In this paper, we take an interest in representation and treatment of imprecision and uncertainty in order to propose an original approach to approximate reasoning. This work has a practical application in supervised learning pattern recognition. Production rules whose conclusions are accompanied by belief degrees, are obtained by supervised learning from a training set. The proposed learning method is multi-featured, it allows to take into account the possible predictive power of a simultaneously considered feature conjunction. On the other hand, the feature space partition allows a fuzzy representation of the features and data imprecision integration. This uncertainty is managed in the learning phase as well as in the recognition one. To introduce more flexibility and to overcome the boundary problem due to the manipulations of membership functions of fuzzy sets, we propose to use a context-oriented approximate reasoning. For this purpose, we introduce an adequate distance to compare neighbouring facts. This distance, measuring imprecision, combined with the uncertainty of classification decisions represented by belief degrees, drives the approximate inference.The proposed method was implemented in a software called SUCRAGE and confronted with a real application in the field of image processing. The obtained results are very satisfactory. They validate our approach and allow us to consider other application fields.