In the petroleum exploration/production context, characterizing the reservoir quality, identifying the main rock types, or predicting their spatial variations is a challenge for the industry. To achieve this purpose, supervised pattern recognition methods, as discriminant analysis are widely used. These methods aim at calibrating, when possible, a relationship between field features -for example a set of borehole measurements, or a set of seismic attributes- and a predefined set of classes -for example, different rock types-. It has yet the major drawback not to take into account the uncertainties on the measurement arrays, which may cause drastic misinterpretations of reservoir characteristics. The methodology developed is an extension of the standard parametric approach to discriminant analysis. The calibration phase follows the same main steps as the standard algorithm, except that all the processed quantities are intervals. The different interval computations are based on interval arithmetic. Eventually, any imprecise object is assigned to a subset of classes, consistent with the measurements and their uncertainties. As a result, the computed reservoir quality model is less precise, but more realistic, taking into account data and its uncertainties.