A method for generating natural color composite of satellite images based on local endmembers of ground features was proposed. First, the reference satellite image which has similar land cover types with the target satellite image is segmented into multiple local patches. Secondly, endmembers of ground features are extracted from each local patch, then we choose better endmembers and gather them into a sample set. Thirdly, we use the sample set to build up the relationship between the spectral values of the blue band and the other bands. Finally, the spectrum relationship is used to generate natural color composite of the target image. The verification experiment on Landsat TM5 images shows that the proposed method is reliable and stable to generate the natural color composite of images. The other experiment on IRS-P6 images shows that our method is able to promote for other satellite images. In experiments, we also compared the existing “reference image-image” method and “spectral library-image” method qualitatively and quantificationally, indicating that our method is more precise to simulate spectrums of all kinds of ground features.