The sonar images provide a rapid view of the seabed in order to characterize it. However, in such as uncertain environment, real seabed is unknown and the only information we can obtain, is the interpretation of different human experts, sometimes in conflict. In this paper, we propose to manage this conflict in order to provide a robust reality for the learning step of classification algorithms. The classification is conducted by a multilayer perceptron, taking into account the uncertainty of the reality in the learning stage. The results of this seabed characterization are presented on real sonar images.