The segmentation of remotely sensed images acquired over tropical forests is of great interest for numerous ecological applications, such as forest inventories or conservation and management of ecosystems, for which species classification techniques and estimation of the number of individuals are highly valuable inputs. In this paper, we propose a method for hyperspectral image segmentation, based on the binary partition tree (BPT) algorithm, and we apply it to two sites located in Hawaiian and Panamean tropical rainforests. Different strategies combining spatial and spectral dimensionality reduction are compared prior to the construction of the BPT. Various superpixel generation methods including watershed transformation and mean shift clustering are applied to decrease spatial dimensionality and provide an initial segmentation map. Principal component analysis is performed to reduce the spectral dimensionality and different combinations of principal components are compared. A non-parametric region model based on histograms, combined with the diffusion distance to merge regions, is used to build the BPT. An adapted pruning strategy based on the size discontinuity of the merging regions is proposed and compared with an already existing pruning strategy. Finally, a set of criteria to assess the quality of the tree segmentation is introduced. The proposed method correctly segmented up to 68% of the tree crowns and produced reasonable patterns of the segmented landscapes.