Study on Spot5 remote sensing imagery automatic registration methods based on texture feature points

Due to the problem of lower-efficiency and subjective error generated by manual operation in the process of Spot5 remote sensing satellite imagery matching, a new efficient automatic registration scheme based on imagery characteristic points was proposed. Firstly, the preprocessed 1A level multispectral imagery and panchromatic imagery were registered coarsely based on affine transformation in this workflow using the ephemeris in metadata files. Secondly, the grid constrained feature points were extracted from the two images by improved Forstner operator, and tie-points pairs were obtained by a coarse-to-fine matching using Euclid distance and correlation. Finally, error pairs were eliminated by least median of squares (LSM). After getting enough control point pairs which were high-precision matched and distributed evenly, coarse-registered multispectral imagery was rectified in local small region based on Delaunay triangulation network. Experiments were performed on many volume images, and the results show that the root mean square errors (RMSE) of the check points are less than 0.5 pixels.