With increasing urbanization and the complexity of modern infrastructures, recognizing complex objects in high and very high spatial resolution satellite images has become a major challenge, especially in urban environments. Deep learning-based methods, while powerful, often require large annotated datasets and struggle to model the complex spatial and contextual relationships between urban objects. This paper proposes a dual-level hybrid approach that integrates evolving ontologies with Constraint Satisfaction Problems (CSP) to model complex objects, while improving their resolution efficiency through a method combining FAC-3, backjumping and particle swarm optimization (PSO). The main objective is to improve complex object recognition by simultaneously modeling spatial and semantic relationships while optimizing the exploration of possible solutions. CSP captures spatial constraints, FAC-3 reduces the search space, and PSO avoids local minima, enhancing overall search efficiency. The formalization of the ontology, CSPs and their translation into SWRL rules, as well as their use, are detailed. Finally, an illustrative application of complex urban object recognition is presented.