Current maritime traffic monitoring systems are essentials for a maritime situational awareness. However, they are not always adapted to the identification of risky behaviours of ships. It is very difficult for operators responsible for monitoring traffic to identify which vessels are at risk among all the shipping traffic displayed on their screen. We present in this paper a hybrid approach for analysing dangerous behaviours of ships based on AIS data. This approach is based on supervised and unsupervised analyses and it was developed in the frame of PhDs. Our approach is based on three complementary methodologies: (1) data mining for knowledge and pattern extraction for behaviour modelling, (2) ontological modelling of behaviour for an unsupervised detection and (3) geovisual analysis of large volume of data for a supervised detection of threats at sea. Three prototypes were developed to test our approach.