Classification is one of the major issue in image analysis and processing for remote sensing applications. Though classification based on texture analysis - land-use, forests, cities, etc - is the purpose of numerous works, classification of curvilinear networks is hardly processed. However, it is of major interest, in particular for image indexing and images matching, because it is a main feature whose global shape does not change with sensors nor point of view. This paper introduces a new approach aiming at (i) building the networks from extracted curvilinear-like features and (ii) classifying them into roads, highways, rivers. The main idea is to use a decision tree taking into account a priori knowledge. Classification and graph building are achieved simultaneously using a hypothesis generation/propagation scheme. The resulting network is encoded as a graph with a multi-scale description. Illustrations given on satellite optical SPOT images show encouraging results.