The optimal exploitation of the information provided byhyperspectral images requires the development of advancedimage processing tools. This paper introduces a new hierarchicalstructure representation for such images using binarypartition trees (BPT). Based on region merging techniques usingstatistical measures, this region-based representation reducesthe number of elementary primitives and allows a morerobust filtering, segmentation, classification or informationretrieval. To demonstrate BPT capabilites, we first discussthe construction of BPT in the specific framework of hyperspectraldata. We then propose a pruning strategy in order toperform a classification. Labelling each BPT node with SVMclassifiers outputs, a pruning decision based on an impuritymeasure is addressed. Experimental results on two differenthyperspectral data sets have demonstrated the good performancesof a BPT-based representation