Sparse unmixing has been proposed for hyperspectral image analysis. It has been shown that improved performance can be achieved when endmembers from a spectral library are used. However, when endmembers from image data have to be employed for unmixing, such a sparse-constrained approach may be problematic due to the fact that endmembers are generally highly coherent, thereby producing unstable sparse solution. In this paper, we propose two approaches to improve the incoherence of the dictionary, i.e., band selection and random projection. Experimental results demonstrate significant improvement when using pixel endmemebers extracted from image data itself.