An automatic procedure for SAR-optical satellite image registration based on multi-layer feature matching strategy

This paper proposes a hierarchical registration strategy for multi-sensor satellite images. The method is a coarse-to-fine registration procedure. In the stage of coarse registration, area features are depicted by both shape matrix and principle component analysis (PCA), which provides not only a unified rule for the shape similarity measure but also the shape orientation consistency of this pair of area features. Therefore, the ambiguity of feature matching can be effectively avoided. For the registration refinement, the parametric description of edge features by nonuniform rational B-spline (NURBS) solves the problem of the "irregular edge" of the SAR image caused by speckle noise when registering SAR-optical data, and avoids mismatching of image features with local deformation, exploiting the local controllability of NURBS curves as well. By combining some constraints on two corresponding curves, the matching of control points of NURBS, which are invariant under affine transformation, can be achieved. Registration experiments on SAR-optical satellite images demonstrate the reliability and effectiveness of the presented method.