We take an iterative, four-step approach for our matching
method. We first calculate a rough offset between the nadir
and aft views (bands 3N and 3B) for each pixel, and then
refine and fine-tune the offset. This is achieved by selecting
tie-points common in both images, generating a polynomial
(usually of degree 1) to relate nadir coordinates to aft
coordinates, and resample the aft scene in nadir space based
on this polynomial. A frequency domain cross-correlation is
then used to estimate the offset between images at each pixel
using a roving kernel, or window of pixels, starting with a
kernel of 16 16 pixels. The cross-correlation (Corr) of a
kernel in the nadir (n) and the aft (a) is given by the relation