In order to solve the problem of destroying the physical meaning of spectral curve of pixels in dimension reduction of traditional N-FINDR algorithm, the best exponential method of band selection was used instead of feature extraction. The dimension reduction method of N-FINDR algorithm was improved. Experiments were carried out using the simulated and real hyperspectral data. The improved N-FINDR algorithm and other two algorithms were used to extract the terminal elements respectively. Full constrained least squares method was used to solve the mixing problem. The results show that the improved N-FINDR algorithm has higher precision and uses less time. It is feasible to use band selection instead of feature extraction to improve the dimension reduction method and retain the physical meaning of spectral curve in N-FINDR algorithm.