Mapping is the technique used by robots to build up a map within an unknown environment, or to update previously build map within a known environment. The problem is related to integrate the information obtained by multiple sensors on a consistent model and describing it by a given representation. The main aspects of mapping are the interpretation of sensor data and the representation of the environment. Topological approaches divide the environment into significant areas, being the aim to capture the connectivity of these areas rather than creating a geometrically accurate map. In this context, this paper proposes a method for mapping generic environments (structured or not) based on several semantic maps. In our implementation, each map can be described as a topological map, which is modeled using self-organizing neural networks. The approach was implemented and validated in a set of environments using Pioneer robots, equipped with an omni directional camera and a GPS. All the results were obtained using the robot simulator We bots, due its facility to test extreme conditions. Issues related to high dimensionality, perceptive correspondence and dynamicity have been evaluated. The results show the capabilities of the method to reduce data dimensionality and the generalization of the proposal.