Traditional N-FINDR algorithm suffers from complicated calculations,and is sensitive to noise and initial endmembers,resulting in wrong extraction by outliers. In order to solve these problems,an improved fast N-FINDR endmember extraction algorithm was proposed,which gets rid of the redundant information and reduces the search area of commonly used N-FINDR algorithm by spectral distance,then smoothes the spectral noise and gets rid of the outliers adaptively. In addition,the initial endmembers of N-FINDR are selected by maximizing the spectral distance,avoiding the blindness of random selection. The synthetic data and real hyperspectral data were used for simulation analysis,and the proposed method was compared with existing algorithms. The experiment results show that the proposed method is able to extract the endmembers correctly in the noise,and has higher extraction efficiency and robustness than existing algorithms.