Data assimilation

Abstract

Predictions of the morphology of coastal areas are used to make decisions on coastal defence and nature conservation. To predict this morphology, simulations made by morphological models are used. To base decisions on this morphology predictions, we want the uncertainties in these predictions to be small. The goal of this research is to get a better insight in the uncertainties in a morphological model. By having a better understanding of these uncertainties, the decisions made using modelled predictions are stronger substantiated. The specific area focused on in this research is the Frisian Inlet, which is located between Ameland and Schiermonnikoog in the Wadden Sea. To achieve the goal of this research, data assimilation is used. Data assimilation combines data with prior knowledge of a model to find the distribution of probabilities of estimates of a true state. In this research, the used data are bed level measurements and the estimates are simulations for the bed level height made by Delft3D. Data assimilation methods make use of a distribution of model outcomes which should cover all possible outcomes. So, to set up data assimilation successfully, we want to use a parameter that induces a significant change in the bathymetry outcomes of Delft3D. Therefore, a sensitivity study is performed which considers the following six parameters: current related roughness, wave related roughness, wave-related suspended load sediment transport factor, wave-related bedload sediment transport factor, the transverse bed slope and the tidal amplitude. For each parameter, ten values are chosen to simulate. Which of the parameters induces the most change in bathymetry is assessed, using the difference between simulation result and the observation and the mean squared error skill score. The transverse bed slope shows the most variation and is further used in the data assimilation method, which is a particle filter. Hundred uniformly distributed values of the transverse bed slope, between 0.5-100, are used to create different Delft3D simulations that give a bed level prediction. This initial distribution is used for three different periods: 1970-1975, 1975-1979 and 1979-1982. These epochs are defined by the availability of bed level measurements. After one iteration of the particle filter, a new distribution of the parameter values is found. This is used in the next iteration in the same epoch. In each epoch, three iterations are made. The methodology as described in this thesis leads to a convergence of the initial distribution. In epoch 1 and 3, the distribution focuses on values of the transverse bed slope in the high range of the initial distribution. The resulting distribution of epoch 2 focuses on values around a transverse bed slope value of 70. This study shows that it is possible to get a better understanding of the distribution of \gls{alfa} values that leads to probable bedlevel predictions by applying a particle filter. The application of this method brings the simulations closer to each other, but the simulations did not get closer to the observations. The contribution to our understanding of data assimilation for morphodynamic models is that it can be used as a calibration tool for a specific parameter.