We　propose　a　novel　method　called　exemplar　finder　(EF)　for　spectral　data　endmember　extraction　problem,　which　is　also　known　as　blind　unmixing　in　remote　sensing　community.　Exemplar　finder　is　based　on　data　self　reconstruction　assuming　that　the　bases　(endmembers)　generating　the　data　exist　in　the　given　data　set.　The　bases　selection　is　fulfilled　by　minimising　a　l2/l1　norm　on　the　reconstruction　coefficients,　which　eliminates　or　suppresses　irrelevant　weights　from　non-exemplar　samples.　As　a　result,　it　is　able　to　identify　endmembers　automatically.　This　algorithm　can　be　further　extended,　for　example,　using　different　error　structures　and　including　rank　operator.　We　test　this　method　on　semi-simulated　hyperspectral　data　where　ground　truth　is　available.　Exemplar　finder　successfully　identifies　endmembers,　which　is　far　better　than　some　existing　methods,　especially　when　signal　to　noise　ratio　is　high.　©　2013　Springer-Verlag.