In order to realize automatic registration of a Synthetic Aperture Radar(SAR) image, an approach of image multi-scale registration based on affine invariant Fast Kernel Independent Component Analysis-Scale Invariant Feature Transform(FKICA-SIFT) features is presented. First, the affine invariant SIFT descriptors are constructed according to the Hessian matrix of feature points. The FKICA is used to extract the independent components of the affine invariant SIFT descriptors to obtain new descriptors (FKICA-SIFT). After filtering the input images by using Steerable pyramid, the new descriptors are used to match the feature points detected from the synthetic images of the band-pass sub-images in each layer. Finally, a coarse-to-fine procedure is adopted for gradual optimizing transformation parameters to achieve the multi-scale registration results. Experimental results show that the correct matching rate of proposed algorithm is more than 85% when the threshold is less than 0.7 and that is more than 90% when the threshold is less than 0.5. The registration accuracy of the proposed algorithm can achieve sub-pixel level and it is better than those of SIFT, PCA-SIFT, ICA-SIFT and SURF methods. It has been applied to accurate detection of the changes of the Tangjiashan waters before and after the Wenchuan earthquake, and obtained results meet the requirements of accurate registration for SAR images.