Nowadays, satellite images are widely exploited in many fields including agriculture, cartography, regional planning, etc. The various advancement in satellite technologies have made satellite images abundantly available. In addition to the specificity and complexity of remote sensing data (spatial and temporal dimensions), the huge amount of this data makes its analysis challenging. Hence, data mining approaches are more adapted to satellite image analysis. In this article, we propose an unsupervised method that combines an object-oriented analysis with a clustering approach to characterize the evolutions in satellite image time series. Our objective is to track the evolutions of land use from satellite image time series. Our framework takes as input raw satellite image time series and their segmentation to highlight spatio-temporal entities that evolve similarly. The method first analyses the segments to detect reference objects among them. The reference objects are meaningful structures of the landscape we want to track. Once they are selected, their evolution over time is then described using a graph-based representation. The spatial extent of the reference object is used to describe its evolution over time. For each reference object, we select the image segments it overlaps and use them to create a graph, named evolution graph. The idea is to represent the state of the reference object on each image in the time series. Finally, the resulting evolution graphs are transformed to synopsis before clustering. The synopsis aims to synthesize the radiometric information contained in the evolution graphs. While the clustering algorithm organizes them in homogenous groups according to their evolution type. The choice of the clustering algorithm and the distance metric is crucial to make the method suitable for handling satellite images time series with their specificity. To evaluate the proposed framework, the experimentations were carried on two study areas: (i) the Lower Aude Valley and (ii) the Pic Saint Loup. These areas are a Natura 2000 sites located in the south of France (https://inpn.mnhn.fr/site/natura2000/listeSites). To track the evolution on these areas, we exploit different satellite images: Spot-2, Spot-4, Spot-5 and Landsat-8. These satellite images are available in the context of the SpotWorld Heritage program (https://www.theia-land.fr/fr/projets/spot-world-heritage), they constitute a multi-annual time series over twenty-six years from 1990 to 2016. Our aim is to perform an inter-site analysis on multi-annual satellite images time series using an object-oriented approach. The results have shown that the proposed method highlights different evolution types according to the different land cover. In our study, the results obtained when gathering the two sites are more accurate than the one obtained analyzing each site independently. The obtained clusters have been compared to an expert classification using a clustering validity index. The identified clusters have also been described and discussed to understand and analyze the behaviour of each group.