Abstract: In the current collaborative sparse unmixing of hyperspectral data, the fractional abundances can not be estimated accurately due to ignoring the differences of endmembers among different pixels. In this paper, a novel unsupervised clustering method is proposed as a preprocessing step to generate several classes of pixels with the same endmember bundles, and then for each class, the collaborative sparse unmixing technique is used to implement spectral unmixing. In terms that the pixels with the same set of active atoms have the smallest values of collaborative sparse coding, the sum of reconstruction errors and sparsity levels are introduced as the distance metric in the unsupervised clustering. As such, the same class pixels can be guaranteed to contain the same endmembers. Finally, the involving optimization problem can be solved by using the algorithm of alternating direction method of multipliers (ADMM). Experimental results on synthetic and real hyperspectral data demonstrate that the our proposed algorithm can identify the actual endmembers effectively and improve the accuracy of the fractional abundance estimation.