The effects of climate change are felt all around the world. An increased sea level goes hand in hand with an increased risk of flooding. To combat this, the coastlines must be reinforced to withstand future sea levels. However, repeatedly reinforcing coastlines to keep up with the sea level rise (SLR) could prove extremely costly. An alternative approach would be to shorten the coastline, as the Netherlands did with the Afsluitdijk to enclose the IJsselmeer. Looking at Europe, Groeskamp and Kjellsson (2020) proposed the construction of the Northern European Enclosure Dam (NEED) — a dam that would disconnect the North and Baltic Seas from the Atlantic Ocean. In this way, it would protect fifteen northern European countries against the accelerated global mean SLR (GMSLR), as it simultaneously shortens the coastline that requires reinforcement. This thesis aims to determine whether the NEED (Adaptation Strategy 2) would be a financially favourable adaptation strategy over raising the coastal defences on a country-by-country basis (Adaptation Strategy 1) around the North Sea to combat future GMSLR, and if so, at which GMSLR.

Through the use of the GLOFRIS model framework it was possible to estimate that by 2080 a total of 15,000 km2 would be inundated, affecting 9.5 million people and resulting in damages up to 1 trillion € for all fifteen countries combined. Reinforcing the regional flood protections (Adaptation Strategy 1) is estimated to have a total cost range of 245 to 335 billion € for a 1-metre GMSLR with an increase in costs between 170 and 235 billion € per metre GMSLR. The construction of the NEED (Adaptation Strategy 2), using an earth-fill dam design with 1:6 slopes on either sides, is estimated to be just under 1.1 trillion €, with an increase in costs of 11 billion € per metre GMSLR.

It was found that the NEED flood protection adaptation strategy will eventually become more financially favourable over the regional strategy. Estimated to be more cost-effective beyond 5.15 metres GMSLR, which is associates with construction costs of roughly 1.15 trillion €. This GMSLR for scenario SSP5-RCP8.5 is expected to occur between 2280 and 2660, approximately. However, as the total costs are greatly contingent upon the core material, modifying slope angles of the NEED design will lead to a significant reduction in volume and, hence, costs. For the alternative designs with a 1:4 and 1:5 slope, the total costs are reduced by 17% and 34%, respectively. For these designs, the NEED will already become favourable at 3.35 and 4.25 metres GMSLR, respectively.

Several cost distributions have been created based on the four aspects that have been investigated, namely (i) coastline reinforcement length, (ii) size of inundated area, (iii) population exposed and (iv) economic damages caused by flooding. Together with the extrapolated regional costs per country, it is possible to determine which distribution is the most and least financially favourable for each country and whether contributing to the NEED is even favourable at all from the perspective of each country.

This research scrutinised the costs, effects and consequences of the most extreme scenario that generated the greatest exposure indicator values (i.e. SSP5-RCP8.5 combined with a return period of 1000 years). Through this
assessment, it was possible to estimate the costs associated with both adaptation strategies, and furthermore, to determine at which GMSLR one strategy surpasses the other in financial attractiveness and when this GMSLR can be expected. However, it should be noted that recent studies have shown that, in reality, the most extreme scenario might, unfortunately, turn out to be even more extreme than the most extreme scenario assumed in this thesis. And as the consequences strongly dependent on how climate change will unfold in the future, the costs to combat and the timing of such GMSLR occuring will differ.

The results retrieved from this research provide insight into when the NEED flood protection adaptation strategy will become a better alternative to regional flood protection reinforcement. However, it should be borne in mind that it is not a matter of ‘either-or’, but rather ‘both-and’, as regional dike reinforcement cannot entirely be omitted when deciding to construct the NEED. Instead, a balance must be found in the extent to which regional dike reinforcement is required to protect the countries while the NEED is under construction. So there are plenty of uncertainties and questions that require additional research to fully comprehend all the effects of this massive operation and making it feasible.

Flood simulations can give insight into the consequences of flood scenario's and can help to create hazard- and risk maps to support decision-making in flood risk management and in crisis management. 2D hydrodynamic simulations give accurate descriptions of the propagation of a flood and rely on advanced numerical methods to solve a set of physics-based mathematical equations. A drawback of these models is that they can be computationally expensive with run times in the order of hours or days depending on the time and spatial resolutions. In this study we explore the use of deep learning techniques in a surrogate model for 2D flood simulation. We propose and test a deep learning-based surrogate modelling framework that can be used to train a deep learning-based surrogate model. Once trained, the surrogate model can be used as a substitute for the hydrodynamic model with the advantage of being much more efficient in terms of run time and can be of great value in for example crisis situations. For training, a data set of expensive 2D hydrodynamic simulations was created using the SOBEK software program. Such simulations require a lot of input data, such as input parameter maps specifying the terrain over the computational grid and boundary conditions. To make training data-efficient, a sampling strategy was used for the input of the flood simulations. Three deep learning architectures were trained and tested. The first two architectures are feed-forward networks and the third architecture is of recurrent network type. These networks contain convolutional neural network (CNN) architectures with an encoder-decoder structure to make patch-level predictions of the flood characteristics in time. These patches contain a small section of the flood prone area and an encoder network is used to extracts coarse feature maps from this data that is then refined by a decoder network to create a prediction of the flood propagation. Using patches has the advantage of making a surrogate model able to create flood simulations over prone areas without restrictions on size or shape by tiling the output patches with flow predictions. Also it allows the surrogate to focus only on regions where the flood has reached and not on the regions where no water has arrived. It was found that with the recurrent architecture, the surrogate model was most capable of emulating the ground truth flood simulations in the test simulation. This trained network architecture was used in a case study where the surrogate was applied to create flood simulations in a small dike ring in the Netherlands. This shows that the surrogate modelling framework can be used to train a deep learning-based surrogate and, once trained, can be used to create flood simulations similar to hydrodynamic simulations. However, two main challenges were identified in using such data-driven deep learning-based surrogates. Firstly, keeping the predictions of the flood characteristics accurate enough to avoid large error propagation. Secondly, accurately generating large amounts of data from relatively little information present in the boundary condition and terrain.

