetric way if necessary, which is often observed in the coastal condition data. In a threshold-based setup, the proposed model is theoretically justified in the limiting case where the threshold tends to infinity. However, the fitting of the model to a given dataset is almost always performed at a high but finite threshold, assuming the bias induced is negligible. This, together with the threshold choice for the marginal distribution with a fitted Generalised Pareto (GP) tail, leads to variability in the parameter estimation, which in turn affects the probability estimation.
In practice, such threshold uncertainty is not always explicitly accounted for as well as the other sources of uncertainty. Quite often the choice is introduced as a fixed input to the whole analysis procedure, assuming the impact would be small. Through this study, we will demonstrate with real data the potential influence of the threshold choice over the joint probability estimation. We will also introduce several model fitting methods that account for the threshold uncertainty and compare their effectiveness in quantifying the threshold uncertainty in a typical coastal flood risk analysis.|
|Item Type:||Conference or Workshop Item (Paper)|
|Subjects:||Floods > General|
Coasts > General
|Deposited On:||29 Jan 2018 15:21|
|Last Modified:||29 Jan 2018 15:21|
Repository Staff Only: item control page