Foresters are faced with difficulties to obtain sub-polygon information with the mapping methods available nowadays. The main objective of this work consisted in the development of new methods able to improve the map accuracy of regenerating forest stands and mature forest stands in the South of Québec, Canada. The Dempster-Shafer Theory (DST) and the Dezert-Smarandache Theory (DSmT) showed their ability to integrate multiple heterogenous data sources to go further than the classical classification procedures like the maximum likelihood or the spectral unmixing, in terms of map accuracy. Improvement on the ability to map regenerating stands, passed from 82.7% with the maximum likelihood method to 91.1% with the Free DSm model with a total transfer of the mass of the"Union" class to the"Intersection" class (+ 8.4%). For the mature stands, the improvement passed from 63.8% with the K nearest neighbour to 79.5% with the DST according to a classical belief structuration and the hybrid decision rule for which the conflict threshold was fixed at 10% (+ 15.7%). Our results with DST and a bayesian belief structuration showed the difficulty to model the uncertainty in the fusion process. This is probably due to the lack of scientific knowledge about the influence of the biophysical and climatic parameters on the mapped forest stands and to the necessity to model specifically the uncertainty for each source. Our work showed concrete improvement when mapping forest stands with DST which is encouraging to continue explorating the fundamental principle of the proposed hybrid decision rule. This means a particular focus on the difference between the fused masses of each potential class after the fusion, to choose the best hypothesis.