The existing hydrodynamic models consider full physics approaches to calculate storm surge at coastal regions. However, due to the complexity of the equations that model these processes, the computational time and power required to run them can be large, compared to models that consider simplified equations. By contrast, simplified hydraulic models lack physical background, what leads to a minor accuracy in the surge estimations with respect to hydrodynamic models. In this project, a stochastic model has been developed with the objective to estimate surge at the coast of Mississippi (United States) at a reasonable accuracy and time, without solving equations that represent complex physical processes. The stochastic model needs to be trained by means of hurricane data, including surge levels. This information must be generated beforehand, by simulating a limited number of hurricanes with an hydrodynamic model. In this project, the hydrodynamic model Delft3D Flexible Mesh has been used for this purpose. The approach followed to build the stochastic model has been based on three main steps. The first step has been setting up and validating the hydrodynamic model in Delft3D FM. Hurricane Katrina (2005) has been simulated to calibrate the input parameters of the model, by comparing the maximum simulated water levels at 41 stations along the shoreline of Mississippi to the high water marks observed at the same locations during the event. Tide, surge and wave setup have been considered in the validation. The results of the validation show a line best fit slope from the origin of 0.912 and an R-squared of 0.996. At Gulfport, the absolute error of the surge estimation is 19 centimeters, equivalent to a relative error of 2.5%. The second step in the construction of the stochastic model has been the generation of a historical hurricane data base. The hurricane best tracks have been retrieved from the HURDAT2 data base. The variables considered have been the forward speed and the forward direction of the hurricane at landfall, the wind speed at landfall, the distance from landfall location to a reference point (Galveston Bay) and the maximum storm surge during the hurricane. In this case, only the hurricane forcing is considered as external action. The storm surge has been recorded at Gulfport Harbour (central coast of Mississippi). The values of the surge have been obtained by using the validated model to simulate the historical hurricanes making landfall in a rectangular domain of 600 kilometers, being Gulfport Harbour the center of the rectangle. Due to the scarce number of hurricanes making landfall in this region, the tracks of the hurricanes making landfall in the North of the Gulf of Mexico but outside the rectangular domain have been shifted inside the domain, in order to generate a sufficiently large data base to train the stochastic model. A data base with 140 hurricanes has been built, from which the 85% (119 hurricanes) have been used for the training of the stochastic model and the other 15% (21 hurricanes) have been used for the validation of the stochastic model. The third and last step has been the setup and validation of the stochastic model, by comparing the storm surge obtained from the stochastic model to the surge obtained from the hydrodynamic simulations. The stochastic model used to estimate storm surge has been a Bayesian Network that assumes normal copulas to represent the joint distribution between nodes of the network. The calculated slope of the best fit line for the mean surge values has been 0.861, with an R-squared of 0.885. Moreover, the average standard deviation of the estimations is 1.16 meters. These results indicate a reasonable estimation of the surge by means of the Bayesian Network. This estimation can be made in the order of seconds.

Over the recent years, flood risks and losses have been increasing for coastal cities due to climate change, subsidence, population and economic growth. Hong Kong and Macau are two cities located in the Pearl River Delta that experience a significant flood risk due to storm surges. The increased losses and risks has sparked interest around the world for efficient and accurate flood forecasting. At the moment coastal flooding events are often simulated with difficult hydrodynamic models that reproduce the physical phenoms. Over the last decades there has been more interest in other methods to forecast storm surges, namely neural networks. Other than hydrodynamic models a neural network is capable is making predictions in seconds, while the model can take hours to finish simulation. fast and accurate storm surge forecasting is of importance for disaster and evacuations management strategies and will only become more important in the future. During this research the main goal is to develop a neural network capable of prediction maximum water levels due to storm surges in case of an approaching tropical cyclone. The neural network is trained with data that is obtained from hydrodynamic simulations. A synthetic storm database is used to provide the necessary data to conduct 1000 simulations of which the results are used in the neural network. The first focusses on developing a hydrodynamic model capable of accurately simulation tropical cyclone induced storm surges in Hong Kong and Macau. The model is calibrated in a way to reproduce the real world as close as possible. It accounts for the real life bathymetry, topography, astronomical tides and wind forcing. The storm surge model is validated extensively based on three historical tropical cyclones. The errors between the observed and simulated water level are below 20 cm, after calibrations of different physical parameters within the model. The second step of this research is to use the validated storm surge model to run synthetic storm simulations. Instead of using historical storms who only have been recorded for 40 years, a synthetic storm database is used containing 10000 years worth of data. From that storm data, 1000 synthetic tropical cyclones are selected that come close to Hong Kong and Macau. With the data obtained from the previous two steps, it is finally possible to training the neural network. In this network a total of seven input parameters (tropical cyclone track parameters) are used to estimate the maximum water level that will occur during the tropical cyclone. The input parameters considered are: latitude, longitude of TC eye, maximum wind speed, minimum eye pressure, radius of maximum winds, forward speed, forward propagation direction. During the development of the network, three different types of configurations are tested. Extensive validation and calibration shows that neural networks are capable of making maximum water level predictions for a large number of cases. However, variety in the quality of the prediction is observed. Improvements still can be made for more accurate predictions.