An affine Invariant Feature is proposed for remote sensing image registration.The key points are extracted from image as local extrema both in scale space differential and 2D Laplacian.For each key point a series of affine parameters templates are applied to transform the neighborhood pixels，which produce several slices so that projection distortion could be mitigated.Local invariant feature descriptors are extracted from these slices using scale-space invariant features transform (SIFT).An initial matching set is obtained by matching SIFT features using a special similarity criterion.The mismatched key points are removed with greedy algorithm iteratively，and then the matching set is modified.The image transformation parameters are estimated based on this modified set.The performance of this algorithm is tested under several interferences including affine transformation，partly barrier，tonal distortion，loss of contrast，and Gaussian noise.This method is confirmed with some filed satellite images.