Visualization is essential in the risk assessment of coastal or river flooding. In this work, we deal with expensive-to-evaluate hydraulic simulators, taking as random scalar inputs offshore meteo-oceanic conditions and dyke breach parameters, whereas the output is a flooding map. The challenge is to display a few prototype maps representing at best the probability law of the flooding event, which is a typical quantization problem. The K-Means algorithm classically serves to minimize the expected squared distance between samples of the random event and their representatives. This clustering technique is adapted to handle three key specificities of our context. First, the quantization is done in the specific space of flooding maps, which requires to define an appropriate distance measure. Second, because of the time-consuming simulators, a Gaussian process based approach adapted to spatial outputs is proposed. Third, flooding being a low-probability event, traditional Monte Carlo approaches are inefficient and an Importance Sampling scheme is introduced to generate the maps. The prototype maps represent the distribution of the floodings and each is associated to a probability mass. To discriminate between the scenarii leading to similar floodings, a second quantization procedure is carried out in the input space within each cluster. The method is evaluated on an analytical example before being extended to the coastal case of Les Boucholeurs on the French Atlantic coast.