This study aims at identifying the best object-based classification strategy that takes advantage of the richness of hyperspectral data, for classifying 8 tree species in an urban area (Toulouse, France). Field spectral measurements at the leaf and canopy level were carried out in a reference site, while airborne hyperspectral Visible Near-Infrared (160 spectral bands, spatial resolution of 0.4 m) and Short-Wavelength Infrared (256 spectral bands, 1.6 m) were acquired over Toulouse. We propose an ensemble classifier approach (at least one classifier per species) such as each classifier uses three vegetation indices, followed by Support Vector Machine (SVM) supervised classification. Then, a decision rule based on the classifiers votes is applied to predict the species. The vegetation indices triplet corresponding to each classifier is chosen in such way that it optimizes the F-score of a certain species, ensuring the complementarity of the classifiers. In this framework, the field data are intended to be used for learning whereas the airborne data are used for testing, in order to assess the potential of field measurements for such classification task. Two baseline approaches are used for comparison. A standard classification procedure using directly the spectral reflectance is chosen in order to evaluate the interest of using vegetation indices. A method which stacks all the selected indices in one feature vector is considered in order to assess the potential of the ensemble classifier. These methods are first compared on the reference site. This allows the best strategy to be selected with a view to introducing the method in an automatic process (tree crown delineation and species classification) on a test site. Concerning the reference site, the proposed method outperforms the baseline approaches in case of leaf level learning with an Overall Accuracy (OA) of 55%, instead of 21% and 32% respectively. Aesculus hippocastanum trees are well classified because of their senescence, caused by the horse-chestnut leaf miner, and highlighted thanks to the vegetation indices. For the test site, the Tilia tomentosa trees of the main alignment are identified with an OA of 81% in case of leaf level learning. In conclusion, the proposed ensemble classifier approach improves the performance. Also, it is shown that leaf level learning gives similar performance in comparison to the use of references from the images.