Sparse hyperspectral unmixing is a relatively new method for automatic endmember detection and abundance estimation using large overcomplete dictionaries. This method suffers from the mutual similarity of the endmembers in the hyperspectral dictionary which has a significant impact on the stability and accuracy of the unmixing result. In this work we introduce a new sparse unmixing algorithm with endmember groups pre- selection. Firstly, the library is clustered into groups with similar endmembers. Secondly, relevant groups are identified based on a preliminary sparse unmixing. Finally the abundance vector is calculated using the dictionary containing selected clusters. We assess our algorithm by comparing to standard sparse unmixing method using both simulated data and real hyperspectral HyMAP imagery.