Abstract: It is difficult to find correspondence points on the satellite images of the high altitude mountainous area of west China because of the geographic and geomorphic conditions. To solve the multispectral deviation of the satellite images, this paper use SIFT-based algorithm, regional gray-based algorithm and artificial registration to do registration experiments on domestic satellite images such as ZY1-02C, ZY-3, GF-1 and so on. Then visual method and error method are used to evaluate precisions, the results of which are contrasted and analyzed. The experimental results show that:artificial method has higher registration precision, but the efficiency is poor; regional gray-based method is affected by luminance difference which extracts least correspondence features and has the lowest registration precision; SIFT-based method has the highest degree of automation which can extract a huge amount of correspondence features and can also screen out false matching points. It has a higher registration precision.

Key words: domestic satellite image, image registration, SIFT algorithm, registration accuracy analysis