Fluvial flooding poses a major threat to mankind and annually leads to major economic losses with many casualties worldwide. The consequences of this can be mitigated when accurate and rapid predictions are available when the water will arrive at which location. Current numerical simulations take a significant amount of time due to their computational demand, which is not affordable during a calamity where each second can make the difference between life and death. Literature has shown promising results in fluvial flood forecasting by using a deep learning model, which generally takes only a fraction of the computation time compared to conventional models. However, these models have thus far only been trained on static landscapes with changing boundary conditions, which prevents the model’s application elsewhere.

This research explores therefore the possibilities of creating and training a generic non-location bound deep learning model which can predict the spatial distribution of fluvial flood arrival times per grid cell. The architecture of the created deep learning model consists of five parallel encoder-decoders, which takes the elevation, slopes, elevated elements, land roughness, and initial water levels into consideration, depending on the dataset. The model is trained, validated and tested on four unique datasets, which consists of 30,000 flooding samples. The degree of complexity within a sample increases with each dataset number.

The average error for the four consecutive test datasets were 0.91, 1.41, 1.25, and 1.76 hours per cell. The differences in the predicted and the groundtruth are relatively small although the deviation tends to become larger at the end of the simulation, in regions with a strong gradient in arrival time, and in hilly and complex landscapes.

In addition, the model has been tested on various benchmark landscapes to examine specific flow phenomena, as well as a realistic test scenario in the Dutch dike ring 48. The model shows satisfactory performances for landscapes other than those present in the dataset, untill it encounters a feature on which the model was not trained for, such as undershots or irregularly shaped waterways.

This research has shown the potential of deep learning in predicting fluvial flood arrival times on unseen before landscapes. Recommendations for further studies include the use of an active dike breach and a variable flood location.

Digital Soil Mapping (DSM) of soil types in geotechnical project areas is a top priority. These maps are often used in decision making and can have significant consequences related to costs and risks. Usually, these maps are generated by digital soil models that interpolate soil types at known locations. In practice, conventional spatial interpolation techniques are still often used for DSM of soil types, such as inverse distance weighting and kriging. However, conventional models are not well suited for predicting or interpolating soil types because of their inability to deal with categorical data properly. Besides, the design of the conventional models does not allow for incorporating the abundance of meaningful covariate information that is available nowadays. The flexibility of machine learning algorithms vanquish both problems and has become increasingly popular for DSM of soil properties in recent years. The results of machine learning techniques for DSM of soil properties are promising and generally outperform conventional models. However, few studies have used machine learning for DSM of soil types and is therefore still a relatively unknown field. Moreover, at the time of writing, there are no studies that use sequence models for DSM of soil properties or types. Hence, the author proposes to introduce a new method for DSM of soil types, namely a Long Short-Term Memory (LSTM) network. The intuition behind this introduction is that the spatial correlation can be captured in sequences and can improve soil type prediction.

Real project data from a cable burial project is used to evaluate and compare the performance of the conventional interpolation methods triangulation and kriging, the machine learning models random forest and XGBoost, and the newly proposed deep learning model LSTM. The project data consist of 757 vibro cores (VC), 718 cone penetration test (CPT), bathymetry data and sub-bottom profilers. The geotechnical data, i.e. VCs and CPTs, is received on separate PDF pages that require to be digitized first. This thesis describes a simple yet precise manner to extract this data from the PDFs. The VCs and CPTs are provided with a soil type interpretation and can be used directly for developing the models. The data is split into a training set to develop/train the models and a test set for evaluation. Ultimately, the best performing model is used to build a 3D stratigraphic soil model for the project area with associated prediction accuracies.

All state-of-the-art techniques outperform the conventional models and especially in predicting minority classes. The best performing model is random forest with an overall accuracy of 85.44\% and is comparable to the performance of XGBoost of 85.11\%. LSTM network achieved a slightly lower accuracy of 84.27\%. The results show that LSTM is suitable for DSM of soil types and has considerable potential for improvement as only a few possibilities of the model have been examined.

In recent years many automatic differentiable programming frameworks have been developed in which numerical programs can be differentiated through automatic differentiation (AD). Examples of these frameworks are Theano, TensorFlow and Pytorch. These frameworks are widely used in Machine Learning. AD also finds applications in the field of computational fluid dynamics (CFD). It is used to develop discrete adjoint CFD code for research concerning for instance sensitivity analysis, data assimilation and design optimization. However, the use of the automatic differentiable programming frameworks in the field of CFD is limited. One can find some examples in the literature on how to find a numerical solution to an initial value problem using a differentiable programming framework. In this work it will be clarified how one can implement an semiimplicit
time integration scheme for a staggered grid to simulate the propagation of long waves in water with a free surface in TensorFlow. A main advantage of the automatic differentiable programming frameworks is the user friendly application
programming interface (API) for AD. No research has been conducted to use this API in the field of CFD. In this work an example will be given how one can use TensorFlow for research concerning sensitivity analysis. AD requires a significant allocation of memory on a CPU/GPU when working with fine meshes and/or long simulations and since CPU/GPU memory is finite, the method checkpointing is
proposed to make it feasible to perform sensitivity analysis when working with fine meshes and/or long simulations. Another main advantage of the differentiable programming framework TensorFlow is the use of compute unified device architecture (CUDA) of a NVIDIA GPU in order to perform computations in parallel, which results in a significant reduction in computation time. A Benchmark will be given that indicates the computational efficiency of TensorFlow compared to a loop over grid implementation in NumPy and a Fortran CPU scalar implementation.