We propose an original data mining method to segment agricultural landscapes based on a temporo spatial modelling of their crop successions. This method consists in (i) choosing the length of the crop succession, (ii) choosing the spatial resolution to sample the territory, (iii) specifying a conceptual model to represent the crop successions by means of second-order Hidden Markov Models, and (iv) finally segmenting the territory into homogeneous areas that will be further investigated.