This paper proposes an approach for pixel unmixing based on possibilistic similarity. The approach exploits possibilistic concepts to provide flexibility in the integration of both contextual information and a priori knowledge. Possibility distributions are first obtained using a priori knowledge given in the form of learning areas delimitated by an expert. These areas serve for the estimation of the probability density functions of different thematic classes also called endmembers. The resulting probability density functions are then transformed into possibility distributions using Dubois-Prade's probability-possibility transformation. The pixel unmixing is then performed based on the possibilistic similarity between a local possibility distribution estimated around the considered pixel and the obtained possibility distributions representing the predefined endmembers in the analyzed image. Several possibilistic similarity measures have been tested to improve the discrimination between endmembers. Results show that the proposed approach represents an efficient estimator of the proportion of each endmember present in the pixel (abundances) and achieves higher classification accuracy. Performance analysis has been conducted using synthetic and real images.