With　the　gradual　maturation　of　imaging　spectroscopy,　the　demand　for　quantitative　analysis　of　hyperspectral　images　grows　with　each　passing　day.　Spectral　unmixing　has　been　considered　as　an　efficient　way　to　extract　detailed　information　about　land　covers.　In　this　paper,　by　introducing　the　co-training　concept　into　the　spectral　unmixing　method　based　on　wavelet　weighted　similarity　(WWS-SU),　a　spectral　unmixing　method　based　on　co-training　(CT-SU)　is　proposed.　Compared　with　the　WWS-SU　method　on　synthetic　hyperspectral　image,　the　CT-SU　method　shows　not　only　more　practical　but　also　more　accurate　in　result.