Satellite image time-series (SITS) are multidimensional signals of high complexity. Their main characteristics are spatio-temporal patterns which describes the scene dynamics. The information contained in SITS was coded using Bayesian methods, resulting in a graph representation [2]. This paper further presents a concept of interactive learning for semantic labeling of spatio-temporal patterns present in SITS. It enables the recognition and the probabilistic retrieval of similar events. Graphs are attached to statistical models for spatio-temporal processes, which at their turn describe physical changes in the observed scene. Therefore, user-specific semantics attached to spatiotemporal events are modeled using combinations of parameters of a distance model between subgraphs. Thus, the learning step is performed by the incremental definition of a spatio-temporal event type via user-provided positive and negative sub-graph examples. From these examples we infer probabilities of the Bayesian network, based on a Dirichlet model, that links user interest to a specific similarity measurement. According to the current state of learning, sub-graph posterior probabilities are estimated. Experiments, performed on a multitemporal SPOT image time-series, demonstrate the presented reasoning concept.