Large areas of the world are protected by flood defence systems. A common part of flood defence systems is a levee with grass cover, with the primary function to protect hinterland against floods. Waveovertopping might lead to erosion of the grass cover, followed by erosion of the inner slope and potentially levee breaching. The wave overtopping simulator (WOS) was developed for full-scale and insitu testing of the erosional resistance of levees against wave overtopping. The first experiment with this simulator was carried out in 2007 and more followed. However, the data gathered during these overtopping experiments has always been considered on individual experiment scale or in small subsets, exception being the Cumulative Overload method. The experiments proved to be difficult to reproduce, seemingly identical experiments gave different results. The objective of this thesis is to combine WOS data and multiple grass erosion models to make a prediction for the failure of the grass cover. Key features that determine the prediction of failure are the location and moment of failure. Therefore, these are included in the main research question, which is identified as: How can grass erosion models and WOS experiment data be combined in a new method to generate a prediction of the moment and location of the grass cover failure due to wave overtopping? A literature study on grass erosion and grass erosion provides the basis for modelling grass erosion, a review of combination techniques gives insight in combining results for a prediction. A selection of models is discussed and used for setting up a prediction method. These include three flow-based models which are semi-realistic simplified representations and time-independent over a single overtopping event: Cumulative Overload Method, Analytical Grass Erosion Model and Dean Stream Power. These models are, where required, adjusted to model grass cover failure and to comply with identical hydrodynamic input. A fourth model, the Wave Impact Approach, is an impact-based approach. Each model is calibrated on each WOS experiment, creating a set of calibrated models, which function as the set of predictors. The number of predictors equals the number of models times the number of experiments. During calibration, the moment of failure has been traced back using the control list of the WOS, and vice-versa, resulting in a moment during the experiment at which the grass cover failed. Detailed review of all factual reports on WOS experiments and the nature of the grass erosion models highlighted the need for a locationdependent resistance parameter to determine the location of failure. Therefore, calibration was based on a location-depended resistance parameter; critical flow velocity for the flow-based models and critical pressure for the impact based model. This set of predictors has been used to create a set of predictions for each of the five validation experiments, input being the geometry, loading and initial condition of the slope. The initial condition given by a registration of anomalies to the average grass cover. Failure is determined by vote and averaging. After validation, none of the final predictions proved to be fully correct. Meaning that none of the final predictions correctly indicated the location and moment of the first grass cover failure. Despite this, in four of the five validations at least one correct failure location was indicated when considering the specific anomalies. The main shortcoming is concluded to be the prediction of number of waves until grass cover failure. Based on this research and given a certain average grass cover quality, it is concluded that the resistance against grass erosion by wave overtopping is described by resistance against erosion of anomalies. For at least 22 of 28 sections in the dataset, grass cover failure occurred at an anomaly of the average grass cover. No grass cover failures occurred related to the average grass quality. Especially mole activity showed a significant 46% decrease in the averaged calibrated critical flow velocity with relative to that for the average grass cover. This leads to the conclusion that design and safety assessments should include conditions other than the average grass quality. For future research a method is recommended that divides mole activity into classes, based on certain characteristics. Each class distinguishable by unique combination of properties and assigned a resistance against erosion. For design, the probability of occurrence of each class must be determined and combined with the corresponding influence on the resistance against erosion. For assessment, an inventory or representative sample of animal activity must be available to asses if the occurrences of animal activity are within the design requirements.